Multimodal 3D object detection for autonomous driving, a task for real-world applications, poses substantial challenges in maintaining robust performance under various perturbations and complex environmental conditions. However, most existing approaches primarily focus on performance optimization under relatively ideal scenarios or focus on one or few disturbing conditions (or adverse conditions), lacking systematic exploration of robustness against real-world factors, including high class imbalance, adverse weather conditions, sensor jitter and failures, and significant scene variations. To address this issue, we propose a robust multimodal 3D detector, termed RobusTor3D, which integrates robustness at both the structural and supervisory levels by blending the knowledge from Vision-Language Models (VLMs). Structurally, textual descriptions are incorporated to enhance the semantic richness and diversity of rare classes. This novel semantic injection operation compensates for the inherent class imbalance and modality weakness in conventional visual features. Furthermore, semantic alignment capability and robust representation by Vision-Language Knowledge Extraction (V-LKE) serve as semantic priors to complement modality-specific representations, significantly improving model adaptability. At the supervisory level, we propose a Scene-level Multimodal Consistency Learning (SMCL) strategy, which jointly enforces global semantic constraints across modalities, encouraging the learning of stable and abundant semantic representations. This special design specifically reduces the impact of spatial alignment, while notably enabling semantic compensation under modality-loss conditions. Extensive robustness experiments conducted on KITTI, KITTI-C, and CADC benchmarks evaluate five robustness aspects, including long-tail problem, adverse weather (rain, snow, fog, strong sunlight), sensor spatial misalignment and motion blur, modality loss, and cross-domain scenarios. The results show that RobusTor3D demonstrates superior robustness across all five evaluated aspects. It consistently outperforms the state-of-the-art methods under various challenging conditions.

Early warning systems for disease outbreaks play a crucial role in public health for management and contingency planning. However, most predictive modeling works focus on flat models that incorporate exogenous inputs (e.g. climate, demographics) to predict future outbreaks at different locations, but do not jointly model multiple spatial aggregation levels. In this paper, we introduce HierarNet, a unique independent-interactive hierarchical forecasting framework that aims to predict disease outbreaks at different levels of spatial resolution, such as provinces, regions, and nations. HierarNet consists of two main phases. In the local phase, we train independent forecasting models for all locations at all levels. In the global phase, all models iteratively interact with others across different levels via their hierarchical relationships under an ensemble fashion to maximize their agreements. This global local hierarchical interactive scheme makes HierarNet a highly effective and flexible method (i.e. it can work with an arbitrary base prediction model and available exogenous data for each location independently). Extensive experiments are conducted on various disease datasets (e.g., Dengue fever, flu, diarrhea, and Bluetongue) in different countries (e.g., France, Vietnam, and USA) to show the performance of HierarNet compared to 19 state-of-the-art (SOTA) methods such as MinT, DYCHEM, WITRAN, SegRNN, TSMixer, PatchTST, or iTransformer. We also illustrate the generability of HierarNet in other domains, e.g., web traffic forecasting.

Explainable AI (XAI) is essential for trust and transparency in deep learning, especially in medical imaging. Existing local explanation methods provide per-instance insights but fail to show whether similar explanations hold across samples of the same class. This limits global interpretability and demands time-consuming manual review by clinicians to trust models in practice. We introduce the Consensus Alignment Score (CAS), a novel metric that quantifies consistency of explanations at the class level. We also present ConsensusXAI, an open-source, model- and method-agnostic framework that evaluates explanation agreement quantitatively (via CAS) and qualitatively (through consensus heatmaps) per class. Unlike prior benchmarks, ConsensusXAI uses a latent-space clustering approach, Latent Consensus, to identify dominant explanation patterns, exposing biases and inconsistencies towards certain classes. Evaluated across two different medical imaging modalities for both correct and incorrect predictions on two different backbones, our method consistently reveals meaningful class-level insights, outperforming traditional consensus method i.e. SSIM, and enabling faster, more confident clinical adoption of AI models. The source code is available at https://github.com/a-haider1992/cas_toolbox.

• Integrates autoencoder with trainable GMM for universal outlier detection. • GMM feedback optimizes autoencoder, enhancing inlier/outlier separation. • Accurately captures complex data patterns to boost detection robustness. • Outperforms state-of-the-art methods across datasets in accuracy and reliability Outlier detection is a critical process in data engineering. Leveraging machine learning techniques for outlier detection enables the handling of large-scale, high-dimensional data, enhancing detection accuracy and efficiency. Traditional methods typically model data directly in the data space. However, these approaches often struggle to accurately distinguish inliers from outliers when dealing with complex data distributions. GMM can flexibly fit complex, multi-peak distributions using multiple Gaussian components and effectively identify outliers through probabilistic modelling. We introduce a novel outlier detection approach, which improves detection efficiency by indirectly modelling data in a latent space using a Gaussian Mixture Model (GMM). This approach aligns with a growing trend in AI, notably advocated by Yann LeCun, that emphasizes decision-making and learning in latent representation spaces, instead of depending on raw token or feature spaces. For this, we design an encoder-decoder neural network with a GMM as the decision layer, enabling effective identification of outliers through probabilistic modelling. Our method not only addresses practical needs in anomaly detection but also contributes to this broader trend of latent space modelling as a step toward more autonomous and generalisable learning systems. Extensive evaluations on public and proprietary datasets demonstrate that our method outperforms existing approaches, including DAGMM and ECOD, highlighting its superiority in accuracy.

Effective detection of financial market manipulation is critically impeded by three fundamental challenges: signal concealment, data sparsity, the boundary vagueness. This paper introduces SD-FMM , a S elf-supervised D etection framework tailored for F inancial M arket M anipulation that addresses these fundamental challenges through three innovative components. First, our Amplification Component extracts and fuses domain-specific features grounded in market microstructure theory, substantially amplifying subtle manipulation signals that would otherwise remain concealed. Second, our Synthesis Component generates realistic synthetic anomalies through few-shot learning and dynamic frequency analysis using Discrete Wavelet Transform, enabling self-supervised training without relying on scarce labeled data. Third, our Detection Component employs a novel Dual-branch Contrastive Detection Neural Network that enhances sensitivity to manipulation boundaries through local contrastive learning and holistic modeling of temporal dependency. We evaluate SD-FMM using a newly collected proprietary dataset of 25 Chinese stock market manipulation cases and a public benchmark of 338 cryptocurrency pump-and-dump schemes. Extensive experiments against 12 state-of-the-art baselines demonstrate the significant superiority of SD-FMM. On the stock dataset, our method outperforms the second-best baseline by 47.61% in average precision metrics and reduces the false alarm rate by 47.46%. Meanwhile, it shortens the mean detection delay by 25.05%, enabling swift regulatory intervention. On the cryptocurrency dataset, SD-FMM exhibits remarkable sensitivity, achieving a Hit Rate@3 of 83.13% and Hit Rate@20 of 97.93%. Overall, our framework offers a generalized solution that can not only accurately distinguish manipulations from normal trading but also deliver a faster and stronger response to manipulations across diverse financial markets.

Fuzzy Cognitive Maps (FCMs) are a prominent soft computing technique for time series forecasting, valued for their ability to effectively model complex temporal dynamics. While FCM learning algorithms improve the performance of FCM-based predictors by capturing causal relationships between nodes, existing approaches predominantly rely on offline time series data stored in static repositories. This limitation hinders their adaptability to dynamic changes in map structures over time, making them unsuitable for real-time streaming data analysis and dynamic modeling of evolving causal relationships. Furthermore, the non-stationary nature of real-world time series presents significant challenges to the predictive performance of FCM-based models. To overcome these limitations, we propose a novel hybrid offline-online FCM learning algorithm that integrates a non-stationarity detection mechanism with a knowledge-guided least squares (KGLS) method. In the offline phase, an initial FCM-based predictor is constructed from historical data, where the recursive least squares (RLS) method is employed to capture long-term causal relationships using a sliding window technique. The online phase incrementally updates the model using streaming data, guided by a non-stationarity detection mechanism based on statistical hypothesis testing. The mechanism classifies data shifts into stable, warning, and drift levels. To mitigate catastrophic forgetting, the KGLS method maintains a compact yet representative memory buffer of past data samples. During training, these samples are replayed alongside new data, enabling the model to reinforce previously learned patterns while adapting to new information. Extensive experiments on stationary and non-stationary datasets demonstrate that our method achieves superior overall prediction performance and accurately forecasts trends in non-stationary time series in real time.

[Context]: It is well known that understanding the evolution of technologies and its cause is essential for more discoveries and innovations. In the Artificial Intelligence (AI) domain, it has also been identified that scrutinising the development context and path of AI will be able to help both academia and industry better understand the current AI limitations, reveal future AI trends, and facilitate AI/digital transformations. [Objectives]: Given the dramatic boom of modern AI, this research aims to unearth the evolution pattern along the recent three AI waves (namely predictive AI, generative AI and agentic AI), and accordingly to guide AI research and development to focus on the most promising directions. [Method]: We employed analogical reasoning as the research method and referred to the existing software architectural styles to inspire our understanding of the architectural evolution of modern AI technologies. [Results]: We see a service-oriented trend in modern AI's working mechanisms, and the offering of AI power seems to be transiting from a heavyweight and monolithic paradigm to an organisational and collaborative paradigm with more and more specific separation of concerns. Following this service-oriented evolution trend, we borrow software architecture lessons and foresee opportunities to grow the current AI wave to a further height, e.g., standardising AI agent-friendly APIs and developing serverless AI agents. [Conclusions]: What is happening in the AI domain has happened before in the software engineering domain. It is worth reusing software architecture knowledge to evolve the architecture of AI technologies.

Recent advancements in question generation (QG) have been significantly propelled by reinforcement learning (RL). Although extensive reward models have been designed to capture the attributes of ideal questions, their associated learning challenges, particularly in sample efficiency and diversity, remain underexplored. This paper introduces a bilevel policy decomposition (BPD) framework and a diversity-seeking RL (DSRL) objective to address these issues. The BPD framework utilizes two cascading policies to divide QG into two more manageable sub-tasks: answer-centric summary generation and summary-augmented QG, facilitating exploration and accelerating policy learning. Concurrently, the DSRL objective preserves the inherent diversity of QG by ensuring the bilevel policies align probabilistically with their reward models rather than merely maximizing returns. Our integrated approach, named BPD-DSRL, demonstrates superior performance over existing baselines on multiple question quality and diversity metrics across various QG benchmarks.

Data-Efficient Generative Adversarial Nets (DE-GANs) have become more and more popular in recent years. Existing methods apply data augmentation, noise injection and pre-trained models to maximumly increase the number of training samples thus improving the training of DE-GANs. However, none of these methods considers the sample quality during training, which can also significantly influence the training of DE-GANs. Focusing on sample quality during training, in this paper, we are the first to incorporate discriminator rejection sampling (DRS) into the training process and introduce a novel method, called quality aware dynamic discriminator rejection sampling (QADDRS). Specifically, QADDRS consists of two steps: (1) the sample quality aware step, which aims to obtain the sorted critic scores, i.e., the ordered discriminator outputs, on real/fake samples in the current training stage; (2) the dynamic rejection step that obtains dynamic rejection number N, where N is controlled by the overfitting degree of discriminator (D) during training. When updating the parameters of D, the N high critic score real samples and the N low critic score fake samples in the minibatch are rejected dynamically based on the overfitting degree of D. As a result, QAD-DRS can avoid D becoming overly confident in distinguishing both real and fake samples, thereby alleviating the over-fitting of D issue during training. Extensive experiments on several datasets demonstrate that integrating QADDRS into different DE-GANs can achieve better performance and deliver state-of-the-art results. Codes are available at https://github.com/zzhang05/QADDRS.

Diffusion-based generative models (diffusion models) often require a large amount of data to train a score-based model that learns the score function of the data distribution through denoising score matching. However, collecting and cleaning such data can be expensive, time-consuming, and even infeasible. In this paper, we present a novel theoretical insight for diffusion models that two factors, i.e., the denoiser function hypothesis space and the number of training samples, can affect the denoising score matching error of all training samples. Based on this theoretical insight, it is evident that minimizing the total denoising score matching error is challenging within the denoiser function hypothesis space in existing methods, when training diffusion models with limited data. To address this, we propose a new diffusion model called Limited Data Diffusion (LD-Diffusion), which consists of two main components: a compressing model and a novel mixed augmentation with fixed probability (MAFP) strategy. Specifically, the compressing model can constrain the complexity of the denoiser function hypothesis space and MAFP can effectively increase the training samples by providing more informative guidance than existing data augmentation methods in the compressed hypothesis space. Extensive experiments on several datasets demonstrate that LD-Diffusion can achieve better performance compared to other diffusion models. Codes are available at https://github.com/zzhang05/LD-Diffusion.

Reliable factuality evaluation is critical for the iterative development of open-domain question answering (ODQA) systems, especially given the rise of large language models (LLMs) and their propensity for hallucination. However, state-of-the-art (SOTA) automatic metrics, which are mostly supervised, remain notably less reliable than humans. In this paper, we find two key challenges behind this gap: (1) length distribution mismatch between lengthy LLM answers and shorter training answers used by current metrics; and (2) reference incompleteness, where current metrics often misjudge valid system answers absent from given references-a challenge worsened by the diversity of LLM outputs. To address these issues, we present a new ODQA factuality evaluation dataset called OBELLA (Open-Book Evaluation for Long-form LLM Answers). OBELLA narrows the length distribution mismatch by significantly increasing the candidate answer length to align with LLM outputs. Moreover, it introduces a neutral class for plausible yet under-supported candidate answers to differentiate reference incompleteness from outright incorrectness, thus enabling flexible reevaluation by consulting external knowledge for more references. Based on OBELLA, we propose a novel metric named OBELLAM (OBELLA Metric). OBELLAM integrates a cross-attention mechanism to enhance long-form candidate answer representations and employs a dynamic closed-open book evaluation strategy to tackle reference incompleteness. Our OBELLAM sets a new SOTA in aligning with human judgments across two ODQA evaluation benchmarks, marking a promising step toward more robust ODQA factuality evaluation.

Idea generation describes a creative process involving reasoning over some knowledge to derive new information. Traditional approaches such as mind-map and brainstorming are limited and often fail due to lack of quality ideas and ineffective methods. The reasoning capability of large language models (LLMs) have been investigated for ideation tasks and have reported interesting performance. However, these models suffer from limited logical reasoning capability which hinders the use of structural and factual real-world knowledge in discovery of latent insight and predict possible outcome when applied to ideation. In addition, the possibility of LLMs regurgitating knowledge learnt from datasets might adversely impact the degree of novel ideas the models can generate. In this paper, a two-stage logical reasoning approach is applied to initiate the search for candidate idea pathways based on the knowledge graphs (KGs) to address the problem of reasoning, domain-specificity and novelty. The divergence stage this reasoning explores utilizes a new node importance estimation (NIE) technique over KGs to discover latent connections supporting idea generation. In the convergence stage of this reasoning, subgraph matching using analogical reasoning (SMAR) is applied to find matching patterns to describe a new idea. The use of SMAR + NIE and KGs helps to achieve an improvement in reasoning over KGs before transferring such reasoning to LLMs for translation of idea into natural language. To evaluate the degree of novelty of ideas generated, a relevance-to-novelty scoring metrics is proposed based on multiple premise entailment (MPE). We combined this metric with other popular metrics to evaluate the performance of SMAR + NIE on benchmark datasets, and as well on the quality of ideas generated. Findings from the study showed that this approach demonstrates competitive performance with mainstream LLMs in idea generation tasks.

Anomaly detection in time series is crucial for applications ranging from finance to industrial monitoring. Effective models need to capture both the inherent characteristics of time series data and the distinct patterns of anomalies. While traditional forecasting-based and reconstruction-based approaches have been successful, they tend to struggle with complex and evolving anomalies. For instance, stock market data exhibits ever-changing fluctuation patterns that defy straightforward modelling. In this paper, we propose a novel method called TDSRL (Time Series Dual Self-Supervised Representation Learning) for robust anomaly detection. TDSRL attach great importance to the frequency domain information throughout the anomaly modelling process. We introduce a data degradation method that simulates real-world anomalies more naturally by operating in both time and frequency domains. Additionally, the key innovations also lie in dual self-supervised pretext tasks: one task characterises anomalies in relation to the entire time series, and the other focuses on local anomaly boundaries using contrastive learning. This significantly improves the network’s discrimination between anomaly and adjacent normal intervals. Consequently, TDSRL is expected to achieve a faster and stronger response to the anomalies, with the potential for early detection. Experimental results show that TDSRL outperforms state-of-the-art methods, making it a promising new direction for time series anomaly detection. The code of our paper is available here: https://github.com/ys-Dai/TDSRL/tree/main.

Link prediction, a key technique for knowledge graph completion, has advanced with transformer-based encoders utilizing high-dimensional embeddings and self-attention mechanisms. However, these approaches often result in models with excessive parameters, poor scalability, and substantial computational demands, limiting their practical applicability. To address these limitations, this paper introduces a low-dimensional link prediction model that leverages cross-attention for improved efficiency and scalability. Our approach employs low-dimensional embeddings to capture essential, non-redundant information about entities and relations, significantly reducing computational and memory requirements. Unlike self-attention, which models interactions within a single set of embeddings, cross-attention in our model captures complex interactions between entities and relations in a compact, low-dimensional space. Additionally, a streamlined decoding method simplifies computations, reducing processing time without compromising accuracy. Experimental results show that our model outperforms most state-of-the-art link prediction models on two public datasets, WN18RR and FB15k-237. Compared to these top-performing methods, our model contains only 18.1 % and 25.4 % of the parameters of these comparable models, while incurring a performance loss of merely 2.4 % and 3.1 %, respectively. Furthermore, it achieves an average 72 % reduction in embedding dimensions compared to five leading models. A case study on drug repurposing further illustrates the model's potential for real-world applications in knowledge graph completion.

Change point detection is crucial for identifying state transitions and anomalies in dynamic systems, with applications in network security, health care, and social network analysis. Dynamic systems are represented by dynamic graphs with spatial and temporal dimensions. As objects and their relations in a dynamic graph change over time, detecting these changes is essential. Numerous methods for change point detection in dynamic graphs have been developed, but no systematic review exists. This paper addresses this gap by introducing change point detection tasks in dynamic graphs, discussing two tasks based on input data types: detection in graph snapshot series (focusing on graph topology changes) and time series on graphs (focusing on changes in graph entities with temporal dynamics). We then present related challenges and applications, provide a comprehensive taxonomy of surveyed methods, including datasets and evaluation metrics, and discuss promising research directions.

As the proliferation of video content continues, and many video archives lack suitable metadata, therefore, video retrieval, particularly through example-based search, has become increasingly crucial. Existing metadata often fails to meet the needs of specific types of searches, especially when videos contain elements from different modalities, such as visual and audio. Consequently, developing video retrieval methods that can handle multi-modal content is essential. In designing our novel video retrieval framework named Multi-modal Video Search by Examples (MVSE)1, we focused on accuracy (precision and recall), efficiency (retrieval time in seconds), interactivity, and extensibility, with key components including advanced data processing and a user-friendly interface aimed at enhancing search effectiveness and user experience. With the advent of Large Language Models (LLMs), the interaction between multimodal data, including image and audio has been transformed with a significant leap forward towards a bigger goal of artificial general intelligence. This workshop aims to bring together experts from diverse domains to explore the possibilities of developing novel ways of multimodal data search, understanding and interaction.

Recently, many studies have highlighted that training Generative Adversarial Networks (GANs) with limited data suffers from the overfitting of the discriminator (D). Existing studies mitigate the overfitting of D by employing data augmentation, model regularization, or pre-trained models. Despite the success of existing methods in training GANs with limited data, noise injection is another plausible, complementary, yet not well-explored approach to alleviate the overfitting of D issue. In this paper, we propose a simple yet effective method called Dual Adaptive Noise Injection (DANI), to further improve the training of GANs with limited data. Specifically, DANI consists of two adaptive strategies: adaptive injection probability and adaptive noise strength. For the adaptive injection probability, Gaussian noise is injected into both real and fake images for generator (G) and D with a probability p, respectively, where the probability p is controlled by the overfitting degree of D. For the adaptive noise strength, the Gaussian noise is produced by applying the adaptive forward diffusion process to both real and fake images, respectively. As a result, DANI can effectively increase the overlap between the distributions of real and fake data during training, thus alleviating the overfitting of D issue. Extensive experiments on several commonly-used datasets with both StyleGAN2 and FastGAN backbones demonstrate that DANI can further improve the training of GANs with limited data and achieve state-of-the-art results compared with other methods. Codes are available at https://github.com/zzhang05/DANI.

Data augmentation (DA) has shown its effectiveness in training Data-Efficient GANs (DE-GANs). However, applying DA in DE-GANs results in transforming the distributions of generated data and real data to augmented distributions of generated data and real data. This augmentation process could produce some out-of-distribution samples, known as the leaking of augmentations problem, which is highly undesirable in DE-GANs training. Although some methods propose "leaking-free" DAs for DE-GANs, we theoretically and practically argue that the leaking of augmentations problem still exists in these methods. To alleviate the leaking of augmentations in DE-GANs, in this paper, we propose a simple yet effective method called adaptive negative data augmentation (ANDA) for DE-GANs, with a negligible computational cost increase. Specifically, ANDA adaptively augments the augmented distribution of generated data using the augmented distribution of negative real data, where the negative real data is produced by applying negative data augmentation (NDA) on the real data. In this case, potential leaking samples can be presented as "fake" instances to the discriminator adaptively, which avoids the generator (G) learning such samples, thus resulting in better performance. Extensive experiments on several datasets with different DE-GANs demonstrate that ANDA can effectively alleviate the leaking of augmentations problem during training and achieve better performance. Codes are available at https://github.com/zzhang05/ANDA

Generative adversarial networks (GANs) have achieved great success and become more and more popular in recent years. However, understanding of the min-max game in GANs training is still limited. In this paper, we first utilize information game theory to analyze the min-max game in GANs and introduce a new viewpoint on the GANs training that the min-max game in existing GANs is unfair during training, leading to sub-optimal convergence. To tackle this, we propose a novel GAN called Information Gap GAN (IGGAN), which consists of one generator (G) and two discriminators (D1and D2). Specifically, we apply different data augmentation methods to D1and D2, respectively. The information gap between different data augmentation methods can change the information received by each player in the min-max game and lead to all three players G, D1and D2in IGGAN obtaining incomplete information, which improves the fairness of the min-max game, yielding better convergence. We conduct extensive experiments for large-scale and limited data settings on several common datasets with two backbones, i.e., BigGAN and StyleGAN2. The results demonstrate that IGGAN can achieve a higher Inception Score (IS) and a lower Fréchet Inception Distance (FID) compared with other GANs. Codes are available at https://github.com/zzhang05/IGGAN

Abstract As the proliferation of video content continues, and many video archives lack suitable metadata, therefore, video retrieval, particularly through example‐based search, has become increasingly crucial. Existing metadata often fails to meet the needs of specific types of searches, especially when videos contain elements from different modalities, such as visual and audio. Consequently, developing video retrieval methods that can handle multi‐modal content is essential. An innovative Multi‐modal Video Search by Examples (MVSE) framework is introduced, employing state‐of‐the‐art techniques in its various components. In designing MVSE, the authors focused on accuracy, efficiency, interactivity, and extensibility, with key components including advanced data processing and a user‐friendly interface aimed at enhancing search effectiveness and user experience. Furthermore, the framework was comprehensively evaluated, assessing individual components, data quality issues, and overall retrieval performance using high‐quality and low‐quality BBC archive videos. The evaluation reveals that: (1) multi‐modal search yields better results than single‐modal search; (2) the quality of video, both visual and audio, has an impact on the query precision. Compared with image query results, audio quality has a greater impact on the query precision (3) a two‐stage search process (i.e. searching by Hamming distance based on hashing, followed by searching by Cosine similarity based on embedding); is effective but increases time overhead; (4) large‐scale video retrieval is not only feasible but also expected to emerge shortly.

Abstract Facial identity is subject to two primary natural variations: time-dependent (TD) factors such as age, and time-independent (TID) factors including sex and race. This study aims to address a broader problem known as variation-invariant face recognition (VIFR) by exploring the question: “How can identity preservation be maximized in the presence of TD and TID variations?" While existing state-of-the-art (SOTA) methods focus on either age-invariant or race and sex-invariant FR, our approach introduces the first novel deep learning architecture utilizing multi-task learning to tackle VIFR, termed “multi-task learning-based variation-invariant face recognition (MTLVIFR)." We redefine FR by incorporating both TD and TID, decomposing faces into age (TD) and residual features (TID: sex, race, and identity). MTLVIFR outperforms existing methods by 2% in LFW and CALFW benchmarks, 1% in CALFW, and 5% in AgeDB (20 years of protocol) in terms of face verification score. Moreover, it achieves higher face identification scores compared to all SOTA methods. Open source code .

Fuzzy rough sets (FRSs) theory is an important granular computing method to deal with incomplete information systems, and the attribute reduction is a basic key issue in FRSs. In this article, we construct a novel framework for selecting the optimal reduct of FRSs with theoretical guarantees by considering the influence of granule size and incorporating the stability theory in machine learning. First, a granule-based soft-margin support vector machine (GSSVM) algorithm is proposed for classification tasks by introducing the$\lambda$-conditional entropy into the hinge loss function, which takes the impact of granule size on data loss into account. Then, according to stability theory, the generalization error bound of the GSSVM algorithm is derived as a theoretical guarantee for selecting the optimal reduct. Finally, an optimization attribute reduction algorithm (RDROAR) based on the relative discernibility relation is presented by removing the attributes with low importance in a reduct while ensuring the generalization ability of GSSVM. Numerical experiments prove the effectiveness of the improved algorithm as well as verify the rationality and effectiveness of the optimal reduct.

Accurate air quality forecasting is crucial for public health and addressing air pollution. However, the dynamic evolution trends, the cross-interference among different air quality indexes, and the error accumulation in the long-term prediction process are still open problems when establishing air quality forecasting models. Thus, we present a long-term interpretable air quality trend forecasting model to address these challenges via directed interval fuzzy cognitive maps, DE-DIFCM. Specifically, we design a time series trend extraction and representation learning module based on the interval fuzzy granules and the Cramer decomposition theorem in the first phase. Next, we formulate the interval information granules' time series forecasting as a DIFCM. In particular, we employ PM$_{2.5}$as a benchmark to validate the performance of the proposed DE-DIFCM. Experimental results on six air quality monitoring datasets demonstrate the model's superior and competitive long-term prediction performance by comparison with some representative baselines.

Cross-domain learning aims to transfer knowledge learned from one or more datasets to other datasets in different domains, so that less data will be required for learning in new tasks and datasets. One big challenge in cross-domain learning is to effectively synergize the knowledge learning between domains. In this paper, we propose a new solution to address this challenge using normalizing flow, named as DomainFlow, which works as a learned mapping to establish knowledge sharing between source and target domains. The learned flow encourages the posterior distributions in multi-domain learning to be better aligned, leading to better performance in the target domain tasks. We conduct extensive experiments on three representative cross-domain learning tasks: unsupervised domain adaptation, domain generalization and zero-shot sketch-based image retrieval, which demonstrates that with DomainFlow, the overall performance on these diverse tasks can all be improved.

Modern deep neural network models have made remarkable success in many areas, supported by large sets of training samples. Yet the hunger for huge data has also become fatal in further expanding the use of deep models. Limited sample learning aims at learning generalized and transferable representations, without requiring large training data. In this paper, we propose FlowReg, a new learnable latent space regularization for limited sample problems. FlowReg modulates the latent space using a Normalizing Flow with a simple prior (such as Gaussian) while maintaining the complexity of the posterior distribution. We conduct thorough experiments on diverse tasks in limited label learning, as well as detailed in-depth analysis to comprehensively demonstrate the effectiveness of FlowReg.

Sign language is the comprehensive medium of mass communication for hearing and speaking impaired individuals. As they cannot speak or hear, they are not able to use sound or vocal signals as an information medium for their communication. Rather, they are bound to exchange visual signals to express their feeling in their day-to-day life. For this, they use various body language mainly hand gestures as sign language. Sign language fundamentals can be largely divided into two parts namely digits (numerals) and characters (alphabetical). In this paper, we proposed a hybrid model consisting of a deep transfer learning-based convolutional neural network with a random forest classifier for the automatic recognition of Bangla Sign Language (numerals and alphabets). The overall performance of the presented system is verified on ‘Ishara-Bochon’ and ‘Ishara-Lipi’ datasets. ‘Ishara-Bochon’ and ‘Ishara-Lipi’ are datasets of isolated numerals and alphabets respectively which are the first complete multipurpose open-access dataset for Bangla Sign Language (BSL). Besides, we also proposed a background elimination algorithm that removes unnecessary features from the sign images. Along with the proposed background elimination technique, the system is able to achieve accuracy, precision, recall, f1-score values of 91.67%, 93.64%, 91.67%, 91.47% for character recognition and 97.33%, 97.89%, 97.33%, 97.37% for digit recognition respectively. The detailed experimental analysis assures the feasibility and effectiveness of the proposed system for BSL recognition.

Linear discriminant analysis (LDA) is a powerful supervised dimensionality reduction method for analysing high-dimensional data. However, LDA cannot use locality information in data, which makes LDA degrade dramatically in performance on multimodal data. A number of LDA variants have been proposed to exploit locality information in data, including subclass-based LDAs. We discover a problem with these variants, which is that subclasses are selected on a within-class basis without considering other classes. This causes the loss of important information at class boundaries. In this paper, we present a novel variant of subclass-based LDA, Global Subclass Discriminant Analysis (GSDA). Unlike other subclass-based LDAs, GSDA selects subclasses from global clusters that may cross class boundaries, thus utilising within-class information and between-class information. More specifically, GSDA applies an effective clustering algorithm to the whole data to construct global clusters. It then utilises the local structure refining strategy on these global clusters to construct subclasses. Finally, GSDA learns a representative data subspace by maximising inter-subclass distance and minimising intra-subclass distance simultaneously. GSDA is extensively evaluated on a wide range of public datasets through comparison with the state-of-the-art LDA algorithms. Experimental results demonstrate its superiority in terms of accuracy and run times.

Granularrules have been extensively used for classification in fuzzy datasets to promote the advancement of artificial intelligence. However, due to the diversity of data types, how to improve the readability of the extracted granular rules while ensuring efficiency is always a challenge. Since granular reduct in granular computing (GrC) can simplify real complex problem and dataset, this article carries out granular rule learning from the perspective of granular reduct by taking formal concept analysis (FCA)-based GrC method as a framework. Specifically, for achieving classification task, we first propose a method to update the granular reduct, and then explore the updating mechanism of fuzzy granular rule in a reduced dataset. Second, a novel fuzzy rule-based classification model named FRCM is presented for fuzzy granular rule learning. In order to verify the effectiveness of the proposed model, some numerical experiments for incremental learning and fuzzy rule mining are conducted to demonstrate that FRCM can achieve the state-of-the-art classification performance.

Abstract Images in video may include both Gaussian noise and geometric rotation. Thus, it is challenging to represent an image sequence in its intrinsically low‐dimensional space in a noise‐robust and rotation‐robust manner. In this paper, we propose a novel‐ordered sparse subspace clustering algorithm based on a p‐norm to achieve an effective clustering of sequential data under heavy noise conditions. We also use the wavelet‐histogram of oriented gradient (HOG) transform in the kernel view to extract both the global features (with the wavelet process) and the local features (with the HOG process) from the image. In addition, we assign different weights to different features to obtain a sparse coefficient matrix that helps to emphasize the global and local correlations in each sample. Similarly, the clustering algorithm based on the p‐norm for sequential images emphasizes the within‐class correlations amongst samples. Therefore, in this paper, we select additional denoising main components under a Laplacian constraint to achieve a better block‐diagonal structure and highlight the independence of different clusters. Extensive experiments performed on various public datasets (including the ordered face dataset, handwritten recognition dataset, video scene segmentation dataset, and object recognition dataset) demonstrate that the proposed method is more resilient to noise and rotation than other representative sparse subspace clustering methods.

Current hashing-based image retrieval methods mostly assume that the database of images is static. However, this assumption is not true in cases where the databases are constantly updated (e.g., on the Internet) and there exists the problem of concept drift. The online (also known as incremental) hashing methods have been proposed recently for image retrieval where the database is not static. However, they have not considered the concept drift problem. Moreover, they update hash functions dynamically by generating new hash codes for all accumulated data over time which is clearly uneconomical. In order to solve these two problems, concept preserving hashing (CPH) is proposed. In contrast to the existing methods, CPH preserves the original concept, that is, the set of hash codes representing a concept is preserved over time, by learning a new set of hash functions to yield the same set of hash codes for images (old and new) of a concept. The objective function of CPH learning consists of three components: 1) isomorphic similarity; 2) hash codes partition balancing; and 3) heterogeneous similarity fitness. The experimental results on 11 concept drift scenarios show that CPH yields better retrieval precisions than the existing methods and does not need to update hash codes of previously stored images.

Generating action proposals in untrimmed videos is a challenging task, since video sequences usually contain lots of irrelevant contents and the duration of an action instance is arbitrary. The quality of action proposals is key to action detection performance. The previous methods mainly rely on sliding windows or anchor boxes to cover all ground-truth actions, but this is infeasible and computationally inefficient. To this end, this article proposes a RecapNet-a novel framework for generating action proposal, by mimicking the human cognitive process of understanding video content. Specifically, this RecapNet includes a residual causal convolution module to build a short memory of the past events, based on which the joint probability actionness density ranking mechanism is designed to retrieve the action proposals. The RecapNet can handle videos with arbitrary length and more important, a video sequence will need to be processed only in one single pass in order to generate all action proposals. The experiments show that the proposed RecapNet outperforms the state of the art under all metrics on the benchmark THUMOS14 and ActivityNet-1.3 datasets. The code is available publicly at https://github.com/tianwangbuaa/RecapNet.

Concept drift is prevalent in non-stationary data environments but is rarely researched in image retrieval. Therefore, more research is needed on image retrieval in non-stationary data environments so that highly relevant images can still be retrieved when concept drifts happen. Hashing is a key technique to allow efficient image retrieval, so incremental hashing technique emerges in recent years for image retrieval in non-stationary environments. A state-of-the-art method isIncremental Hashing(ICH). ICH trains new hash tables on new data without considering the performance of previous hash tables, so the dependency of successive hash tables is ignored. To make use of this dependency in order to improve the performance of image retrieval in non-stationary environments,Complementary Incremental Hashing with query-adaptive Re-ranking(CIHR) is proposed in this paper. CIHR trains multiple hash tables incrementally, one for each data chunk of images. A new hash table is trained on a new data chunk of images as well as those images badly hashed by previous hash tables, thus the new hash table is complementary to the previous hash tables. To use the hash tables more effectively, a query-adaptive re-ranking method is used to weight all hash functions in each hash table according to their retrieval performance with respect to a given query. Weighted Hamming distance is finally used to evaluate the similarity between the query and the images in the database, as the basis of image retrieval. Experimental results on simulated non-stationary scenarios show that the proposed CIHR method achieves higher retrieval accuracy than all methods being compared, thus setting a new state of the art in image retrieval in non-stationary data environments.

Abstract In many real-world classification problems there exist multiple subclasses (or clusters) within a class; in other words, the underlying data distribution is within-class multimodal. One example is face recognition where a face (i.e. a class) may be presented in frontal view or side view, corresponding to different modalities. This issue has been largely ignored in the literature or at least under studied. How to address the within-class multimodality issue is still an unsolved problem. In this paper, we present an extensive study of within-class multimodality classification. This study is guided by a number of research questions, and conducted through experimentation on artificial data and real data. In addition, we establish a case for within-class multimodal classification that is characterised by the concurrent maximisation of between-class separation, between-subclass separation and within-class compactness. Extensive experimental results show that within-class multimodal classification consistently leads to significant performance gains when within-class multimodality is present in data. Furthermore, it has been found that within-class multimodal classification offers a competitive solution to face recognition under different lighting and face pose conditions. It is our opinion that the case for within-class multimodal classification is established, therefore there is a milestone to be achieved in some machine learning algorithms (e.g. Gaussian mixture model) when within-class multimodal classification, or part of it, is pursued.

We present an extension of our previous work in [1] by investigating the use of Local Septenary Patterns (LSP) for breast density classification in mammograms. The LSP operator is a variant of Local Binary Patterns (LBP) inspired by Local Ternary Patterns (LTP) and Local Quinary patterns (LQP). The main extensions in our work are i) we investigate the use of a multi-resolution technique when extracting micro texture information, ii) we investigate different neighbourhood topologies as different ways of extracting texture features, and iii) we use an additional dataset called InBreast as well as the most popular dataset in the literature, which is the Mammographic Image Analysis Society (MIAS) to further evaluate the performance of the LSP operator.

Loss function plays an important role in CNNs. However, the recent loss functions either do not apply weight and feature normalisation or do not explicitly follow the two targets of improving discriminative ability: minimising intra-class variance and maximising inter-class variance. Besides, all of them consider only the feedback information from the current mini-batch instead of the distribution information from the whole training set. In this paper, we propose a novel loss function - Global Information-based Cosine Optimal loss (Gico loss). Gico loss is applied with weight and feature normalisation, designed explicitly following the aforementioned two targets of improving discriminative ability, and is guided by the distribution information from the whole training set. Extensive experiments are conducted on multiple public datasets, which confirms the effectiveness of the proposed Gico loss and shows that we achieve state-of-the-art performance.

Softmax loss is arguably one of the most widely used loss functions in CNNs. In recent years some Softmax variants have been proposed to enhance the discriminative ability of the learned features by adding additional margin constraints, which significantly improved the state-of-the-art performance of face recognition. However, the `margin' referenced in these losses does not represent the real margin between the different classes in the training set. Furthermore, they impose a margin on all possible combinations of class pairs, which is unnecessary. In this paper we propose the Precise Adjacent Margin loss (PAM loss), which gives an accurate definition of `margin' and has precise operations appropriate for different cases. PAM loss has better geometrical interpretation than the existing margin-based losses. Extensive experiments are conducted on LFW, YTF, MegaFace and FaceScrub datasets, and results show that the proposed method has state-of-the-art performance.

Text matching is important for a variety of natural language processing tasks, such as paraphrase identification and natural language inference. Recent studies have achieved very promising results under the compare-aggregate framework. A limitation of previous approaches following this framework is that they solely conduct matching at word level. In this paper, we propose a multi-level compare-aggregate model (MLCA), which matches each word in one text against the other text at three different levels, word level (word-by-word matching), phrase level (word-by-phrase matching) and sentence level (word-bysentence matching). Then the results of these levels of matching are aggregated for making final matching decision. We evaluate our model on two different tasks: paraphrase identification and natural language inference. Experimental results show that our model achieves the state-of-the-art performance on both tasks.

Text matching aims to establish the matching relationship between two texts. It is an important operation in some information retrieval related tasks such as question duplicate detection, question answering, and dialog systems. Bidirectional long short term memory (BiLSTM) coupled with attention mechanism has achieved state-of-the-art performance in text matching. A major limitation of existing works is that only high level contextualized word representations are utilized to obtain word level matching results without considering other levels of word representations, thus resulting in incorrect matching decisions for cases where two words with different meanings are very close in high level contextualized word representation space. Therefore, instead of making decisions utilizing single level word representations, a multi-level matching network (MMN) is proposed in this paper for text matching, which utilizes multiple levels of word representations to obtain multiple word level matching results for final text level matching decision. Experimental results on two widely used benchmarks, SNLI and Scaitail, show that the proposed MMN achieves the state-of-the-art performance.

Most of current hashing methods are proposed based on the assumption that the database is stationary. However, this assumption is not always true as the data environment is sometimes non-stationary. When new images being added to the database, data distributions of existing classes may change and new classes may also appear which result in concept drifts. The problem of concept drifts is unavoidable in non-stationary data environments. Incremental Hashing (ICH) is an effective method for image retrieval in non-stationary data environments with concept drifts using multiple hash tables. In ICH, new concept is adapted by training new hash table using the most updated data chunks. However, images in the new data chunk may not be all informative for updating. To enhance the efficiency of ICH, ICH with Undersampling (ICHUS) is proposed to select informative samples in the new data chunk for the training of new hash table to adapt to the non-stationary data environment. Experimental results show that ICHUS yields a better retrieval performance than ICH and state-of-art non-stationary hashing methods.

Gaussian Mixture Model (GMM) is a probabilistic model for representing normally distributed subpopulations within an overall population. It is usually used for unsupervised learning to learn the subpopulations and the subpopulation assignment automatically. It is also used for supervised learning or classification to learn the boundary of subpopulations. However, the performance of GMM as a classifier is not impressive compared with other conventional classifiers such as k-nearest neighbors (KNN), support vector machine (SVM), decision tree and naive Bayes. In this paper, we attempt to address this problem. We propose a GMM classifier, SC-GMM, based on the separability criterion in order to separate the Gaussian models as much as possible. This classifier finds the optimal number of Gaussian components for each class based on the separability criterion and then determines the parameters of these Gaussian components by using the expectation maximization algorithm. Extensive experiments have been carried out on classification tasks from general data mining to face verification. Results show that SC-GMM significantly outperforms the original GMM classifier. Results also show that SC-GMM is comparable in classification accuracy to three variants of GMM classifier: Akaike Information Criterion based GMM (AIC-GMM), Bayesian Information Criterion based GMM (BIC-GMM) and variational Bayesian gaussian mixture (VBGM). However, SC-GMM is significantly more efficient than both AIC-GMM and BIC-GMM. Furthermore, compared with KNN, SVM, decision tree and naive Bayes, SC-GMM achieves competitive classification performance.

The performance of a brain–computer interface (BCI) will generally improve by increasing the volume of training data on which it is trained. However, a classifier’s generalization ability is often negatively affected when highly non-stationary data are collected across both sessions and subjects. The aim of this work is to reduce the long calibration time in BCI systems by proposing a transfer learning model which can be used for evaluating unseen single trials for a subject without the need for training session data. A method is proposed which combines a generalization of the previously proposed subject-specific “multivariate empirical-mode decomposition” preprocessing technique by taking a fixed band of 8–30[Formula: see text]Hz for all four motor imagery tasks and a novel classification model which exploits the structure of tangent space features drawn from the Riemannian geometry framework, that is shared among the training data of multiple sessions and subjects. Results demonstrate comparable performance improvement across multiple subjects without subject-specific calibration, when compared with other state-of-the-art techniques.

Fuzzy similarity relation is a function to measure the similarity between two samples. It is widely used to learn knowledge under the framework of fuzzy machine learning. The selection of a suitable fuzzy similarity relation is important for the learning task. It has been pointed out that fuzzy similarity relations can be brought into the framework of kernel functions in machine learning. This fact motivates us to study fuzzy similarity relation selection for fuzzy machine learning utilizing kernel selection methods in machine learning. Kernel alignment is a kernel selection method that is effective and has low computational complexity. In this paper, we present novel methods for fuzzy similarity relation selection based on the kernel alignment, and their use in attribution reduction for heterogeneous data. First, we define an ideal kernel for classification problems, based on which a novel fuzzy kernel alignment model is proposed. Second, we present a method for the fuzzy similarity relation selection based on the minimization of the fuzzy alignment between the defined ideal kernel and a kernel for the learning problem at hand. In order to show the correctness of this selection method, we prove that the lower bound of the classification accuracy of a support vector machine will increase with the decrease of the fuzzy alignment value. Furthermore, we apply the proposed fuzzy similarity relation selection to attribute reduction for heterogeneous data. Finally, we present experimental results to show that the proposed method of fuzzy similarity relation selection based on the fuzzy kernel alignment is effective.

The paper presents quantitative results of a preliminary study undertaken as part of Decision Support and Information Management System for Breast Cancer (DESIREE). DESIREE is a European-funded project to improve the management of primary breast cancer through image-based, guideline-based, experience-based, and case-based information systems. In this study we explore the use of ensemble deep learning for breast mass classification in mammograms. The proposed method is based on AlexNet with some modifications in order to adapt it to our classification problem. Subsequently, model selection is performed to select the best three results based on the highest validation accuracies during the validation phase. Finally, the prediction is based on the average probability of the models. Experimental evaluation shows that accuracy from individual models ranges between 75% and 77%, but combining the best models (ensemble networks) results in over 80% classification accuracy and aura under the curve.

Breast cancer is the most common type of cancer in women worldwide, with incidence rate being second highest to other types of cancer. In the current clinical setting, multidisciplinary breast units are introduced to improve the quality of the therapeutic decision based on the best evidence-based practices. DESIREE project aims to provide a web-based software ecosystem for personalized, collaborative and multidisciplinary management of primary breast cancer by specialized breast units, from diagnosis to therapy and follow-ups. In order to provide a multi-model decision support to clinicians in present clinical settings, the project develops and integrates three modalities of decision support, namely guideline-based decision support system (DSS), experience-based DSS and case-based DSS. Visual analytics GUI are developed to properly adapt the results of the DSSs and graphically represent them to the clinician in a user-friendly manner. DESIREE information management system (DESIMS), serves as the interface between the user and DSSs for entering patient data and viewing the results in the visual analytics GUI. In this paper, we present the overall architecture, workflow and integration of the three DSSs in the DESIREE platform.

Computer-aided diagnosis (CAD) has great potential in providing real benefits to doctors and patients. Recent studies have, however, found lack of trust in CAD by radiologists in clinical diagnostic decision making. One of the main reasons is the lack of an appropriate confidence measure. This paper presents the first-ever study of classification confidence in the context of breast mass classification. We evaluated 11 state-of-the-art classification algorithms on breast mass image data using their confidence of classification metric, in addition to other standard evaluation metrics including accuracy and area under the curve (ROC). Experimental results show that although most classifiers produced very similar results with less than 2% difference in terms of accuracy and ROC, their performances are significantly different in terms of confidence levels. We suggest that the confidence measure should be used in conjunction with the existing performance metrics such as accuracy and ROC.

Most of the existing hashing methods for image retrieval are based on the assumption the image database is stationary. However, in the real world data environments are always changing or non-stationary, therefore the underlying data distribution may change from time to time which will result in the problem of concept drift. Incremental Hashing (ICH) is the only existing method to handle image retrieval with concept drift in non-stationary data environments. It builds hash codes for the database through increments. At each increment, a set of new hash functions is built with the new chunk of data, which is utilized to update the multi-hashing system to generate multiple sets of hash codes for all data. However, only the newest data chunk is used to train individual hash functions, while the semantic similarity information of previous data is missed. In this paper, we present a new hashing method based on ICH for image retrieval with concept drift, Incremental Hashing with Dynamic Semantic Pool (ICH-DSP). It builds a semantic pool to collect representative labeled data for each existing class. The semantic pool is updated incrementally and is used as the supervisory information for the training of hash functions. Experimental results on three real world image databases show that ICH-DSP outperforms the original ICH and other state-of-the-art hashing methods.

Linear discriminant analysis (LDA) is a classical method for discriminative dimensionality reduction. The original LDA may degrade in its performance for non-Gaussian data, and may be unable to extract sufficient features to satisfactorily explain the data when the number of classes is small. Two prominent extensions to address these problems are subclass discriminant analysis (SDA) and mixture subclass discriminant analysis (MSDA). They divide every class into subclasses and re-define the within-class and between-class scatter matrices on the basis of subclass. In this paper we study the issue of how to obtain subclasses more effectively in order to achieve higher class separation. We observe that there is significant overlap between models of the subclasses, which we hypothesise is undesirable. In order to reduce their overlap we propose an extension of LDA, separability oriented subclass discriminant analysis (SSDA), which employs hierarchical clustering to divide a class into subclasses using a separability oriented criterion, before applying LDA optimisation using re-defined scatter matrices. Extensive experiments have shown that SSDA has better performance than LDA, SDA and MSDA in most cases. Additional experiments have further shown that SSDA can project data into LDA space that has higher class separation than LDA, SDA and MSDA in most cases.

Feature selection based on fuzzy rough sets is an effective approach to select a compact feature subset that optimally predicts a given decision label. Despite being studied extensively, most existing methods of fuzzy rough set based feature selection are restricted to computing the whole dataset in batch, which is often costly or even intractable for large datasets. To improve the time efficiency, we investigate the incremental perspective for fuzzy rough set based feature selection assuming data can be presented in sample subsets one after another. The key challenge for the incremental perspective is how to add and delete features with the subsequent arrival of sample subsets. We tackle this challenge with strategies of adding and deleting features based on the relative discernibility relations that are updated as subsets arrive sequentially. Two incremental algorithms for fuzzy rough set based feature selection are designed based on the strategies. One updates the selected features as each sample subset arrives, and outputs the final feature subset where no sample subset is left. The other updates the relative discernibility relations but only performs feature selection where there is no further subset arriving. Experimental comparisons suggest our incremental algorithms expedite fuzzy rough set based feature selection without compromising performance.

Palmprint recognition has received in the last 20 years a great deal of the research community's attention. In this paper a new palmprint matching approach based on corner feature point extraction is proposed. A 72-element fixed-length descriptor is used to capture distinctive information of each feature point neighborhood and to build a measure of similarity whilst their coordinates provide a measure of proximity between the points. Matching two images takes into account both similarity and proximity measures which converts into a cost minimization problem. Our experiments carried out on a database of 250 prints from the Poly U database have yielded very good results evidenced by an EER of 0.31%.

Estimating the initial background of a scene is a key prerequisite for several applications in video analytics. In this paper, we present a simple approach that takes into account spatio-temporal motion intensities while estimating the true background. We tested the algorithm on real video sequences from the Scene Background Initialization (SBI) benchmark dataset, and the results show that the algorithm is competitive compared to the state of the art.

Pupillary behaviour and dilation have been considered in the literature as an effective input for the measurement of cognitive workload and stress. In this work, we explore the correlation between pupil dilation and features extracted from low quality video frames that have been captured using a normal webcam during a set of computer-based tasks. The methodology presented herein attempts to develop an alternative, cost effective technique for the representation of pupil dilation in order to track pupillary behaviour from images instead of employing specialised, high-cost eye-tracking devices, which typically require specialist expertise during setup and calibration. A description of the data collection protocol and subsequent data analysis is presented. The results obtained indicate that there is a moderate correlation achieved through the use of a linear regression model, which employs fiducial point features as independent variables, and pupil size measured by an infrared-based eye-tracker as the dependent variable. Furthermore, an example of the pupil size variation within a game-based task context is shown, whereby one can easily relate the engagement and the amount of mental processing during gameplay.

One of the challenges in virtual environments is the difficulty users have in interacting with these increasingly complex systems. Ultimately, endowing machines with the ability to perceive users emotions will enable a more intuitive and reliable interaction. Consequently, using the electroencephalogram as a bio-signal sensor, the affective state of a user can be modelled and subsequently utilised in order to achieve a system that can recognise and react to the user’s emotions. This paper investigates features extracted from electroencephalogram signals for the purpose of affective state modelling based on Russell’s Circumplex Model. Investigations are presented that aim to provide the foundation for future work in modelling user affect to enhance interaction experience in virtual environments. The DEAP dataset was used within this work, along with a Support Vector Machine and Random Forest, which yielded reasonable classification accuracies for Valence and Arousal using feature vectors based on statistical measurements and band power from the α , β , δ , and 𝜃 waves and High Order Crossing of the EEG signal.

Attribute reduction with rough sets is an effective technique for obtaining a compact and informative attribute set from a given dataset. However, traditional algorithms have no explicit provision for handling dynamic datasets where data present themselves in successive samples. Incremental algorithms for attribute reduction with rough sets have been recently introduced to handle dynamic datasets with large samples, though they have high complexity in time and space. To address the time/space complexity issue of the algorithms, this paper presents a novel incremental algorithm for attribute reduction with rough sets based on the adoption of an active sample selection process and an insight into the attribute reduction process. This algorithm first decides whether each incoming sample is useful with respect to the current dataset by the active sample selection process. A useless sample is discarded while a useful sample is selected to update a reduct. At the arrival of a useful sample, the attribute reduction process is then employed to guide how to add and/or delete attributes in the current reduct. The two processes thus constitute the theoretical framework of our algorithm. The proposed algorithm is finally experimentally shown to be efficient in time and space.

Due to the increase in electronic documents, automatic text classification has gained a lot of importance as manual classification of documents is time-consuming. Machine learning is the main approach for automatic text classification, where texts are represented, terms are weighted on the basis of the chosen representation and a classification model is built. Vector space model is the dominant text representation largely due to its simplicity. Graphs are becoming an alternative text representation that have the ability to capture important information in text such as term order, term co-occurrence and term relationships that are not considered by the vector space model. Substantially better text classification performance has been demonstrated for term weighting schemes which use a graph representation. In this paper, we introduce a graph-based term weighting scheme, tw-srw, which is an effective supervised term weighting method that considers the co-occurrence information in text for increasing text classification accuracy. Experimental results show that it outperforms the state-of-the-art unsupervised term weighting schemes.

In previous work we have presented a class of algebras enhanced with two contact relations representing rough set style approximations of a spatial contact relation. In this paper, we develop a class of relational systems which is mutually interpretable with that class of algebras, and we consider a relational logic whose semantics is determined by those relational systems. For this relational logic we construct a proof system in the spirit of Rasiowa-Sikorski, and we outline the proofs of its soundness and completeness.

As digital technologies advance, video has become ubiquitous and hence a rich source of information. Video analytics (or video content analysis) is an important area of computer vision which is concerned with the process of making sense of video content in order to ultimately understand video. Video analytics appears in different forms, such as activity recognition, motion detection, object detection and recognition, person detection and recognition, event and scenario recognition, anomaly detection and identity recognition and verification. Video analytics can be applied in a wide range of domains including healthcare, retail, transport, smart homes, safety and security. The aim of this Special Issue is to raise the awareness of the importance of video analytics. The specific objectives are: (1) to report the latest developments; (2) to identify major research challenges and; (3) to provide visions of future development. A total of 21 papers were submitted to this Special Issue including invited papers and, following a rigorous peer-review process, a total of 11 papers were accepted for inclusion in this Special Issue. The invited paper “Video Analytics Revisited” by Ayesha Choudhary and Santanu Chaudhury presents a concise yet in-depth survey of video analytics. Research problems are discussed and important current applications of video analytics are reviewed. The paper “Human Action Recognition Using Histogram of Motion Intensity and Direction from Multiple Views” by SungYong Chun et al. presents an approach to human activity recognition from multiple views based on estimation of local motion from multiple camera views. A new motion descriptor, histogram of motion intensity and direction, is proposed to capture local motion characteristics of human activity. Classification is done using a support vector machine. Experimental evaluation has demonstrated superior performance of their approach, outperforming 3D optical flow-based approaches with lower computational requirements. The paper “Video Anomaly Detection Using Deep Incremental Slow Feature Analysis Network” by Xing Hu et al. presents an approach to anomaly detection using automatically learned features instead of hand-crafted features. A Deep Incremental Slow Feature Analysis (D-IncSFA) network is proposed, which learns progressively abstract and global high-level features from raw data. The D-IncSFA network has the functionalities of both feature extractor and anomaly detector so anomaly detection can be completed in one step. Their approach can detect global anomaly such as crowd panic and local anomaly and is intended to be universal in order to work in different scenarios, with little human intervention and low memory and computational requirements. The paper “Multiple Deep Features Learning for Object Retrieval in Surveillance Videos” by Haiyun Guo et al. aims to address the challenge of efficiently indexing and retrieving objects of interest from large-scale surveillance videos. A multiple deep features learning approach to object retrieval in surveillance videos is proposed, which is based on the discriminative convolutional neural network (CNN). The CNN model is pre-trained on ImageNet ILSVRC12 and then fine-tuned on their dataset. To improve the retrieval performance, the deep features are encoded into short binary codes by Locality-Sensitive Hash and fused to retrieve the object of interest. Experiments on a dataset of 100k objects extracted from multi-camera surveillance videos have demonstrated good performance of the proposed approach, compared with other common visual features. The paper “A Two-layer Discriminative Model for Human Activity Recognition” by Mouna Selmi et al. studies human activity recognition with a focus on the role of local interest point features like spatio-temporal interest points. This paper presents a new approach that explicitly models the sequential aspect of activities. A support vector machine provides a vector of conditional class probabilities for each window that summarises all discriminant information that is relevant for sequence recognition. The sequence of these stochastic vectors is then fed to a hidden conditional random field for inference at the sequence level. Experiments on various human activity datasets have demonstrated that the proposed approach compares favourably with current state-of-the-art. The paper “A New Fusional Framework Combining Sparse Selection and Clustering for Key Frame Extraction” by Mengjuan Fei et al. studies key frame extraction, a type of video summarisation, which facilitates rapid browsing and efficient video indexing. This paper proposes a syncretic key frame extraction framework (SS-MIAHC) that combines sparse selection and mutual information-based agglomerative hierarchical clustering (MIAHC) to generate effective video summaries. The proposed framework overcomes issues such as information redundancy and computational complexity. The experiments conducted on two benchmark datasets demonstrate that the proposed SS-MIAHC framework is superior to conventional methods. The paper “Multi-Object Tracking using Dominant Sets” by Yonatan T. Tesfaye et al. studies multi-object tracking and addresses the challenges of identity switches and difficulties in handling long-term occlusions by formulating the tracking task as a problem of finding dominant sets in an auxiliary edge weighted graph. This is a novel approach to multi-object tracking, which has been demonstrated to have superior performance compared with several state of the art methods in experiments on three different challenging datasets. The paper “Contextualized Learning-free 3D Body Pose Estimation from 2D Body Features in Monocular Images” by Luis Unzueta et al. presents a method for 3D human body pose estimation from a monocular camera based on a learning-free hierarchical optimisation procedure and contextual information. This approach explicitly considers and preserves the relations between the 3D subject's overall scale; its depth with respect to the camera; and its configuration related to the reference floor. Thus, it can obtain more coherent reconstructions with respect to the shared 3D world, compared to other state-of-the-art approaches, efficiently and without the need for learning 2D/3D mapping models from training data. Therefore, it is not affected by data characteristic differences between training and deployment stages. The paper “‘Owl’ and ‘Lizard’: Patterns of Head Pose and Eye Pose in Driver Gaze Classification” by Lex Fridman et al. studies gaze tracking in the car through estimating head pose and eye pose from monocular video. New research questions are asked, which are answered by evaluating data drawn from an on-road study of 40 drivers. The main insight of the paper is conveyed through the analogy of an “owl” and “lizard” which describes the degree to which the eyes and the head move when shifting gaze. When the head moves a lot (“owl”), not much classification improvement is attained by estimating eye pose on top of head pose. On the other hand, when the head stays still and only the eyes move (“lizard”), classification accuracy increases significantly from adding in eye pose. The paper “Forensic Video Solution Using Facial Feature Based Synoptic Video Footage Record” by B.Yogameena et al. proposes a solution to identify a specific person quickly which is valuable in analysing incidents/crimes. The main idea of this paper is to reduce the enormous volume of video data by using an object based video synopsis. SVM is used to classify the weak and strong features. These strong features are used to recognise the person. The algorithm works well even in complicated situations such as expression changes, pose, illumination variations and even if the face is partially or fully occluded in few frames. The advantage of synoptic video helps to recognize the person who is not occluded in some other frames. Experimental results on benchmark and real time datasets demonstrate the effectiveness of the proposed algorithm. The paper “Facial Video based Detection of Physical Fatigue for Maximal Muscle Activity” by Mohammad A. Haque et al. studies video based detection of physical fatigue. This paper presents an efficient noncontact system for detecting non-localised physical fatigue from maximal muscle activity using facial videos acquired in a realistic environment with natural lighting where subjects were allowed to voluntarily move their head, change their facial expression and vary their pose. Experimental results show that the proposed system outperforms video based existing system for physical fatigue detection.

Spatial information describes the relative spatial position of an object in a video.Such information may aid several video analysis tasks such as object, scene, event and activity recognition.This paper studies the effect of spatial information on video activity recognition.The paper firstly performs activity recognition on KTH and Weizmann videos using Hidden Markov Model and k-Nearest Neighbour classifiers trained on Histogram Of Oriented Optical Flows feature.Histogram of Oriented Optical Flows feature is based on optical flow vectors and ignores any spatial information present in a video.Further, in this paper, a new feature set, referred to as Regional Motion Vectors is proposed.This feature like Histogram of Oriented Optical Flow is derived from optical flow vectors; however, unlike Histogram of Oriented Optical Flows preserves any spatial information in a video.Activity recognition was again performed using the two classifiers, this time trained on Regional Motion Vectors feature.Results show that when Regional Motion Vectors is used as the feature set on the KTH dataset, there is a significant improvement in the performance of k-Nearest Neighbour.When Regional Motion Vector is used on the Weizmann dataset, performances of the k-Nearest Neighbour improves significantly for some of the cases and for the other cases, the performance is comparable to when oriented optical flows is used as a feature set.Slight improvement is achieved by Hidden Markov Model on both the datasets.As Histogram of Oriented Optical Flows ignores spatial information and Regional Motion Vectors preserves it, the increase in the performance of the classifiers on using Reginal Motion Vectors instead of Histogram of Oriented Optical Flows illustrates the importance of spatial information in video activity recognition.

In this paper, we present a new filtering method based on the empirical mode decomposition (EMD) for classification of motor imagery (MI) electroencephalogram (EEG) signals for enhancing brain-computer interface (BCI). The EMD method decomposes EEG signals into a set of intrinsic mode functions (IMFs). These IMFs can be considered narrow-band, amplitude and frequency modulated (AM-FM) signals. The mean frequency measure of these IMFs has been used to combine these IMFs in order to obtain the enhanced EEG signals which have major contributions due to μ and β rhythms. The main aim of the proposed method is to filter EEG signals before feature extraction and classification to enhance the features separability and ultimately the BCI task classification performance. The features namely, Hjorth and band power features computed from the enhanced EEG signals, have been used as a feature set for classification of left hand and right hand MIs using a linear discriminant analysis (LDA) based classification method. Significant superior performance is obtained when the method is tested on the BCI competition IV datasets, which demonstrates the effectiveness of the proposed method.

Systematic identification of protein complexes from protein-protein interaction networks (PPIs) is an important application of data mining in life science. Over the past decades, various new clustering techniques have been developed based on modelling PPIs as binary relations. Non-binary information of co-complex relations (prey/bait) in PPIs data derived from tandem affinity purification/mass spectrometry (TAP-MS) experiments has been unfairly disregarded. In this paper, we propose a Biased Random Walk based algorithm for detecting protein complexes from TAP-MS data, resulting in the random walk with restarting baits (RWRB). RWRB is developed based on Random walk with restart. The main contribution of RWRB is the incorporation of co-complex relations in TAP-MS PPI networks into the clustering process, by implementing a new restarting strategy during the process of random walk. Through experimentation on un-weighted and weighted TAP-MS data sets, we validated biological significance of our results by mapping them to manually curated complexes. Results showed that, by incorporating non-binary, co-membership information, significant improvement has been achieved in terms of both statistical measurements and biological relevance. Better accuracy demonstrates that the proposed method outperformed several state-of-the-art clustering algorithms for the detection of protein complexes in TAP-MS data.

Data incompleteness and inconsistency are common issues in data-driven decision models. To some extend, they can be considered as two opposite circumstances, since the former occurs due to lack of information and the latter can be regarded as an excess of heterogeneous information. Although these issues often contribute to a decrease in the accuracy of the model, most modeling approaches lack of mechanisms to address them. This research focuses on an advanced belief rule-based decision model and proposes a dynamic rule activation (DRA) method to address both issues simultaneously. DRA is based on “smart” rule activation, where the actived rules are selected in a dynamic way to search for a balance between the incompleteness and inconsistency in the rule-base generated from sample data to achive a better performance. A series of case studies demonstrate how the use of DRA improves the accuracy of this advanced rule-based decision model, without compromising its efficiency, especially when dealing with multi-class classification datasets. DRA has been proved to be beneficial to select the most suitable rules or data instances instead of aggregating an entire rule-base. Beside the work performed in rule-based systems, DRA alone can be regarded as a generic dynamic similarity measurement that can be applied in different domains.

The KTH video dataset [1] contains three activities - walking, jogging and running - which are very similar but are carried out at a different natural pace. We show that explicit inclusion of a feature which may be interpreted as a measure of the overall state of motion in a frame improves a classifier's ability to discriminate between these activities.

This demonstration shows the integration of video analysis and search tools to facilitate the interactive retrieval of video segments depicting specific activities from surveillance footage. The implementation was developed by members of the SAVASA project for participation in the interactive surveillance event detection (SED) task of TRECVid 2012. This year, for the first time, the purpose of the interactive SED task was to evaluate systems' ability to support users in identifying video segments that depict a specific activity (event) in a large collection of surveillance video footage. Project partners worked together to analyse video and provide a query interface enabling users to search and identify matching video segments. The collaborative integration of components from multiple partners and the participation of end user partners in evaluating the system are the novel aspects of this work.

This paper presents work on integrating multiple computer vision-based approaches to surveillance video analysis to support user retrieval of video segments showing human activities. Applied computer vision using real-world surveillance video data is an extremely challenging research problem, independently of any information retrieval (IR) issues. Here we describe the issues faced in developing both generic and specific analysis tools and how they were integrated for use in the new TRECVid interactive surveillance event detection task. We present an interaction paradigm and discuss the outcomes from face-to-face end user trials and the resulting feedback on the system from both professionals, who manage surveillance video, and computer vision or machine learning experts. We propose an information retrieval approach to finding events in surveillance video rather than solely relying on traditional annotation using specifically trained classifiers.

Real-world decision problems are usually associated with a certain geographical area, and therefore can and should be geographically referenced in most of cases. While traditional Decision Support Systems (DSSs) ignore the spatial dimension of the problem, most Geographic Information System (GIS)-based Spatial Decision Support Systems (SDSSs) focus mainly on the spatial analysis of the problem, avoiding other relevant factors like uncertainty and incompleteness of data sets. This research is based on a recently developed intelligent belief rule-based DSS, called RIMER+, which is shown to be capable of capturing vagueness, incompleteness, uncertainty, and nonlinear causal relationships in an integrated way. The main contribution of this research is to explore the possibilities of achieving a higher degree of integration of DSSs in a GIS environment, i.e., integration of RIMER+ within GIS system by using an embedded approach, which not only enhances further the capability and applicability of the RIMER+ by integrating the spatial component of the problem into the decision making process, but also takes advantage of the GIS software capabilities in terms of spatial analysis and visualization. Finally, this research employs a comparative case study to demonstrate performance of the proposed Spatial RIMER+ methodology against the well-known Geographically Weighted Regression (GWR) methodology.

Conventional human activity recognition approaches are mainly based on machine learning methods, which are not working well for composite activity recognition due to the complexity and uncertainty of real scenarios. We propose in this paper an automated reasoning based hierarchical framework for human activity recognition. This approach constructs a hierarchical structure for representing the composite activity by a composition of lower-level actions and gestures according to its semantic meaning. This hierarchical structure is then transformed into logical formulas and rules, based on which the resolution based automated reasoning is applied to recognize the composite activity given the recognized lower-level actions by machine learning methods.

In this paper we review Lattice Machine, a learning paradigm that “learns” by generalising data in a consistent, conservative and parsimonious way, and has the advantage of being able to provide additional reliability information for any classification. More specifically, we review the related concepts such as hyper tuple and hyper relation, the three generalising criteria (equilabelledness, maximality, and supportedness) as well as the modelling and classifying algorithms. In an attempt to find a better method for classification in Lattice Machine, we consider the contextual probability which was originally proposed as a measure for approximate reasoning when there is insufficient data. It was later found to be a probability function that has the same classification ability as the data generating probability called primary probability. It was also found to be an alternative way of estimating the primary probability without much model assumption. Consequently, a contextual probability based Bayes classifier can be designed. In this paper we present a new classifier that utilises the Lattice Machine model and generalises the contextual probability based Bayes classifier. We interpret the model as a dense set of data points in the data space and then apply the contextual probability based Bayes classifier. A theorem is presented that allows efficient estimation of the contextual probability based on this interpretation. The proposed classifier is illustrated by examples.

This paper presents a framework for incorporating semantic similarities in the detection of protein complexes from Affinity Purification/Mass Spectrometry (AP-MS) data. AP-MS data is modeled as a bipartite network, where one set of nodes consist of bait proteins and the other set are prey proteins. Pair-wise similarities of bait proteins are computed by combining similarities based on topological features and functional semantic similarities. A hierarchical clustering algorithm is then applied to obtain `seed clusters' consisting of bait proteins. Starting from these `seed' clusters, an expansion process is developed to recruit prey proteins which are significantly associated with bait proteins, to produce final sets of identified protein complexes. In the application to real AP-MS datasets, we validate biological significance of predicted protein complexes by using curated protein complexes. Six statistical metrics have been applied. Results show that by integrating semantic similarities into the clustering process, the accuracy of identifying complexes has been greatly improved. Meanwhile, clustering results obtained by the proposed framework are better than those from several existent clustering methods.

Decision makers are required to make critical decisions throughout all stages of a life-cycle in large-scale projects. These decisions are important as they impact upon the outcome and the success of projects. In this paper we present an evidential reasoning framework to aid decision-makers in the decision making process. This approach utilizes the Dezert-Smarandache Theory (DSm) to fuse heterogeneous evidence sources that suffer from levels of uncertainty, imprecision and conflicts to provide beliefs for decision options. To analyze the impact that source reliability and priority has upon the decision making process a reliability discounting technique along with a priority discounting technique are applied. Application of the evidential reasoning framework is illustrated using a Case Study based in the Aerospace domain.

Hand gesture recognition aims to recognize the meaningful expressions of hand motion. It is widely used in information visualization, robotics, sign language understanding, medicine and healthcare. Some methods have been proposed for hand gesture recognition. But no single algorithm can handle all kinds of situations, because of the complex environment. In this study, we propose a hybrid method for hand gesture recognition, which extends our previous work on a gesture recognition method based on concept learning by the addition of an association learning process. We use association learning to reveal the frequent patterns in gesture sequences, and then use such patterns to help recognize incomplete gesture sequences. Experiments show the use of association learning does indeed improve recognition accuracy. Experiments also show the hybrid method is comparable to two state of the art methods (HMMs and DTW) for hand gesture recognition, but outperforms them in the larger datasets.

Activity modelling is required to support activity recognition and further to provide activity assistance for users in smart homes. Current research in knowledge-driven activity modelling has mainly focused on single activities with little attention being paid to the modelling of composite activities such as interleaved and concurrent activities. This paper presents a hybrid approach to composite activity modelling by combining ontological and temporal knowledge modelling formalisms. Ontological modelling constructors, i.e. concepts and properties for describing composite activities, have been developed and temporal modelling operators have been introduced. As such, the resulting approach is able to model both static and dynamic characteristics of activities. Several composite activity models have been created based on the proposed approach. In addition, a set of inference rules has been provided for use in composite activity recognition. A concurrent meal preparation scenario is used to illustrate both the proposed approach and associated reasoning mechanisms for composite activity recognition.

This paper introduces a knowledge-driven approach to real-time, continuous activity recognition based on multisensor data streams in smart homes. The approach goes beyond the traditional data-centric methods for activity recognition in three ways. First, it makes extensive use of domain knowledge in the life cycle of activity recognition. Second, it uses ontologies for explicit context and activity modeling and representation. Third and finally, it exploits semantic reasoning and classification for activity inferencing, thus enabling both coarse-grained and fine-grained activity recognition. In this paper, we analyze the characteristics of smart homes and Activities of Daily Living (ADL) upon which we built both context and ADL ontologies. We present a generic system architecture for the proposed knowledge-driven approach and describe the underlying ontology-based recognition process. Special emphasis is placed on semantic subsumption reasoning algorithms for activity recognition. The proposed approach has been implemented in a function-rich software system, which was deployed in a smart home research laboratory. We evaluated the proposed approach and the developed system through extensive experiments involving a number of various ADL use scenarios. An average activity recognition rate of 94.44 percent was achieved and the average recognition runtime per recognition operation was measured as 2.5 seconds.

Tree is one of the most common and well-studied data structures in computer science. Measuring the similarity of such structures is key to analyzing this type of data. However, measuring tree similarity is not trivial due to the inherent complexity of trees and the ensuing large search space. Tree kernel, a state of the art similarity measurement of trees, represents trees as vectors in a feature space and measures similarity in this space. When different features are used, different algorithms are required. Tree edit distance is another widely used similarity measurement of trees. It measures similarity through edit operations needed to transform one tree to another. Without any restrictions on edit operations, the computation cost is too high to be applicable to large volume of data. To improve efficiency of tree edit distance, some approximations were introduced into tree edit distance. However, their effectiveness can be compromised. In this paper, a novel approach to measuring tree similarity is presented. Trees are represented as multidimensional sequences and their similarity is measured on the basis of their sequence representations. Multidimensional sequences have their sequential dimensions and spatial dimensions. We measure the sequential similarity by the all common subsequences sequence similarity measurement or the longest common subsequence measurement, and measure the spatial similarity by dynamic time warping. Then we combine them to give a measure of tree similarity. A brute force algorithm to calculate the similarity will have high computational cost. In the spirit of dynamic programming two efficient algorithms are designed for calculating the similarity, which have quadratic time complexity. The new measurements are evaluated in terms of classification accuracy in two popular classifiers (k-nearest neighbor and support vector machine) and in terms of search effectiveness and efficiency in k-nearest neighbor similarity search, using three different data sets from natural language processing and information retrieval. Experimental results show that the new measurements outperform the benchmark measures consistently and significantly.

Taking into account the need of handling hybrid information with uncertainty in human decision making, a new belief rule-base inference methodology (RIMER) has been recently proposed. RIMER approach and its relevant extensions have proved to be highly positive solving decision problems. However, for an end user it is difficult to implement the methods and algorithms from the raw equations in order to solve a specific problem. This paper presents a decision support tool based on the RIMER approach that facilitates its implementation and use to end-users. The overall structure and main functionalities of the tool are outlined, followed by an example to illustrate the use of this tool for applications.

In this paper we propose an evidential fusion approach to combining the decisions of text classifiers. These text classifiers are generated by four widely used learning algorithms: Support Vector Machine (SVM), kNN (Nearest Neighbour), kNN model-based approach (kNNM), and Rocchio on two text corpora. We first model each classifier output as a list of prioritized decisions and then divide it into the subsets of 2 and 3 decisions which are subsequently represented by the evidential structures in terms of triplet and quartet. We also develop the general formulae based on the Dempster- Shafer theory of evidence for combining such decisions. To validate our method various experiments have been carried out over the data sets of 20-newsgroup and Reuters-21578, and a comparative analysis with an alternative dichotomous structure and with majority voting have also been conducted to demonstrate the advantage of our approach in combining text classifiers.

Glove-based systems are an important option in the field of gesture recognition. They are designed to recognize meaningful expressions of hand motion. In our daily lives, we use our hands for interacting with the environment around us in many tasks. Our glove-based gesture recognition is focused on developing technologies for studying the motion and interaction with a data glove which can augment the capabilities of some users to perform some tasks. This idea is relevant to many research areas, for example: design and manufacturing, information visualization, robotics, sign language understanding, medicine and health Care. In this paper, we proposed a new concept grounding approach for glove-based gesture recognition. We record the data from finger sensors and then abstract and extract concepts from the data. This allow us to construct conceptual levels which we can use to study interaction and manipulation for users during their activities.

In recent years, there is an increasing interest in the research community in finding community structure in complex networks. The networks are usually represented as graphs, and the task is usually cast as a graph clustering problem. Traditional clustering algorithms and graph partitioning algorithms have been applied to this problem. New graph clustering algorithms have also been proposed. Random walk based clustering, in which the similarities between pairs of nodes in a graph are usually estimated using random walk with restart (RWR) algorithm, is one of the most popular graph clustering methods. Most of these clustering algorithms only find disjoint partitions in networks; however, communities in many real-world networks often overlap to some degree. In this paper, we propose an efficient clustering method based on random walks for discovering communities in graphs. The proposed method makes use of network topology and edge weights, and is able to discover overlapping communities. We analyze the effect of parameters in the proposed method on clustering results. We evaluate the proposed method on real world social networks that are well documented in the literature, using both topological-based and knowledge-based evaluation methods. We compare the proposed method to other clustering methods including recently published Repeated Random Walks, and find that the proposed method achieves better precision and accuracy values in terms of six statistical measurements including both data-driven and knowledge-driven evaluation metrics.

The Latent Dirichlet Allocation model is an unsupervised generative model that is widely used for topic modelling in text. We propose to add supervision to the model in the form of domain knowledge to direct the focus of topics to more relevant aspects than the topics produced by standard LDA. Experimental results demonstrate the effectiveness of our method. We also propose a novel Twofold-LDA model to improve the current output of LDA in order to visualize results in graphical form, which can ultimately be used by potential customers. Experiments show the benefit of this new output, with the ability to produce topics focused on our desired aspects in a user friendly chart.

Natural Language Processing (NLP) techniques are believed to hold the potential to assist "bag-of-words" Information Retrieval (IR) in terms of retrieval accuracy. In this paper, we report a natural language based IR approach where the common syntactic structures between documents and the query is regarded to as a query-dependent feature for documents. Specifically, a "structural weight" is proposed for query terms, which can be seen as a weight to model the degree of term's involvement in the common syntactic structures. This structural weight is used together with the TF-IDF weighting scheme, which results in a new ranking function. The accumulation of this structural weight of all the query terms in the new ranking function will be seen as a measure of how much a document and a query share the common syntactic structures. The experimental results show that by using this ranking function, significant improvements in the retrieval performance are achieved.

The number of intelligent environment implementations such as smart homes is set to increase dramatically within the next 40 years. This is predicted using forecasts of demographic data which indicates an expansion of the aged population. It has also been predicted that governments will struggle to meet the demand for resources such as sensor technology due to costs. Optimisation of limited resources involves physically positioning devices to maximise pertinent data gathering potential. Currently the most utilised methodology of distributing limited spatial detection sensors such as pressure mats within smart homes is via ad-hoc deployments performed by a human being. In this study idiosyncratic inhabitant spatial-frequency data was processed using a Pure Random Search (PRS) algorithm to uncover probabilistic future regions of interest, alluding to optimal sensor distributions under resource constraint. With PRS a null hypothesis was stated: `using lower iteration stopping criteria produce less optimal sensor distributions than when using higher iteration stopping criteria'. A student t-test between 1000 and 5000 iterations was statistically significant at 5% (p = 0.016852) whereby the null hypothesis was rejected. Similar results were obtained between other iteration criteria. These data demonstrate that the iteration stopping criterion is not as critical as sensor size or number of sensors; and that comparable results could be obtained when lower stopping parameters are specified when using PRS.

In this paper, a new approach to improve least squares support vector machines is presented. We consider the membership of every sample in constraints, that is to say, every sample are not fully assigned to one class. The membership is computed by employing the technique of fuzzy rough sets, and then a new least squares support vector machine algorithm based on fuzzy rough sets is proposed, experiments are carried out to show that our idea in this paper is feasible and valid.

The neighborhood counting measure is a similarity measure based on the counting of all common neighborhoods in a data space. The minimum risk metric (MRM) is a distance measure based on the minimization of the risk of misclassification. The paper by Argentini and Blanzieri refutes a remark about the time complexity of MRM, and presents an experimental comparison of MRM and NCM. This paper is a response to the paper by Argentini and Blanzieri. The original remark is clarified by a combination of theoretical analysis of different implementations of MRM and experimental comparison of MRM and NCM using straightforward implementations of the two measures.

Time series data abound and analysis of such data is challenging and potentially rewarding. One example is financial time series analysis. Most of the intelligent data analysis methods can be applied in principle, but evolutionary computing is becoming increasingly popular and powerful. In this paper we focus on one task of financial time series analysis - stock price forecasting based on historical data. The premise of this task is that the current price of a stock is dependent on the price of the same stock in the past. Here we consider an additional assumption, i.e., time dependency relevance, that the price in the nearer past is more relevant to the current price than that in the more distant past. This assumption appears intuitively sound, but needs formally validated. In this paper we set to test this assumption by introducing time weighting into similarity measures, as similarity is one of the key notions in time series analysis methods including evolutionary computing. We consider the generic neighborhood counting similarity as it can be specialized for various forms of data by defining the notion of neighborhood in a way that satisfies different requirements. We do so with a view to capturing time weights in time series. This results in a novel time weighted similarity for time series. A formula is also discovered for the similarity so that it can be computed efficiently. Experiments show that this similarity outperforms the standard Euclidean distance and a time weighted variant of it. We conclude that the time dependency relevance assumption is sound.

The neighborhood counting measure is the number of all common neighborhoods between a pair of data points. It can be used as a similarity measure for different types of data through the notion of neighborhood: multivariate, sequence, and tree data. It has been shown that this measure is closely related to a secondary probability G, which is defined in terms of a primary probability P of interest to a problem. It has also been shown that the G probability can be estimated by aggregating neighborhood counts. The following questions can be asked: What is the relationship between this similarity measure and the primary probability P, especially for the task of classification? How does this similarity measure compare with the euclidean distance, since they are directly comparable? How does the G probability estimation compare with the popular kernel density estimation for the task of classification? These questions are answered in this paper, some theoretically and some experimentally. It is shown that G is a linear function of P and, therefore, a G-based Bayes classifier is equivalent to a P-based Bayes classifier. It is also shown that a weighted k-nearest neighbor classifier equipped with the neighborhood counting measure is, in fact, an approximation of the G-based Bayes classifier. It is further shown that the G probability leads to a probability estimator similar in spirit to the kernel density estimator. New experimental results are presented in this paper, which show that this measure compares favorably with the euclidean distance not only on multivariate data but also on time-series data. New experimental results are also presented regarding probability/density estimation. It was found that the G probability estimation can outperform the kernel density estimation in classification tasks.

The mathematical theory of evidence is a generalization of the Bayesian theory of probability. It is one of the primary tools for knowledge representation and uncertainty and probabilistic reasoning and has found many applications. Using this theory to solve a specific problem is critically dependent on the availability of a mass function (or basic belief assignment). In this paper, we consider the important problem of how to systematically derive mass functions from the common multivariate data spaces and also the ensuing problem of how to compute the various forms of belief function efficiently. We also consider how such a systematic approach can be used in practical pattern recognition problems. More specifically, we propose a novel method in which a mass function can be systematically derived from multivariate data and present new methods that exploit the algebraic structure of a multivariate data space to compute various belief functions including the belief, plausibility, and commonality functions in polynomial-time. We further consider the use of commonality as an equality check. We also develop a plausibility-based classifier. Experiments show that the equality checker and the classifier are comparable to state-of-the-art algorithms.

The longest common subsequence is a well known and popular method for measuring similarity between sequences. It advocates the use of information contained in the longest common subsequence as an indication of similarity. In this paper we consider the count of all common subsequences as a measure of sequence similarity with the view that all common information is captured. This measure is inspired and derived from a generic similarity measure, neighborhood counting metric. The close connection of the neighborhood counting metric with probability and the Bayes classifier helps gain an insight from the probabilistic perspective into the all-common subsequences measure. We design algorithms to calculate this measure and we also carry out an experiment in the framework of k-nearest neighbors on a gene sequence classification task. The experiment shows that the all-common subsequences measure and the longest common subsequence measure have little difference for small k values, but differ significantly for large k values. The performance of the all-common subsequences measure remains steady as k gets larger, but the performance of the longest common subsequence measure drops sharply as k gets larger. Such a property may be useful for those tasks where we are interested not only in the nearest neighbor, but also in the first k nearest neighbors. The main contribution of this paper is a suite of algorithms for finding all common subsequences.

Finding nearest neighbors is a general idea that underlies many artificial intelligence tasks, including machine learning, data mining, natural language understanding, and information retrieval. This idea is explicitly used in the k-nearest neighbors algorithm (kNN), a popular classification method. In this paper, this idea is adopted in the development of a general methodology, neighborhood counting, for devising similarity functions. We turn our focus from neighbors to neighborhoods, a region in the data space covering the data point in question. To measure the similarity between two data points, we consider all neighborhoods that cover both data points. We propose to use the number of such neighborhoods as a measure of similarity. Neighborhood can be defined for different types of data in different ways. Here, we consider one definition of neighborhood for multivariate data and derive a formula for such similarity, called neighborhood counting measure or NCM. NCM was tested experimentally in the framework of kNN. Experiments show that NCM is generally comparable to VDM and its variants, the state-of-the-art distance functions for multivariate data, and, at the same time, is consistently better for relatively large k values. Additionally, NCM consistently outperforms HEOM (a mixture of Euclidean and Hamming distances), the "standard" and most widely used distance function for multivariate data. NCM has a computational complexity in the same order as the standard Euclidean distance function and NCM is task independent and works for numerical and categorical data in a conceptually uniform way. The neighborhood counting methodology is proven sound for multivariate data experimentally. We hope it will work for other types of data.

Many algorithms for query rewriting using views have been proposed. Most of them are used for rewriting conjunctive queries with conjunctive views only. Only one inverse rule-based rewriting algorithm has been presented to deal with query rewriting with disjunctive views, in which a set of disjunctive inverse rules are created and used to generate the query rewritings. However, there has no been bucket-based algorithm for this issue. In this paper we apply the ideas of the previous bucket-based algorithms to deal with query rewriting using views in the presence of disjunctions in view definitions. We create a set of buckets over either a conjunctive view or a disjunctive view. Specially, when a bucket is over a disjunctive view, we present a method to remove the tuples that are in the view but not required by a given query. The rewritings obtained by our algorithm are contained in the original query.

An evidence theoretic classification method is proposed in this paper. In order to classify a pattern we consider its neighbours, which are taken as parts of a single source of evidence to support the class membership of the pattern. A single mass function or basic belief assignment is then derived, and the belief function and the pignistic (‘betting rates’) probability function can be calculated. Then the (posterior) conditional pignistic probability function is calculated and used to decide the class label for the pattern. It is shown that such a classifier extends the standard majority voting based k-nearest neighbour classifier, and it is an approximation to the optimal Bayes classifier. In experiments this classifier performed as well as or better than the voting and distance weighted k-nearest neighbours classifiers with best k, and its performance became stable when the number of neighbours considered was >4.

This paper presents a novel method for classification, which makes use of models built by the lattice machine (LM). The LM approximates data resulting in, as a model of data, a set of hyper tuples that are equilabelled, supported and maximal. The method presented uses the LM model of data to classify new data with a view to maximising the density of the model. Experiments show that this method, when used with the LM, outperforms the C2 algorithm and is comparable to the C5.0 classification algorithm.

The formalization of the “part – of” relationship goes back to the mereology of S. Leśniewski, subsequently taken up by Leonard & Goodman (1940), and Clarke (1981). In this paper we investigate relation algebras obtained from different notions of “part-of”, respectively, “connectedness” in various domains. We obtain minimal models for the relational part of mereology in a general setting, and when the underlying set is an atomless Boolean algebra.

Relevance has traditionally been linked with feature subset selection, but formalization of this link has not been attempted. In this paper, we propose two axioms for feature subset selection-sufficiency axiom and necessity axiom-based on which this link is formalized: The expected feature subset is the one which maximizes relevance. Finding the expected feature subset turns out to be NP-hard. We then devise a heuristic algorithm to find the expected subset which has a polynomial time complexity. The experimental results show that the algorithm finds good enough subset of features which, when presented to C4.5, results in better prediction accuracy.