Multi-modal Entity Alignment (MMEA) aims to identify equivalent entities across diverse knowledge graphs by leveraging structural, attribute, and visual information. However, real-world datasets frequently suffer from missing modalities, necessitating feature imputation. A critical yet underexplored issue is that not all imputed modalities are inherently trustworthy. Ignoring the aleatoric uncertainty of such modalities introduces severe noise which propagates through the fusion process and degrades alignment performance. To address this challenge, we propose a novel MMEA framework, namely SURE, to Suppress Uncertainty for tRustworthy Entity alignment. Specifically, SURE introduces an uncertainty-aware variational imputation module to estimate the aleatoric uncertainty of generated features. Crucially, rather than using these imputation features blindly, SURE leverages the estimated uncertainty to suppress noise propagation via a confidence-gated multi-modal fusion and an adaptive contrastive learning objective. Extensive experiments on DBP15K datasets demonstrate that SURE significantly outperforms state-of-the-art baselines, exhibiting exceptional robustness particularly in scenarios with high modality missing rates.

In recent years, large language models (LLMs) have emerged as powerful tools for link prediction in knowledge graphs (KGs) due to their strong capabilities in understanding and generation. However, many LLM-based methods still heavily rely on textual descriptions of KGs, limiting their ability to capture structural information and to model complex relational patterns. Although some methods integrate structural embeddings into LLMs, their ability to harness the complementary strengths of both modalities and dynamically prioritize candidate entities based on query context remains limited. In this paper, we propose ST-KGLP, a novel framework that improves link prediction by aligning structural knowledge with textual knowledge and employing query-aware adaptive weighting for candidate selection. Specifically, our proposed ST-KGLP employs a knowledge aligner to bridge the information gap between structural and textual knowledge, and then utilizes a query-aware adaptive weighting strategy that dynamically computes attention weights between query representations and candidate entities, enabling contextually relevant candidate re-ranking for more accurate prediction. Extensive experiments on various datasets show that our ST-KGLP outperforms state-of-the-art approaches, achieving average improvements of 3.81%, 11.52%, 2.22%, and 1.55% across four evaluation metrics. Our code and datasets are available at https://github.com/shijielaw/ST-KGLP.

To achieve multi-hop question answering over knowledge graphs (KGQA), many studies have explored converting retrieved subgraphs into textual form and feeding them into large language models (LLMs) to leverage their reasoning capabilities. However, due to the linear and discrete nature of text sequences, model performance may degrade when handling complex questions. To this end, we propose a novel structure-text knowledge synergistic method, BrikQA, which bridges the knowledge gap between knowledge graphs (KGs) and LLMs for multi-hop KGQA. LLMs and KGs complement each other by leveraging explicit topological patterns and implicit knowledge mining to enhance knowledge understanding and address sparsity issues. Experimental results on various datasets demonstrate that BrikQA outperforms state-of-the-art baselines. Our source code is available at https://github.com/shijielaw/BrikQA.

Blood glucose levels, when analyzed as time series data, contain vital information about a patient's glucose fluctuations and trends, which are crucial for assessing pancreatic function and insulin sensitivity. Leveraging these insights through feature extraction and classification can help determine the optimal insulin dosage for patients. Traditional methods of manual feature selection and extraction are laborintensive and inefficient. In this paper, we propose a novel local-global context encoding architecture to extract deep features from blood glucose time series data. We employ a weighted k-nearest neighbors (KNN) classifier to provide dynamic and personalized treatment recommendations. Our experiments, conducted using the UVA/Padova type 1 diabetes simulator and the ShanghaiT2DM real-world diabetes dataset, demonstrate that deep features significantly outperform manually extracted features in recommending insulin dosages. Our approach offers a more accurate and personalized method for managing multiple daily insulin injections, potentially reducing the workload of medical practitioners and improving the efficiency and effectiveness of diabetes management. Code is available at https://github.com/EricWvi/ykw.

Recently, deep reinforcement learning (DRL) has been employed for intelligent traffic‐light control and demonstrated promising results. However, state‐of‐the‐art DRL‐based systems still rely on discrete decision‐making, which can lead to unsafe driving practices. Additionally, existing feature representations of the environment often fail to capture the complex dynamics of traffic flows, resulting in imprecise predictions of traffic conditions. To address these issues, we propose a novel DRL framework based on the multiagent deep deterministic policy gradient algorithm. Our method offers several key innovations: it suggests employing a transitional phase before changing the current phase for safer traffic management, integrates local road network topology into feature representation to enhance the accuracy of traffic flow predictions, and uses two‐layer regional features to improve coordination among agents within the region. Our extensive evaluations using simulation of urban mobility, a widely used multimodal traffic simulation package, demonstrated that the proposed method outperformed previous methods and reduced the number of emergency stops, queue lengths, and waiting times.

Using Graph Neural Networks (GNNs) to simulate complex physical systems is currently a popular field. Finite Element Analysis (FEA) is a numerical analysis solution in physical simulation and involves dividing a physical structure into finite elements. A mesh-based graph network is formed by interconnecting the nodes of the finite elements. Mesh-based simulations are central to the modeling of physical systems, and GNNs are used to overcome the high cost of the simulation. However, most existing GNNs-based methods lack the ability to generalize when small changes in structure occur. The cantilever beam, which is a fundamental component of many physical structures, is a classic study case for FEA stress analysis. In this paper, we propose a Mesh-based Graph Transfer Learning (MGTL) method for beams, in which the knowledge learned on a mesh-based graph network of one beam structure via contrastive learning, and then is transferred to another. Furthermore, an augmentation approach is proposed for the mesh-based graph in contrastive learning. Our experiments demonstrate the effectiveness of transfer learning in the MGTL method, as well as its generalization ability across different mesh sizes and structure types.

With the burgeoning prevalence of diabetes and the urgent imperative for effective dietary management, harnessing the power of image analysis and computer vision technology presents a promising solution to the automatic nutrition estimation challenge. However, constructing a comprehensive frame-work for this endeavor presents its own set of challenges, encompassing multiple intricate procedures, including data collection, image segmentation, food recognition, and volume estimation. In this paper, we introduce a sophisticated and automated framework for Chinese food nutrition estimation, leveraging food images sourced from diabetic patients. In addition to developing models for the various procedures, the framework necessitates label images for model training. These label images encompass ground truth masks delineating precise object boundaries for image segmentation, food type labels for food recognition, and comprehensive nutrition tables for each food item. Through rigorous experimentation and validation, our framework has demonstrated its efficacy, offering a convenient and practical tool for managing dietary requirements in diabetes care.

The constrained shortest path ($\mathtt {CSP}$) problem has wide applications in travel path planning, mobile video broadcasting and network routing. Existing works do not work well on large dynamic graphs and suffer from either ineffectiveness or low scalability issues. To overcome these issues, in this paper, we propose an efficient and effective solution framework, namely$\mathtt {CSP\_{G}S}$. The solution framework includes two key components: (1) the techniques to decompose a large$\mathtt {CSP}$instance into multiple small sub-instances and (2) the developed learning model$\mathtt {CSP\_{D}QN}$to solve small$\mathtt {CSP}$instances. The evaluation result on real road network graphs indicates that our approach$\mathtt {CSP\_{G}S}$performs well on large dynamic graphs by rather high quality and reasonable running time, and particularly adapt to significant graph changes even with broken edges. To the best of our knowledge, this is the first learning-based model to well solve the$\mathtt {CSP}$problem on large dynamic graphs.

Multivariate time series forecasting has wide applications such as traffic flow prediction, supermarket commodity demand forecasting and etc., and a large number of forecasting models have been developed. Given these models, a natural question has been raised: what theoretical limits of forecasting accuracy can these models achieve? Recent works of urban human mobility prediction have made progress on the maximum predictability that any algorithm can achieve. However, existing approaches on maximum predictability on the multivariate time series fully ignore the interrelationship between multiple variables. In this article, we propose a methodology to measure the upper limit of predictability for multivariate time series with multivariate constraint relations. The key of the proposed methodology is a novel entropy, named Multivariate Constraint Sample Entropy (McSE), to incorporate the multivariate constraint relations for better predictability. We conduct a systematic evaluation over eight datasets and compare existing methods with our proposed predictability and find that we get a higher predictability. We also find that the forecasting algorithms that capture the multivariate constraint relation information, such as GNN, can achieve higher accuracy, confirming the importance of multivariate constraint relations for predictability.

For multi-class food images, an excellent segmentation method has a great influence on accuracy of recognition result, and hence improve the effectiveness of dietary management for diabetics. For Chinese food images, there are some challenges during the processing, such as blurred outlines, rich colors, and varied appearances due to various cooking methods. To overcome these difficulties, we propose a ChineseFoodSeg approach to obtain accurate and efficient multi-class segmentation. A food recognition model, Two-Path Global Local Network (TPGLNet), is also introduced to jointly learn complementary global and local features of the bounding box and the segment. Experiments on the ChineseDiabetesFood187 dataset collected by us demonstrate that the new methods are competitive compared to existing segmentation and recognition methods.

The threshold digital signature technique is important for decentralized finance (DeFi) applications such as asset custody and cross-chain interoperations. The Edwards-curve digital signature algorithm (EdDSA) is widely used in blockchains, e.g., Libra/Diem; however, no suitable threshold solution exists. Therefore, to bridge this gap, we propose a threshold EdDSA that allows n parties to generate keys in a decentralized and distributed manner. Any t + 1-of-n parties can generate standard EdDSA signatures. This scheme supports an arbitrary threshold (t, n) and has been proven to be secure against at most t malicious adversaries. The theoretical analysis (computation complexity and communication footprints) and experimental results demonstrate that the proposed scheme performs efficiently on cloud servers and embedded devices. Furthermore, the proposed scheme is integrated with Tendermint, a blockchain framework that uses EdDSA, to generate keys and sign transactions in a decentralized manner, which indicates that this scheme is compatible with blockchains for supporting DeFi applications.

Content-based Pub/Sub paradigm has been widely used in many distributed applications and existing approaches suffer from high redundancy subscription index structure and low matching efficiency. To tackle this issue, in this paper, we propose a learning framework to guide the construction of an efficient in-memory subscription index, namely PMIndex, via a multi-task learning framework. The key of PMIndex is to merge redundant subscriptions into an optimal number of partitions for less memory cost and faster matching time. Our initial experimental result on a synthetic dataset demonstrates that PMindex outperforms two state-of-the-arts by faster matching time and less memory cost.

Web users provide rich multi-modal and heterogeneous data. Taking use of the data for extracting the representation of web users is a vital prerequisite of further profiling tasks. In this paper, we study the web user profiling based on the user-entity interaction data integrating the taxonomy information of the entity. To overcome the challenges of the representation of a web user and the distance definition of two web users, we introduce TaxoVec, an interpretable method to represent a user’s interaction history by vectors with the help of the entities’ category information. The TaxoVec is employed in clustering-based web user profiling on their transaction data and geographical data, which provide the information on their interests and activity area. Experiments show that the TaxoVec performs well on representing characteristics of web users.

Deep learning has become a key technology on modeling large amounts of multi-sourced data. For privacy concerns, the data sharing among companies and organizations is increasingly difficult. In this paper, we present a crowd-sourced federated learning solution to train neural networks with a hybrid blockchain architecture. Smart contracts are used to share data authentications on the main chain, where the proxy re-encryption is for the privacy preserving. A consensus-based asynchronous practical byzantine federated learning (APBFL) algorithm is proposed on the side chains, to improve the model reliability and security. Experiments show that our solution is efficient, secure and robust.

Recently deep reinforcement learning (DRL) has been used for intelligent traffic light control. Unfortunately, we find that state-of-the-art on DRL-based intelligent traffic light essentially adopts discrete decision making and would suffer from the issue of unsafe driving. Moreover, existing feature representation of environment may not capture dynamics of traffic flow and thus cannot precisely predict future traffic flows. To overcome these issues, in this paper, we propose a DDPG-based DRL framework to learn a continuous time duration of traffic signal phases by introducing 1) a transit phase before the change of current phase for better safety, and 2) vehicle moving speed into feature representation for more precise estimation of traffic flow in next phase. Our preliminary evaluation on a well-known simulator SUMO indicates that our work significantly outperforms a recent work by much smaller number of emergency stops, queue length and waiting time.

Image aesthetic captioning is a multi-modal task that is to generate aesthetic critiques for images. In contrast to common image captioning tasks, where different captions aimed at providing factual descriptions of a same image are always similar, captions with respect to different aesthetic attributes of the same image can be totally different in an aesthetic captioning task. Such inter-aspect differences are always overlooked, which leads to the lack of diversity and coherence of the captions generated by most of the existing image aesthetic captioning systems. In this paper, we propose a novel model to generate aesthetic captions for food images. Our model redefines food image aesthetic captioning as a compositional task that consists of two separated modules, i.e., a single-aspect captioning and an unsupervised text compression. The first module is guaranteed to generate the captions and learn feature representations of each aesthetic attribute. Then, the second module is supposed to study the associations among all feature representations and automatically aggregate captions of all aesthetic attributes to a final sentence. We also collect a dataset which contains pair-wise image-comment data related to six aesthetic attributes. Two new evaluation criteria are introduced to comprehensively assess the quality of the generated captions. Experiments on the dataset demonstrate the effectiveness of the proposed model.

Multivariate time series forecasting has wide applications such as traffic flow prediction, supermarket commodity demand forecasting and etc. In literature, Due to the complex temporal patterns and inter-dependencies among multivariate time series, a large number of forecasting models have been developed. However, one question still remains unclear: how these models perform on a certain forecasting task, and there is lack of comprehensive performance comparison of these models on different tasks. To this end, in this paper, we conduct a systematic evaluation of eight representative forecasting models over eight multivariate time series datasets, and have the following findings: 1) When the datasets exhibit strong periodic patterns, deep learning models perform best. Otherwise on the datasets in a non-periodic manner, the statistical models such as ARIMA perform best. 2) For the long term prediction involving a high horizon value, the direct prediction strategy could lead to lower errors than the recursive one, but at the cost of higher training time. 3) For the multivariate time series explicitly involving graphic inter-dependencies among the multivariates, e.g., the road network topology in the spatio-temporal time series of traffic volumes in multiple routes, the Graph Convolution Network can incorporate the graphic inter-dependencies into their forecasting models for smaller prediction errors.

Nowadays, transaction data of users on the network platform has become important raw materials to analyze their purchase behaviors and consumption patterns. In fields such as advertising and marketing, it is important to identify certain groups of users to target. Therefore, an automatic user categorization method is meaningful. In this paper, we introduce a user categorization method on e-commerce transaction data based on the products that users are interested in. The categorization method introduces a novel distance metric for two users and an efficient clustering algorithm based on it. We conduct our experiments on the yelp website, and the dataset with 1.2 million users are tested. Compared with two commonly used clustering algorithms, the proposed method shows the best efficiency, which is suitable for large data sets. Furthermore, the clusters of users obtained are summarized by keywords which are extracted from their review comments.

Accurate prediction of traffic congestion at the granularity of road segment is important for planning travel routes and optimizing traffic control in urban areas. Previous works often calculated only the average congestion levels of a large region covering many road segments and did not take into account spatial correlation between road segments, resulting in inaccurate and coarse-grained prediction. To overcome these issues, we propose in this paper CPM-ConvLSTM, a spatiotemporal model for short-term prediction of congestion level in each road segment. Our model is built on a spatial matrix which incorporates both the congestion propagation pattern and the spatial correlation between road segments. The preliminary experiments on the traffic data set collected from Helsinki, Finland prove that CPM-ConvLSTM greatly outperforms 6 counterparts in terms of prediction accuracy.

The volume of trajectory data has become tremendously large in recent years. How to efficiently maintain and compute such trajectory data becomes a challenging task. In this paper, we propose a trajectory spatial and temporal compression framework, namely CLEAN. The key of spatial compression is to mine meaningful trajectory frequent patterns on road networks. By treating the mined patterns as dictionary items, we have the chance to encode a long trajectory by shorter paths, thus leading to smaller space cost. Meanwhile, we design an error-bounded temporal compression on top of the identified spatial patterns for much low space cost. Extensive experiments on real trajectory datasets validate that CLEAN significantly outperforms existing state-of-art approaches in terms of both space saving and runtime of trajectory compression.

Time-Sync Comment (TSC) is a type of crowdsourced user review embedded in online video websites, which provides better real-time user interaction than traditional user comment type. Various TSC-related problems and approaches have been studied to improve user experience by taking advantage of special characteristics of TSCs such as strong time reliance. However, there are three major drawbacks to these TSC researches. First, they did not explicitly show advantage of TSC features over the traditional features in terms of users' experience. Second, the experiments were conducted on some inconsistent TSC datasets crawled from different source, which makes the effectiveness of their methods less convincing. Third, the methods were manually evaluated by a limited number of so-called "experts" in these experiments, so it is hard for other researchers to obtain the data labels and reproduce the results. In order to overcome these drawbacks, this paper aims to explore the usefulness of TSC data for for the improvement of user experience online by exploiting the TSC pattern inside a new dataset. Specifically, we present a larger-scale TSC dataset with four-level structures and rich self-labeled attributes and formally define a group of TSC-related research problems based on this dataset. The problems are solved by adapted state-of-the-art methods and evaluated through crowdsourced labels in the dataset. The result can be regarded as a baseline for further research.

As Internet is developing rapidly, new information is generated with an incredibly high speed. In this situation, publish/subscribe services provided by Cloud providers are used to recommend useful information to the users in the Internet. However, the middleware models and routing algorithms in existing publish/subscribe approaches cannot perform efficiently in the P2P Cloud, which is one of the novel Cloud types that belongs to the distributed Cloud. The P2P Cloud brings challenge to publish/subscribe services because it has larger scale and more complicated network architecture than the traditional Cloud. In this paper, a Potential Field based service framework is proposed to provide high-quality publish/subscribe services in P2P Cloud. A middleware is designed in the framework to match subscribe requests and published information. Furthermore, a Potential Field Model is constructed. Under the model, event brokers are organized as P2P overlay for routing on request message. The simulation results show our proposed method has advantages in recall rate, response time, and bandwidth cost against other methods.

The development of location-based applications raises a new challenge to manage and visualize large amounts of geo-tags presented on a web map. The visualization of the geo-tags often leads to a clutter problem, especially in web-mapping systems. We present a new clustering method to reduce the amount of visual clutter. A split smart swap strategy, which has the advantage that it can be applied to a certain data only once at all map scales, is employed in the method. We compare the proposed method to several other methods. Taking the advantage of the one-time running offline, the proposed method is more applicable for the clutter problem.

In different types of feature selection algorithms, feature clustering is an emerging subset generation paradigm. In this paper, a Minimum spanning tree based Feature Clustering (MFC) algorithm is proposed. In the algorithm, an information-theoretic based measure, i.e., Variation of information, is utilized as the feature redundancy and relevance metric. At the clustering phase, the sum of pair wise feature redundancy is minimized. Then, a representative feature is selected from each cluster, where the relevance between representative features and the target label is maximized. The algorithm is supervised since it is designed for various supervised learning problems, such as classification and regression. The proposed MFC is compared with three conventional feature selection algorithms, two of which are also feature clustering method. The MFC obtains well separated feature clusters in the experiment and considerable better classification accuracies applied on several real data sets.

Clustering algorithm and cluster validity are two highly correlated parts in cluster analysis. In this paper, a novel idea for cluster validity and a clustering algorithm based on the validity index are introduced. A Centroid Ratio is firstly introduced to compare two clustering results. This centroid ratio is then used in prototype-based clustering by introducing a Pairwise Random Swap clustering algorithm to avoid the local optimum problem of k -means. The swap strategy in the algorithm alternates between simple perturbation to the solution and convergence toward the nearest optimum by k -means. The centroid ratio is shown to be highly correlated to the mean square error (MSE) and other external indices. Moreover, it is fast and simple to calculate. An empirical study of several different datasets indicates that the proposed algorithm works more efficiently than Random Swap, Deterministic Random Swap, Repeated k-means or k-means++. The algorithm is successfully applied to document clustering and color image quantization as well.

Expectation maximization (EM) algorithm, being a gradient ascent algorithm depends highly on the initialization. Repeating EM multiple times with different initial solutions and taking the best result is used to attack this problem. However, the solution space is searched inefficiently in Repeated EM, because after each restart it can take a long time to converge without any guarantee that it leads to an improved solution. A random swap EM algorithm utilizes random swap strategy to improve the problem in a more efficient way. In this paper, a theoretical and experimental comparison between RSEM and REM is conducted. Based on GMM estimation theory, it is proved that RSEM reaches the optimal result faster than REM with high probability. It is also shown experimentally that RSEM speeds up REM from 9% to 63%. A study in color-texture images demonstrates an application of EM algorithms in a segmentation task.

External validity measures in cluster analysis evaluate how well the clustering results match to a prior knowledge about the data. However, it is always intractable to get the prior knowledge in the practical problem of unsupervised learning, such as cluster analysis. In this paper, we extend the external validity measures for both hard and soft partitions by a resampling method, where no prior information is needed. To lighten the time burden caused by the resampling method, we incorporate two approaches into the proposed method: (i) extending external validity measures for soft partitions in a computational time of O(M2N); (ii) an efficient sub-sampling method with time complexity of O(N). The proposed method is then applied and reviewed in determining the number of clusters for the problem of unsupervised learning, cluster analysis. Experimental results has demonstrated the proposed method is very effective in solving the number of clusters.

The expectation-maximization (EM) algorithm is a popular tool in estimating model parameters, especially mixture models. As the EM algorithm is a hill-climbing approach, problems such as local maxima, plateau and ridges may appear. In the case of mixture models, these problems involve the initialization of the algorithm and the structure of the data set. We propose a random swap EM algorithm (RSEM) to overcome these problems in Gaussian mixture models. Random swaps are repeatedly performed in our method, which can break the configuration of the local maxima and other problems. Compared to the strategies in other methods, the proposed algorithm has relative improvements on log-likelihood value in most cases and less variance than other algorithms. We also apply RSEM to the image segmentation problem.

The main challenge of cluster analysis is that the number of clusters or the number of model parameters is seldom known, and it must therefore be determined before clustering. Bayesian Information Criterion (BIC) often serves as a statistical criterion for model selection, which can also be used in solving model-based clustering problems, in particular for determining the number of clusters. Conventionally, a correct number of clusters can be identified as the first decisive local maximum of BIC; however, this is intractable due to the overtraining problem and inefficiency of clustering algorithms. To circumvent this limitation, we proposed a novel method for identifying the number of clusters by detecting the knee point of the resulting BIC curve instead. Experiments demonstrated that the proposed method is able to detect the correct number of clusters more robustly and accurately than the conventional approach.