Zero-shot object navigation presents a highly challenging task in embodied AI, requiring an agent to interpret natural language instructions, perceive complex visual environments, and plan actions without any task-specific training. While recent approaches have introduced large language models (LLMs) as high-level planners, they often rely on static, one-shot inference and struggle with ambiguous or partially observable scenes. This paper proposes DebateNav, a novel multi-agent decision framework that integrates multiple vision-language model (VLM) experts under the supervision of a central LLM controller. Each VLM is assigned a unique expert role (e.g., object detection, risk assessment, spatial reasoning), and together they engage in structured multi-round debates when perception conflicts arise. The LLM controller performs task decomposition, memory-guided exploration, and final arbitration based on expert arguments. To enhance the perception and decision process, DebateNav incorporates a multimodal image fusion module combining RGB, depth, and segmentation inputs, as well as a map memory and trajectory tracking system that helps avoid redundant exploration and supports long-horizon planning. The system is evaluated on a subset of the HM3D dataset with approximately 5,000 tasks, achieving a$\mathbf{5 2. 3 \%}$success rate and$\mathbf{1 5. 5}$SPL under strict zero-shot conditions. Extensive ablation studies confirm the effectiveness of the expert debate mechanism, multi-modal fusion, memory system, and LLM-based arbitration. The results demonstrate that DebateNav outperforms several recent baselines and establishes a new perspective on collaborative, interpretable planning for zero-shot embodied navigation.

In task-oriented dialogue systems, intent recognition and entity extraction are key for driving system understanding and state updates. However, traditional structured systems often show limited robustness and generalization when facing complex expressions. To improve understanding in these scenarios, this paper proposes a hybrid dialogue system framework. The framework fuses a traditional pipeline architecture with a large language model (LLM). It is built on a conventional pipeline. It evaluates the uncertainty of pipeline predictions online. When confidence falls below a preset threshold, it automatically invokes the LLM for semantic enhancement and reanalysis. This dynamic step compensates for the shortcomings of the structured process. To ensure controllability and structural alignment of generative outputs, we design a dual-control mechanism. The mechanism integrates a domain-adaptive prompt selection strategy with an output normalization protocol. This design significantly improves LLM accuracy and consistency in entity extraction. Experiments on the MultiWOZ 2.2 dataset show that the hybrid framework outperforms traditional structured methods. It handles complex inputs and colloquial queries more effectively.

Accurate identification of N7-methylguanosine (m7G) modification sites plays a critical role in uncovering the regulatory mechanisms of various biological processes, including human development, tumor initiation, and progression. However, existing prediction methods still suffer from limited representational power, redundant feature fusion, insufficient utilization of biological prior knowledge, and poor interpretability. In this study, we propose a novel deep learning model named MCAMEF-BERT. This model adopts a parallel architecture that integrates both a DNABERT-2-based pretrained model branch and multiple traditional feature encoding branches, enabling comprehensive multi-perspective sequence feature extraction. To address the redundancy issue in feature fusion, we introduce a multi-channel attention module. Our model demonstrates superior accuracy and effectiveness on datasets from m7GHub, outperforming other state-of-the-art classifiers. Furthermore, we validate the interpretability of MCAMEF-BERT through in silico saturation mutagenesis experiments, and confirm its robustness in motif recognition. Moreover, its generalization capability is validated across diverse RNA modification site prediction tasks.

Incorporating unmanned aerial vehicles (UAVs) and distributed sensors into aerial-terrestrial networks (ATNs) improves the ability to monitor remote and challenging environments. However, in such harsh environments, the UAV must return to the base station multiple times to be recharged to ensure that all sensors can be accessed. Neglecting the energy consumption associated with the backhaul recharging path would lead to a decrease in task efficiency and an increase in overall energy consumption. To address the challenge, this article formulates a UAV back-haul trajectory planning (UBTP) problem, aiming to optimize both the data collection and charging paths for UAVs. At the same time, a two-stage deep reinforcement learning (DRL)-based method is developed to address the UBTP problem. The first stage of the proposed method determines the navigation sequence taking into account the energy consumption of the data collection and charging path. The subsequent stage refines the trajectory by considering the sensing ranges of the sensors. The numerical results demonstrate that the proposed approach provides solutions that are more energy-efficient compared to benchmark methods, thereby enhancing the UAV’s ability to harvest data sustainably and effectively from distributed sensors in challenging environments.

Aerial-terrestrial network (ATN) framework is currently the dominant method of the Internet of unmanned agents (IUAs) for integrating both aerial vehicles and terrestrial sensors. In ATN scenarios, existing work shows that route planning utilizing deep reinforcement learning (DRL) is important as it can conserve energy for aerial vehicles or diminish network latency. However, utilizing DRL approaches in a multiagent ATN environment may lead to inefficiencies when sharing original interaction data. In addition, direct data exchange between agents can raise significant privacy concerns. To address these challenges, this work develops a novel federated reinforcement learning (FRL) approach called FRIGE, which takes advantage of ensemble learning for DRL-based methods. Specifically, this article develops a prioritized gradient ensemble technique for local DRL agents, which constructs twin networks to obtain prioritized gradients without incurring additional interactive costs. After aggregating local models on the server, the latest iteration of the global model is used to update the networks of both agents and their twin networks for subsequent rounds of training. Extensive experiments are conducted on two typical ATN route planning tasks to validate FRIGE’s advancement and generalization capabilities. The numerical results demonstrate that FRIGE not only enhances data sharing efficiency among all agents but also maintains privacy while delivering superior solutions for ATN route planning tasks.

Partial multi-label learning (PML) addresses scenarios where each training sample is associated with multiple candidate labels, but only a subset are ground-truth labels. The primary difficulty in PML is to mitigate the negative impact of noisy labels. Most existing PML methods rely on sufficient samples to train a noise-robust multi-label classifier. However, in practical scenarios, such as privacy-sensitive domains or those with limited data, only a few training samples are typically available for the target task. In this paper, we propose an approach called FsPML-CNL (Few-shot Partial Multi-label Learning with Credible Non-candidate Label) to tackle the PML problem with few-shot training samples. Specifically, FsPML-CNL first utilizes the sample features and feature-prototype similarity in the embedding space to disambiguate candidate labels and to obtain label prototypes. Then, the credible non-candidate label is selected based on label correlation and confidence, and its prototype is incorporated into the training samples to generate new data for boosting supervised information . The noise-tolerant multi-label classifier is finally induced with the original and generated samples, along with the confidence-guided loss. Extensive experiments on public datasets demonstrate that FsPML-CNL outperforms competitive baselines across different settings.

Transformers based on Self-Attention (SA) mechanism have demonstrated unrivaled superiority in numerous areas. Compared to RNN-based networks, Transformers can learn the temporal dependency representation of an entire sequence in parallel, while efficiently dealing with long-range dependencies. However, the $\mathcal {O}(L^{2})$O(L2) ($L$L denotes the length of the sequence) computational complexity of the SA mechanism and the high memory usage make the construction cost of the Transformer-based model prohibitively expensive. To address these challenges, we propose a Transformer-like model, HPformer: Low-Parameter Transformer with Temporal Dependency Hierarchical Propagation. HPformer first chunks the sequence into $K$K ($K = \left\lceil \log {L} \right\rceil + 1$K=logL+1, $\left\lceil \cdot \right\rceil$· denotes ceiling operation) sequence segments, then leverages the hierarchical propagation mechanism with $\mathcal {O}(L)$O(L) computational complexity to learn the temporal dependencies between the segments and within the segments, and ultimately generates $K$K vectors as $Key$Key matrices. This reduces the complexity of the SA mechanism from $\mathcal {O}(L^{2})$O(L2) to $\mathcal {O}(L\log {L})$O(LlogL). In addition, we employ a strategy of sharing $Key$Key and $Value$Value matrices between layers to build the HPformer, thus reducing memory usage. Extensive experiments based on public health informatics benchmark and Long-Range Arena (LRA) benchmark have demonstrated that HPformer has advantages over Transformer-based models in terms of memory usage and efficiency.

Federated Learning is a data privacy-protected distributed machine learning framework, but malicious clients can damage it. Byzantine-robust federated learning aims to learn an accurate global model despite the presence of malicious clients. Most current defenses assume that clients have identically and independently distributed (i.i.d.) data and thus cannot perform well in canonical non-i.i.d. scenarios. Several non-i.i.d. statistic-based defenses have been recently proposed to identify malicious clients through gradient statistics without any auxiliary techniques. They thus can only ensure robustness under certain attacks with specific characteristics. We propose Sophon, a comprehensive defense using auxiliary data to combat an arbitrary number of malicious clients in both i.i.d. and non-i.i.d. cases. Specifically, Sophon first normalizes received client gradients to reduce the dominance of malicious gradients. Then it introduces a dual trust mechanism to assign the aggregation weight for each gradient. The dual trust mechanism estimates the consistency-based and diversity-based trust scores of client gradients and integrates the two scores as the aggregation weight to effectively suppress the impact of malicious gradients. Extensive experimental results on three datasets from different domains, with diverse models and FL scenarios, show that Sophon is robust in maintaining the overall accuracy of the training model.

At present, the microservice architecture is widely used in modern software development for its flexibility and scalability. However, the huge data scale and complex invocation relationships between services make the root cause localization of faults in microservice systems extremely difficult. Some of the current root cause localization methods based on metrics data use history fault data to effectively improve the localization results, but there are challenges such as not representing history faults effectively and limiting the localization results to repetitive faults that have occurred in history. In this paper, we propose an automatic root cause localization framework MicroHFRCL based on the history fault library to address the above issues. MicroHFRCL constructs an instance causal graph based on metric data for causal analysis. The instance causal graph is weighted by encoding the anomalous subgraph and calculating the similarity of history faults. The PageRank algorithm is used to locate the root cause of faults. Among them, MicroHFRCL learns the structure and feature information of fault anomalous subgraphs through GCN and Transformer models, achieving effective representation of history faults and fast calculation of similarity in history fault codes, improving the efficiency of repetitive fault localization, and effectively solving the problem of the limitation of using history faults for root cause localization results. We implemented MicroHFRCL and tested it on the fault dataset collected by a benchmark microservice test system. Compared with the latest baseline models, MicroHFRCL has significantly improved the localization accuracy, and can also achieve good results in the case of small-scale history faults.

In recent years, with the emergence of large models that have extremely high parameter counts and require vast amounts of training data, the number of tasks demanding multiple GPUs surges, promoting the development of computing platforms equipped with substantial GPU resources. A trusted GPU allocation scheme is an essential part of building a trusted deep learning computing platform. Quantifying the relationship between the number of GPUs allocated and the efficiency of deep learning training tasks helps the platform’s scheduling system provide users with more convincing GPU resource allocation plans, improving task execution efficiency, ensuring fairness among tasks, and reducing and stabilizing energy consumption. However, this topic does not receive significant attention and research. In this paper, we conduct an empirical study by deploying different types of deep learning training tasks on the multi-GPU server to explore the relationship between GPU allocation and task execution efficiency. We summarize the efficiency change patterns of various types of tasks under different GPU allocations. Experiments on various training tasks indicate that models with different parameter sizes exhibit different trends as the number of allocated GPUs increases. This study analyzes these trends and uses function fitting methods to describe the efficiency changes of similar tasks.

As a new design mechanism for improving service quality and user experience, a common practice is to deploy the delay-sensitive Internet of Things (IoT) control systems in a public edge cloud. However, this method is also faced with the challenge of serving the fluctuating resource demands through timely acquisition of enough instances. In this article, we systematically study the problem of proactively purchasing cloud resources for delay-sensitive IoT control systems over the public edge clouds which are under a flexible pricing model, so that the total cost can be minimized over the long run. We formulate the proactive cloud resource scaling cost minimization (PCRSCM) problem, in which we take the prediction, purchasing, and deployment cost into consideration. This problem can be proved to be NP-hard and we propose the proactive online instance purchase (POIP) algorithm to solve the problem. We prove that the competitive ratio of POIP is 2. We also evaluate the performance of POIP through real trace-driven simulations and real testbed. Our evaluation shows that POIP significantly reduces delay by more than 40% when compared to the reactive method and it costs about 20%, 25%, and 30% less than state-of-the-art methods, such as POLAR, cost minimization for provisioning virtual server, and ARMA prediction with on-demand priority, respectively.

Existing Cross Modal Hashing (CMH) methods are mainly designed for balanced data, while imbalanced data with long-tail distribution is more general in real-world. Several long-tail hashing methods have been proposed but they can not adapt for multi-modal data, due to the complex interplay between labels and individuality and commonality information of multi-modal data. Furthermore, CMH methods mostly mine the commonality of multi-modal data to learn hash codes, which may override tail labels encoded by the individuality of respective modalities. In this paper, we propose LtCMH (Long-tail CMH) to handle imbalanced multi-modal data. LtCMH firstly adopts auto-encoders to mine the individuality and commonality of different modalities by minimizing the dependency between the individuality of respective modalities and by enhancing the commonality of these modalities. Then it dynamically combines the individuality and commonality with direct features extracted from respective modalities to create meta features that enrich the representation of tail labels, and binaries meta features to generate hash codes. LtCMH significantly outperforms state-of-the-art baselines on long-tail datasets and holds a better (or comparable) performance on datasets with balanced labels.

Approaches predicting the results of mutation testing by machine learning have been proposed to reduce the cost of mutation testing. The predictive approaches based on PIE theory and approaches based on natural language have been proposed. However, both PIE-based and natural language-based approaches have disadvantages, leading to a reduction in effectiveness at the test case level prediction. In order to predict at the test case level and improve the effectiveness of prediction, we propose Natural Language and PIE Predictive Mutation Testing (NLPIE-PMT), which combines advantages of PIE-based and natural language-based approaches and predict whether each test case kills each mutant in the cross-version scenario. The experimental results on subjects in Defects4J show that NLPIE-PMT can predict whether each test case kill each mutant with the average F1-score of 0.811, which is 0.135 and 0.046 higher than the PIE-based baseline and the natural language-based baseline respectively. NLPIE-PMT also performs better than the baselines in predicting mutation score.

An online service system may experience various performance faults during operation. Detecting and locating these faults after they occur can significantly impact the user experience and lead to significant losses. Therefore, it is necessary to predict faults before they occur. Existing methods for fault prediction typically only predict the possibility of fault, without providing more granular predictions, such as the type of fault. This can make troubleshooting more difficult for developers. In this paper, we propose a fine-grained fault prediction method called FSFP, which not only predicts the possibility of fault but also identifies the type of fault that may occur. The method initially collects performance monitoring metrics from the runtime system, including two types: normal operation and abnormal conditions. It then utilizes cross-attention to capture the interdependencies between these two types of monitoring metrics, followed by the construction of a multi-label classification model. We evaluated FSFP by injecting faults into a benchmark microservice system. In terms of predicting the possibility of fault, FSFP achieved a precision of 0.999, a recall of 0.998, and an F1 score of 0.999. In terms of predicting the type of fault, FSFP achieved an exact match ratio of 0.955 and a Hamming loss of 0.017. In terms of predicting six specific types of faults, FSFP achieved four optimal F1 scores.

The purpose of drug recommendation is to predict the effective and safe drug combinations required for the current visit based on the historical medical data of patients. How to better mine the hidden relationship in the medical data and effectively improve the accuracy of drug recommendation are research hotspots in the medical field. This paper proposes a Collaborative Cross-attention Drug Recommendation model (CCDR) based on patient and medical relationship representations, which mines medical data from two aspects to enhance the representation ability of the model. CCDR obtains patient representation vectors by modeling the patients’ historical sequence data using Bidirectional Gated Recurrent Unit. Meanwhile, CCDR designs a medical graph structure data learning method based on relationship division to better capture the complex association relationships among diagnoses, procedures, and drugs. Finally, the representation capability of the model is enhanced by introducing a collaborative cross-attention mechanism to fuse the information obtained from both medical sequence and graph structure data. The experimental results show that the CCDR model can effectively improve the performance of drug recommendation.

Federated learning is a new type of artificial intelligence technology. During the training process, the client transmits model parameter information instead of local data to ensure their privacy and security. But it also incurs higher communication costs. This article proposes a new federated learning pruning method, FedADP, with the aim of adaptively determining pruning ratios for each layer in each client model without infringing on client privacy, and achieving more accurate pruning effects. Our method not only reduces communication costs during the training process, but also maintains accuracy similar to the original model. We conducted experimental validation using classic models and datasets, and evaluated our scheme and traditional federated learning scheme in terms of model accuracy, communication cost, and computational cost.

Human trajectories have explicit linkages to urban dynamics and functions. Therefore, learning urban region representations from human trajectories could help understand cities, and be used in various downstream tasks, e.g., land use classification and crime prediction, and ultimately assist city planning and management. However, previous works fall short in deeply mining the complete information embodied in human trajectories and neglect the data sparsity issue of some regions. We propose a multi-view graph contrastive learning (MVGCL) framework to learn urban region representations from human trajectories and the spatial adjacency between urban regions. First, we construct an outflow view based on human trajectories and devise data augmentation techniques to construct an inflow view. Then, we construct a spatial view based on the spatial adjacency between regions. Moreover, we use an MLP to extract spatial features from spatial view and model two graph encoders to learn complete region representation. Finally, we utilize a dual-multiplet loss function based on graph contrastive learning to maximize node consistency. Extensive experiments in Manhattan regions demonstrate that the performance of MVGCL outperforms the state-of-the-art methods.

As a new design mechanism for improving service quality and user experience, a common practice is to deploy the delay-sensitive services in a public cloud. However, this method is also faced with the challenge of serving the fluctuating resource demands through timely acquisition of enough instances. In this paper, we systematically study the problem of proactively purchasing cloud resources for delay-sensitive service providers over the public clouds which are under a flexible pricing model, so that the total cost can be minimized over the long run. We formulate the Proactive Cloud Resource Scaling Cost Minimization (PCRSCM) problem, in which we take the prediction, purchasing and deployment cost into consideration. This problem can be proved to be NP-hard and we propose the Proactive Online Instance Purchase (POIP) algorithm to solve the problem. We also evaluate the performance of POIP through real trace-driven simulations and real testbed. Our evaluation shows that POIP significantly reduces delay by more than 40% when compared to the reactive method and it costs about 20%, 25%, 30% less than state-of-the-art methods, such as POLAR, CMPVS and APOP, respectively.

Discovery the causal structure graph among a set of variables is a fundamental but difficult task in many empirical sciences. Reinforcement learning based causal discovery from observed data achieves prominent results. However, previous algorithms lack interpretability and efficiency, and ignore the prior knowledge of causal structure. To solve these problems, we propose GARL that leverages graph attention network to embed the structure information and the prior knowledge, and reinforcement learning to search the variable ordering with the best score. GARL takes the structure information and prior knowledge as the computational skeleton of attention to obtain the embedded representation of variables, and then generates variable orderings through the designed ordering model. In addition, the structure information is used to form the DAG corresponding to the variable ordering, which reduces the computational difficulty and improves the efficiency. GARL generates DAGs in the reinforcement learning framework, and uses the score of DAG as the reward to optimize the network structure to search the DAG with the best score. Experimental results on synthetic and real datasets show that our GARL has obvious advantages in multi-node operation efficiency, and competitive results with competitive baselines.

As auxiliary collaborative signals, the entity connectivity and relation semanticity beneath knowledge graph (KG) triples can alleviate the data sparsity and cold-start issues of recommendation tasks. Thus many works consider obtaining user and item representations via information aggregation on graph-structured data within Euclidean space. However, the scale-free graphs (e.g., KGs) inherently exhibit non-Euclidean geometric topologies, such as tree-like and circle-like structures. The existing recommendation models built in a single type of embedding space do not have enough capacity to embrace various geometric patterns, consequently, resulting in suboptimal performance. To address this limitation, we propose a KG-aware recommendation model with mixed-curvature manifolds interaction learning, namely CurvRec. On the one hand, it aims to preserve various global geometric structures in KG with mixed-curvature manifold spaces as the backbone. On the other hand, we integrate Ricci curvature into graph convolutional networks (GCNs) to capture local geometric structural properties when aggregating neighbor nodes. Besides, to exploit the expressive spatial features in KG, we incorporate interaction learning to ensure the geometric message passing between curved manifolds. Specifically, we adopt curvature-aware geodesic distance metrics to maximize the mutual information between Euclidean space and non-Euclidean spaces. Through extensive experiments, we demonstrate that the proposed CurvRec outperforms state-of-the-art baselines.

Predicting differential gene expression (DGE) from Histone modifications (HM) signal is crucial to understand how HM controls cell functional heterogeneity through influencing differential gene regulation. Most existing prediction methods use fixed-length bins to represent HM signals and transmit these bins into a single machine learning model to predict differential expression genes of single cell type or cell type pair. However, the inappropriate bin length may cause the splitting of the important HM segment and lead to information loss. Furthermore, the bias of single learning model may limit the prediction accuracy. Considering these problems, in this paper, we proposes an Ensemble deep neural networks framework for predicting Differential Gene Expression (EnDGE). EnDGE employs different feature extractors on input HM signal data with different bin lengths and fuses the feature vectors for DGE prediction. Ensemble multiple learning models with different HM signal cutting strategies helps to keep the integrity and consistency of genetic information in each signal segment, and offset the bias of individual models. Besides the popular feature extractors, we also propose a new Residual Network based model with higher prediction accuracy to increase the diversity of feature extractors. Experiments on the real datasets from the Roadmap Epigenome Project (REMC) show that for all cell type pairs, EnDGE significantly outperforms the state-of-the-art baselines for differential gene expression prediction.

Accurate diagnosis of cancer subtypes is crucial for precise treatment, because different cancer subtypes are involved with different pathology and require different therapies. Although deep learning techniques have made great success in computer vision and other fields, they do not work well on Lung cancer subtype diagnosis, due to the distinction of slide images between different cancer subtypes is ambiguous. Furthermore, they often over-fit to high-dimensional genomics data with limited samples, and do not fuse the image and genomics data in a sensible way. In this paper, we propose a hybrid deep network based approach LungDIG for Lung cancer subtype Diagnosis by fusing Image-Genomics data. LungDIG first tiles the tissue slide image into small patches and extracts the patch-level features by fine-tuning an Inception-V3 model. Since the patches may contain some false positives in non-diagnostic regions, it further designs a patch-level feature combination strategy to integrate the extracted patch features and maintain the diversity between different cancer subtypes. At the same time, it extracts the genomics features from Copy Number Variation data by an attention based nonlinear extractor. Next, it fuses the image and genomics features by an attention based multilayer perceptron (MLP) to diagnose cancer subtype. Experiments on TCGA lung cancer data show that LungDIG can not only achieve higher accuracy for cancer subtype diagnosis than state-of-the-art methods, but also have a high authenticity and good interpretability.

The rapid development of communication technology improves the feasibility of modern communication networks, such as wireless networks and satellite communication networks, etc. As a result, the complexity of networks increases significantly in terms of topology, increasing the difficulty of network resource allocation (i.e., network flow scheduling) and causing network congestion. It is worth noting that congestion is extremely destructive to the network, causing the network's overall performance to decline and even crashes and paralysis in severe cases. Therefore, timely identification and dealing with the network congestion are hot and difficult points of current research. Currently, most works (i.e., the active and passive measurement) identify the congestion on the node level, which may cause impractical issues with the increasing number of nodes (i.e., increasing detection probes will exacerbate the network congestion). To this end, this paper innovatively develops an active detection method based on the improved ant colony algorithm (IAC) for congestion identification on the path level. The developed method uses the IAC algorithm to find the optimal path and identifies the congested path by comparing the optimal path before and after congestion. Besides, the queuing theory model is utilized to obtain the performance metrics (i.e., queuing waiting time) to evaluate the optimal path. Experimental results show that the developed method can accurately identify the congested path in the network. This method makes identifying congestion less expensive and has higher operational feasibility, which is meaningful for the practical reference.

Partial multi-label learning (PML) models the scenario where each training sample is annotated with a set of candidate labels, but only a subset of them corresponds to the ground-truths. The key challenge for PML is how to minimize the negative impact of incorrect labels concealed within the candidate ones. Most existing PML solutions require abundant samples to train a noise-robust multi-label predictor. However, due to privacy, safety or ethic issues, we more often have a handful of training samples for the target task. In this paper, we propose an approach named FsPML-DA (Few-shot Partial Multi-Label Learning with Data Augmentation) to simultaneously estimate label confidence, perform data augmentation and induce multilabel classifier. Specifically, FsPML-DA disambiguates the label confidence vector of each PML sample by jointly modeling the feature and semantic similarity, label credibility of other samples and label co-occurrence. Next, FsPML-DA introduces a synthetic feature network to generate more training samples from pairs of given samples with label confidence values. FsPML-DA then leverages original and generated samples to train a noise-tolerant multi-label classifier. Extensive experiments on benchmark datasets show that FsPML-DA performs better than recent competitive PML baselines and few-shot solutions. FsPML-DA can dislodge noisy labels by mining PML data in a sensible way and the proposed data augmentation strategy effectively combats with the scarcity of few-shot training samples.

In recent years, there has been a lot of research on deep learning for survival prediction in EHR (Electronic Health Record). At present, EHR usually contains multivariate time series data with missing values. How to better predict mortality based on such data is what many studies are currently doing. Most of the mortality prediction methods based on deep learning only pay attention to missing value filling or adjusting the model structure to enhance the mortality prediction performance, but they do not combine these two aspects well. In this paper, we propose MTSSP (Multivariate Time Series for Survival Prediction), a new method that combines missing value filling and time series classification. It takes two representations of the loss pattern: the mask and the time interval, which are combined into the recurrent neural network for the interpolation of the missing value of patient characteristics. When the missing data values are interpolated, the model combines bidirectional RNN and one-dimensional CNN to jointly capture the patient's medical behavior from a global and local perspective to enhance the representation ability of data information, thereby improving the prediction accuracy of the model. In the end, we conducted our mortality prediction experiments on the real-world emergency MIMIC-III dataset and MIMIC-IV dataset. The experimental results demonstrate that the proposed approach has been shown to significantly outperform other approaches.

The explosive growth of machine learning (ML) methods is overloading users with choices for learning tasks. Method recommendation aims to alleviate this problem by selecting the most appropriate ML methods for given learning tasks. Recent research shows that the descriptive and structural information of the knowledge graphs (KGs) can significantly enhance the performance of ML method recommendation. However, existing studies have not fully explored the descriptive information in KGs, nor have they effectively exploited the descriptive and structural information to provide the necessary supervision. To address these limitations, we distinguish descriptive attributes from the traditional relationships in KGs with the rest as structural connections to expand the scope of KG descriptive information. Based on this insight, we propose the Cross-modal Knowledge Graph Contrastive learning (CKGC) approach, which regards information from descriptive attributes and structural connections as two modalities, learning informative node representations by maximizing the agreement between the descriptive view and the structural view. Through extensive experiments, we demonstrate that CKGC significantly outperforms the state-of-the-art baselines, achieving around 2% higher accurate click-through-rate (CTR) prediction, over 30% more accurate top-10 recommendation, and over 50% more accurate top-20 recommendation compared to the best performing existing approach.

Predicting differentially expressed genes (DEGs) from epigenetics signal data is the key to understand how epigenetics controls cell functional heterogeneity by gene regulation. This knowledge can help developing 'epigenetics drugs' for complex diseases like cancers. Most of existing machine learning-based methods suffer defects in prediction accuracy, interpretability or training speed. To address these problems, in this paper, we propose a Multiple Self-Attention model for predicting DEGs on Epigenetic data (Epi-MSA). Epi-MSA first uses convolutional neural networks for neighborhood bins information embedding, and then employs multiple self-attention encoders on different input epigenetics factors data to learn which locations of genes are important for predicting DEGs. Next it trains a soft attention module to pick out which epigenetics factors are significant. The attention mechanism makes the model interpretable, and the pure matrix operation of self-attention enables the model to be parallel calculated and speeds up the training. Experiments on datasets from the Roadmap Epigenome Project and BluePrint Data Analysis Portal (BDAP) show that the performance of Epi-MSA is better than existing competitive methods, and Epi-MSA also has a smaller standard deviation, which shows that Epi-MSA is effective and stable. In addition, Epi-MSA has a good interpretability, this is confirmed by referring its attention weight matrix with existing biological knowledge.

Abstract Nowadays recording and sharing personal lives using mobile devices on the Internet is becoming increasingly popular, and successive POI recommendation is gaining growing attention from academia and industry. In mobile scenarios, multiple influencing factors including the diversity of user preferences, the changeability of user behavior and the dynamic of spatiotemporal context bring great challenges to the POI recommender system. In order to accurately capture both the stable and the contextual preferences of mobile users in dynamic contexts, we propose a fusion framework JANICP (Joint Attention Networks with Inherent and Contextual Preferences) for successive POI recommendation by jointly training an offline/nearline user inherent interest perception model and an online user contextual interest prediction model. The offline model is trained based on the global historical behavior data to achieve stable interest representation, while the online model is trained based on the instantly selected context-sensitive data to achieve dynamic interest perception. An attention aggregation and matching module is used to fully connect the two kinds of preference representations and generate the final POI recommendation. Extensive experiments were conducted on three real datasets and experimental results show that the proposed JANICP outperforms existing state-of-the-art methods.

With the current raging spread of the COVID19, early forecasting of the future epidemic trend is of great significance to public health security. The COVID-19 is virulent and spreads widely. An outbreak in one region often triggers the spread of others, and regions with relatively close association would show a strong correlation in the spread of the epidemic. In the real world, many factors affect the spread of the outbreak between regions. These factors exist in the form of multimodal data, such as the time-series data of the epidemic, the geographic relationship, and the strength of social contacts between regions. However, most of the current work only uses historical epidemic data or single-modal geographic location data to forecast the spread of the epidemic, ignoring the correlation and complementarity in multi-modal data and its impact on the disease spread between regions. In this paper, we propose a Multimodal InformatioN fusion COVID-19 Epidemic forecasting model (MINE). It fuses inter-regional and intra-regional multi-modal information to capture the temporal and spatial relevance of the COVID-19 spread in different regions. Extensive experimental results show that the proposed method achieves the best results compared to state-of-art methods on benchmark datasets.

Partial multi-label learning (PML) aims at learning a robust multi-label classifier by training on ambiguous data, where each sample is associated with a set of candidate labels, among which only a subset are valid labels. A basic premise of existing PML solutions is to obtain enough partial multi-label samples for inducing the classification model. However, when dealing with new tasks, we may only have a few PML samples for those tasks. Furthermore, existing few-shot learning approaches assume the support (training) samples are precisely labeled; as such, irrelevant labels in the candidate label set may seriously mislead the meta-learner and thus result in a compromised performance. How to achieve PML with limited few-shot support samples is an important and practical problem, but not yet well studied. In this paper, we propose an approach called FsPML (Few-shot PML) to tackle this problem. Specifically, FsPML first performs adaptive distance metric learning via an embedding network using both sample features and label semantics in the embedding space. Next it rectifies the positive and negative prototypes of each new label of the target task in the embedding space. An unseen example can then be classified via its distances to the positive and to the negative prototypes. Experimental results on widely-used multi-label datasets (MS COCO and NUS-WIDE) demonstrate that our FsPML outperforms competitive baselines across different settings, and it can quickly generalize to new tasks with fewer training samples.

Partial-label learning (PLL) generally focuses on inducing a noise-tolerant multi-class classifier by training on overly-annotated samples, each of which is annotated with a set of labels, but only one is the valid label. A basic promise of existing PLL solutions is that there are sufficient partial-label (PL) samples for training. However, it is more common than not to have just few PL samples at hand when dealing with new tasks. Furthermore, existing few-shot learning algorithms assume precise labels of the support set; as such, irrelevant labels may seriously mislead the meta-learner and thus lead to a compromised performance. How to enable PLL under a few-shot learning setting is an important problem, but not yet well studied. In this paper, we introduce an approach called FsPLL (Few-shot PLL). FsPLL first performs adaptive distance metric learning by an embedding network and rectifying prototypes on the tasks previously encountered. Next, it calculates the prototype of each class of a new task in the embedding network. An unseen example can then be classified via its distance to each prototype. Experimental results on widely-used few-shot datasets demonstrate that our FsPLL can achieve a superior performance than the state-of-the-art methods, and it needs fewer samples for quickly adapting to new tasks.

With the development of internet technology, the proliferation of network-based applications leads to large number of multi-source multi-sink traffic transmission. Such as wireless sensor network (WSN), to deal with actuator nodes or support high-level programming abstractions, it naturally calls for a many-to-many communication. But the existing algorithms or solutions are not able to solve the scenarios effectively, they face many difficulties and challenges when dealing with multi-source multi-sink network problems. In this paper, we develop a new traffic scheduling algorithm suitable for multi-source and multi-sink networks. According to traffic transmission rate and network structure information, it selects the optimal path to transmit traffic and achieve load balance. When the traffic or source nodes change in the network, the paths and traffic are adjusted as needed to ensure the overall optimum. To evaluate the performance on efficiency, we perform a series of simulations and the results indicate the advantages of the proposed algorithm.

The huge number of machine learning (ML) methods has resulted in significant information overload. Faced with an overwhelming number of ML methods, it is challenging to select appropriate ones for the given dataset and task. In general, the names of ML methods or datasets are rather condensed, thus lacking specific explanations, while the rich latent relationships between ML entities are not fully explored. In this paper, we propose a description-enhanced machine learning knowledge graph-based approach - DEKR - to help recommend appropriate ML methods for given ML datasets. The proposed knowledge graph (KG) not only includes the connections between entities but also contains the descriptions of the dataset and method entities. DEKR fuses the structural information with the description information of entities in the knowledge graph. It is a deep hybrid recommendation framework, which incorporates the knowledge graph-based and text-based methods, overcoming the limitations of previous knowledge graph-based recommendation systems that ignore the description information. There are two key components of DEKR: 1) a graph neural network aggregating information from multi-order neighbors with attention to enrich the seed (i.e. dataset or method) node's own representation, and 2) a deep collaborative filtering network based on the description text to obtain the linear and nonlinear interactions of description features. Through extensive experiments, we demonstrated the efficiency of DEKR, which outperforms the current state-of-the-art baselines by a large margin.

Multi-instance multi-label (MIML) Active Learning (M2AL) aims to improve the learner while reducing the cost as much as possible by querying informative labels of complex bags composed of diverse instances. Existing M2AL solutions suffer high query costs for scrutinizing all relevant labels of MIML samples, querying excessive bag–label or instance–label pairs. To address these issues, a Cost-effective M2AL solution (CM2AL) is presented. CM2AL first selects the most informative bag–label pairs by leveraging uncertainty, label correlations, label space sparsity, and informativeness from queried instances of the bag, and thus avoids scrutinizing all labels. Next, it queries the most probably positive instance–label pairs of the selected bag–label pair. Particularly, if the feedback is positive, the bag is positively annotated with the label. For negative feedback, it further leverages the label of the neighborhood bags and the label of the nearby instances of queried instances of this bag, if the suggested labels from bag- and instance-levels disagree, CM2AL temporally gives up querying this bag–label pair and moves to another most informativeness one; otherwise, it takes the agreed label to annotate the bag, which further saves the cost by avoiding the excessive query. Extensive experiments on MIML data sets from diverse domains show that CM2AL can more reduce the cost while managing a better performance than state-of-the-art methods, the collaboration between bags and instances contributes to the saved cost.

Mining Electronic Health Records (EHRs) is of great significance to improve the efficiency and quality of medical services. In recent years, researchers have used deep recurrent neural networks to predict patients' next-period prescriptions. The main challenges of predicting next-period prescriptions are as follows: i) The latent interdependence information which can be utilized to enhance the representation of the data between heterogeneous features is dynamic. However, most existing approaches do not consider capturing and utilizing this information. ii) Conventional recurrent neural networks cannot store historical interdependent information between heterogeneous features and take advantage of it. To address these challenges, We propose a novel attention mechanism named Cross Attention (CA) that can capture the interdependence between two sequences. We further propose three types of recurrent neural networks that can capture and utilize current and historical latent interdependence between the two sequences. Extensive experiments on real-world data demonstrate that DSCA networks can capture the interdependence information between sequences and outperform state-of-the-art methods.

Nowadays, prediction for medical treatment migration has become one of the interesting issues in the field of health informatics. This is because the medical treatment migration behavior is closely related to the evaluation of regional medical level, the rational use of medical resources, and the distribution of medical insurance. Therefore, a prediction model for medical treatment migration based on medical insurance data is introduced in this paper. First, a medical treatment graph is constructed based on medical insurance data. The medical treatment graph is a heterogeneous graph, which contains entities such as patients, diseases, hospitals, medicines, hospitalization events, and the relations between these entities. However, existing graph neural networks are unable to capture the time-series relationships between event-type entities. To this end, a prediction model based on Graph Convolutional Network (GCN) is proposed in this paper, namely, Event-involved GCN (EGCN). The proposed model aggregates conventional entities based on attention mechanism, and aggregates event-type entities based on a gating mechanism similar to LSTM. In addition, jumping connection is deployed to obtain the final node representation. In order to obtain embedded representations of medicines based on external information (medicine descriptions), an automatic encoder capable of embedding medicine descriptions is deployed in the proposed model. Finally, extensive experiments are conducted on a real medical insurance data set. Experimental results show that our model's predictive ability is better than the best models available.

With the development of health insurance, the consequent problem of health insurance fraud has become increasingly serious. Outlier detection is a common method for health insurance anomaly detection. Researchers use some prior knowledge to assume patterns and indicators of interest. However, fraud pattern is concealed and changeable. The method of mining anomalies from fixed patterns is difficult to meet the current needs. In order to overcome this limitation, this paper combines the rich expression ability of heterogeneous information networks to model the complex relationships between entities, and establish a health insurance business representation model. The paper explores all possible business patterns, interrelated business portfolio patterns and related indicators in the field of health insurance. Considering the dynamic nature of a network, anomaly mining is carried out from both horizontal and vertical perspectives. Among them, the horizontal comparison adopts a fixed time period. The vertical comparison dynamically adjusts the time period according to the frequency of occurrence of the health insurance pattern instance, and then performs indicator calculation and outlier detection. Finally, the experimental results on the real data set show that our approach can narrow the scope of professional review and find more records of possible fraud than traditional methods.

Power demand response aims to clip the peak and fill the valley of power load by interruptible load management. Its closely related to the safety and economic benefit of the power system. At present, various provinces in China have carried out the pilot work of interruptible load management. The power companies signed contracts with power users to stipulate power users to adjust the power load consumption during peak hours or in emergency situations. The singed contracts should be fair for all power users. At the same time, the specific information about the signed contracts should be protected for privacy preservation, which is a typical federated learning scenario. A fair and privacy-preserving incentive pricing model should be proposed to stimulate power uses to participate in thee interruptible load management. To this end, a power demand response incentive pricing model is proposed. Based on the multi-attribute sealed auction game, the bidding model between the power company and the power users is established. In the single bidding model, the risk of the power user's demand response is evaluated firstly. The power users could be classified into different group according to their power load characteristics. Then a user classification selection algorithm' is proposed, which enables both the power company's revenue and power user risk to be considered, and enables the selected users to achieve balanced peak clipping. A fair mechanism based on integrals is proposed to assure the fairness between all power users. The case simulations demonstrate that effectiveness of the proposed incentive pricing model.

Diabetes is the 5thleading cause of death in the world, which threatens people's physical and mental health. However, in the course of studying diabetes, the missing medical data has caused great trouble for researchers. There are few researches focusing on sparse medical data. Common method of filling missing values is not desirable. In this study, we select and analyze a large physical check dataset containing more than 900 records with missing data. Using these data, a model for predicting diabetes is designed and implemented. Before predicting, logistic regression is utilized to remain the most significant attributes as the input variables of prediction model. Concept lattice is considered an effective mathematical tool for conceptual data analysis and knowledge processing in mathematics. Experimental results demonstrate that concept lattice illustrates great capacity of predicting diabetes while remaining missing data.

With the advancement and popularization of healthcare informatization, healthcare fraud has happened frequently. Doctors, patients and pharmacies are involved, and these various subjects cooperate with each other to form a complicated fraud relationship. Healthcare fraud has caused great damage to the safety of the health insurance fund. However, fraudster detection problem has not been solved properly at present. Therefore, Fraudster Detection Based on Modularity Optimization Algorithm (FDMOA) approach is developed to find doctor fraudsters in a good accuracy, which models the medical insurance records as a heterogeneous weighted network of doctors and medicines, and uses a modularity optimization algorithm to divide doctors and drugs into corresponding communities on a heterogeneous network. Finally, anomaly detection strategy was used in the community to find an abnormal doctor. Experiments show that the proposed method outperforms distance-based, clustering-based, and LOF algorithms.

In real-world, name ambiguity mainly arises when many people share the same name or express their names in the same way, which often causes erroneous aggregation of records of multiple persons with the same name. This name ambiguity problem deteriorates the performance of information retrieval in digital libraries, web search etc. It is nontrivial to distinguish those name references, especially when there is very limited information about them. Most existing studies uses features like email address, frequent words etc. However, the information is not always available because of privacy or too expensive to get. In this paper, we utilize a graph node embedding approach to solve author name disambiguation problem, where a graph is constructed only using the collaborator relationships. In the methodological aspect, the proposed method uses random walk and a graph node representation learning method to embed each node into a low dimensional vector space. Finally, we solve this problem by partitioning the records associated with a name reference such that each partition contains records pertaining to a unique real-world person. We evaluate our method on the real world CiteSeerX dataset, and the experimental results demonstrate that the proposed method is significantly better than most of the existing name disambiguation methods working in a similar setting.

Crowdsourcing emerges as a promising means of solution generation, which creates tremendous value by leveraging the intelligence of crowds in the web services. With the rise of the business of crowdsourcing services, both crowdsourcers and workers expect to gain better Quality of Experience, as well as more profits. But there is a contradiction between the incentives of the crowdsourcer and the quality of result of the crowds. In order to balance this conflict, the paper develops a profit optimization model for all parties in crowdsourcing by employing Tullock Contests. The model consists of two parts. Firstly, optimized incentives are provided by crowdsourcer to encourage workers to achieve a better quality of result. Secondly, an optimal fee schedule is provided as guidance to workers. The visualization of the equilibria of benefits is helpful to reach a win-win situation for crowdsourcer and crowds, which in turn impacts the development of crowdsourcing services. In addition, we simulate the acquisition of Nash-equilibrium as a repeated crowdsourcing task. The effectiveness of our model is testified by the experimental results.

Fraud detection is an important issue in the area of data science, and it has a lot of practical applications in related fields, such as business, health, and environment. Most traditional methods detect fraud based on rulemaking. Unfortunately, it is not always useful in the medical field since the boundary of fraud detection is vague. As a result, outlier detection is a promising method. This paper develops an outlier detection method of analyzing the correlation of patients to detect fraud. We construct a heterogeneous information network which bridges the medicines used and diseases of patients. In light of the network, we calculate the correlation score of different patients and design a discriminant rule. Through the discriminating rule, fraudulent patients represented by the abnormal nodes can be found. Our experiments use real medical insurance data sets and the results confirm that our method is accurate and effective.

Early detection of diseases is a tricky problem for doctors due to the concealment of the disease itself and the patient's negligence. Electronic health records (EHRs) contain a wealth of patient information that can be used to assist in the early detection of diseases. The EHRs have the characteristics of high dimensionality and timing. The traditional non-sequential method using EHRs is not accurate for the prediction of disease and the calculation efficiency is relatively low. We use previous medical records of patients based on recurrent neural networks (RNNs) to present a prediction model named “RNN-INFO” for multiple diseases. The model can predict the probability and diagnosis of clinical visits simultaneously in a time window. The model reduces the dimension of the data through the hidden layer of the neural network and extracts the hidden representation between the medical records. In addition, in order to further improve the performance of the model, we introduce demographic information about patients. After data preprocessing, the final dataset of experiment we used contains more than 28 thousand patients and 289 thousand clinical visits. Our experiments are finished with the final dataset. The experimental results show that our model has better performance than non-temporal models on both probability prediction and diagnostic prediction. “RNN - INFO” gets better performance than RNN model without demographic information on probability prediction. It shows that the addition of demographic information can indeed be improved performance of probability prediction.

Process frameworks contain the main logic of business. Frameworks extracting and filling can formulate a new process efficiently and accurately. Aiming at the construction guidance of new business processes with framework, a direct-viewing solution based on the similarity analysis of process diagrams is proposed. In this solution, the process diagrams are abstracted into process trees to get process sequences and then the similarity is calculated out of process skeletons and process entities, thus the processes which are similar both in structure and function can be found. The process trees and process sequences refer to simple data structures that are easy to analyse and enhance process comprehensibility. In addition, similarity measurement from both the skeletons and the entities can control process analysis from the overall situation and avoid information omission and deviation. Meanwhile, general process frameworks are extracted out for reference by clustering the processes through improving similarity cluster algorithm, and the instructive recommendation to the processes under construction can be given according to the existing processes. Based on this solution, economic expenditure can be reduced to a large extent and existing resources can be reused.

The rapid growth of data makes it possible for us to study human behavior patterns. Knowing the patterns of human behavior is of great use to help us detect the unusual fraud human behavior. Existing fraud detection methods can be divided into two categories: pattern based and outlier detection based methods. However, because of the sparsity and complex granularity of big data, these methods have high false positive in fraud detection. In this paper, we propose an effective hybrid fraud detection method. We propose SSIsomap which improves isomap to cluster behaviors into behavior classes and propose SimLOF which improves LOF to conduct outlier detection, then we use Dempster-Shafer evidence Theory for combining behavior pattern evidence and outlier evidence, which yields a degree of belief of fraud to the new coming claim. The experiment result shows our method has significantly higher accuracy than exsiting methods in medical insurance fraud detection.

With the tremendous growth of the Web, it has become a huge challenge for the single-process crawlers to locate the resources that are precise and relevant to some topics in an appropriate amount of time, so it is increasingly important to use the parallel crawler. However, due to the parallelism of crawlers, one headache problem we have to face is how to distribute the URLs to crawlers to make the parallel system work coordinately and thereby make sure that the Web pages fetched are of high quality. In this paper, a novel URL assignment model for the parallel crawler is described, which is based on multi-objective decision making method and considers multiple factors synthetically such as load balance, overlap and so on. Extensive experiments test and validate our techniques.

As an important supporting technology of Market Intelligence (MI), Web data integration is facing new challenges, such as the integrity of data acquisition, the quality of data extraction and data consolidation. To solve such problems, we propose an MI-oriented web data integration system (MI-WDIS), which achieves excellent performances in integrating Surface Web and Deep Web data with much less manual work. Based on MI-WDIS, we have developed a platform for intelligent analysis of job data. The platform collects tens of thousands of job data daily and provides personalized services for job seekers through diversified channels. Besides, it provides other advanced services, including intelligence analysis, automatic monitoring and alerting, for various organizations, such as enterprises, training institutions and recruitment agencies.