Large language models (LLMs) have transformed recommender systems through strong semantic understanding and generalization. However, the design of item identifiers remains a critical bottleneck that directly affects recommendation quality. Traditional metadata-based identifiers introduce length variability and semantic ambiguity, whereas existing collaborative indexing (CID) approaches often neglect item attributes, show limited cross-dataset generalizability, and incur high computational cost at scale. To address these limitations, we propose a Graph Neural Network (GNN)–based item indexing framework with three coordinated innovations. First, we construct attribute-enriched co-occurrence graphs and use a GNN encoder to fuse item features with collaborative signals, yielding semantically informed representations that work well for attribute-rich catalogs. Second, we replace recursive spectral clustering with hierarchical agglomerative clustering on GNN embeddings, enabling direct control of index length via tree depth and reducing hyperparameter tuning across datasets. Third, we exploit localized message passing rather than global eigendecomposition, which provides considerably better runtime efficiency and is amenable to mini-batch training, supporting online index updates as interactions evolve. Across five benchmarks, GID achieves strong average ranking performance, showing larger improvements on sparse and attribute-rich datasets while remaining competitive in dense settings. The framework is robust under both seen and unseen prompt templates, which supports practical LLM-based recommendation. On sequential recommendation, GID improves HR@10 by 7.9% on average over the strongest baseline in each dataset.

In location-based social networks, next point of interest (POI) recommendation predicts the next-visited POIs of users by mining their behavioral patterns. However, existing POI recommendation methods based on graph neural networks and attention mechanisms fail to adequately capture: 1) the local structural features influenced by the trajectories of other users (i.e., the relationships between POIs visited by users); and 2) the dynamic visitation preferences and channel relationships among POIs (i.e., interdependencies between contextual features). We propose a next-POI recommendation model based on graph structure and sequential pattern to address these limitations. The model generates graph representations that reflect the real-time preferences of users by extracting local structures from a global POI graph. In addition, we design a temporal-aware self-attentive graph convolutional network and a channel attention mechanism to capture the structural features of users’ sequential visitation tendencies and the latent feature-channel relationships between POIs, respectively. These components enhance the ability of the model to characterize dynamic user preferences and behavioral changes. The results demonstrate that our model outperforms baseline methods on three real-world datasets, validating its effectiveness in capturing both global and local information.

Accurate prediction of pedestrian trajectories is essential for safety in autonomous driving. Pedestrian movements are shaped by individual behaviors, interactions with surrounding agents, and implicit social relationships. The primary challenge lies in effectively modeling these intricate social interactions while integrating temporal dependencies. Existing approaches typically aggregate agent states but often fail to account for hidden social relationships and trajectory multimodality. To address these challenges, this article proposes the social temporal multihead attention gated recurrent unit (Social-TMAGRU), a model that integrates two key submodules: social multihead attention (S-MHA) and temporal social attention GRU (TSA-GRU). The S-MHA submodule captures implicit social relationships between pedestrians by modeling their latent interactions, while TSA-GRU handles temporal dependencies and evolving dynamics in pedestrian trajectories. Experimental evaluations of multiple public datasets demonstrate that the proposed model significantly outperforms existing state-of-the-art methods in predictive accuracy and computational efficiency. The model excels in complex and high-density environments, demonstrating robust performance in dynamically changing scenarios, making it highly suitable for real-world autonomous driving applications.

Graph Neural Networks (GNNs) have been widely adopted in recommendation systems. When integrated into a federated learning framework, GNNs can enhance the model’s expressive capability. However, challenges arise in personalized representation and graph expansion due to the heterogeneity and locality of user data in federated recommendation systems. To address these challenges, we propose a federated recommendation model based on personalized attention and privacy-preserving dynamic graphs. The method first matches neighbor users for each selected client. Subsequently, it counts the interaction frequencies of items for both local and neighbor users to construct personalized weights, which captures the unique characteristics of different users. Additionally, we designs a method for constructing privacy-preserving dynamic graphs. In each round of federated training, the selected client adds pseudo-interaction items to its own interaction subgraph, perturbing the real interactions. After completing local training, the noisy interaction subgraph is incorporated into the global graph to capture higher-order connectivity information among users while safeguarding their interaction privacy. We conduct extensive experiments on three benchmark datasets, and the results demonstrate that the proposed PADG method achieves superior performance while effectively protecting privacy.

Cross-domain recommendation (CDR) advances recommendation accuracy through knowledge migration from auxiliary domains to a target domain. Traditional CDR methods require centralized storage of user data for training, which poses potential privacy risks. Additionally, when transferring knowledge from multiple domains, finding a transfer strategy that balances the contributions from various source domains is crucial. To overcome these limitations, this work introduces AWT-FCDR, a federated learning approach for cross-domain recommendation that employs adaptive weight transfer. First, user embeddings are learned within each domain through a federated learning framework with deep matrix factorization. Second, adaptive weights for cross-domain users are generated based on rating timestamps, rating ratios, and user rating similarity. Then, consensus encryption is employed to transfer the weighted user embeddings, which are then fused with target domain embeddings. Finally, the transferred user embeddings are combined with item embeddings from the target domain for rating prediction and recommendation. Comprehensive evaluations on three publicly available datasets validate the superior performance of our approach over conventional cross-domain recommendation techniques as measured by HR@10 and NDCG@10 metrics.

Federated learning, owing to its distributed characteristic, is particularly susceptible to data poisoning attacks. To address this issue, a wealth of defenses has been developed that aim to mitigate such attacks by limiting the adverse effects of malicious models on the aggregated result. However, most existing defense methods are designed only from the perspective of data filtering or model weighting, which leads to poor robustness and an exclusive reliance on a single server-side defense mechanism. To mitigate these vulnerabilities, we propose FedD2, a federated defense framework that integrates data augmentation and residual-based weighted aggregation. Firstly, FedD2 applies data augmentation techniques to regenerate and mix local datasets, enhancing data generalization. During the server aggregation stage, FedD2 leverages model residuals to detect abnormal updates and adaptively assign aggregation weights to local models, thereby reducing the influence of malicious clients. Comprehensive experiments performed on benchmark datasets demonstrate that FedD2 significantly enhances classification performance and exhibits strong robustness against data poisoning attacks.

To address the challenge of non-independent and identically distributed data in federated learning, clustering federated learning extracts local model features from clients and groups clients with high local model similarity into the same cluster to optimize the global model for the corresponding data distribution. However, current federated clustering algorithms bring additional communication cost for the accuracy of cluster division, which increases the communication burden of some performance-limited edge devices. To address the communication efficiency issue of clustering federated learning, we proposed an optimized communication efficiency federated average clustering learning that uses time weights to select partial clients to participate in training randomly. At the same time, a time-asynchronous similarity calculation method is proposed to improve the accuracy of local model similarity calculation for randomly selected clients. Extensive experimental evaluations show that our federated average clustering learning can achieve or even surpass the model accuracy of existing federated clustering algorithms. In the communication evaluation experiment, we can achieve the specified model accuracy using 5% to 30% of the communication of existing federated clustering algorithms.

In the modern interconnected world, the popularization of networks and the rapid development of information technology led to the increasing security risks and threats in network systems. The existing intrusion detection system is constantly challenged by various malicious intrusion attacks. Machine learning algorithms have been widely used in intrusion detection. However, the model training requires the support of a sufficient high-quality samples, especially attack traffic data. Network intrusion detection datasets may not be shared between organizations due to data security and some privacy policy concerns. The federated learning framework is an optimal approach to address this issue, in which organizations collaborate to train a global model shared by multiple parties while keeping the data local to the client, guaranteeing the data privacy and security of all parties. However, there is a problem of class imbalance in the network traffic data owned by the organizations, which seriously affects the detection performance of the model and leads to a high consumption of model training time. Therefore, this study proposed a novel federated undersampling learning framework with Gini impurity, namely Fed-UGI. The framework is based on the hash-based block undersampling method to rebalance the client, which can solve the influence of imbalanced training data on the model detection performance and improve the model training efficiency. Moreover, the client weighted aggregation strategy based on Local Gini impurity can further optimize the effect of global model aggregation and reduce the impact of the dispersion degree and information difference in client data on model aggregation. In addition, extensive experiments on intrusion detection datasets show that compared to SOTA methods, the proposed Fed-UGI method has a good detection effect on the three metrics of F1-score, G-mean and AUC, the training time of the model is reduced by 51.76%-92.58%, especially in highly class imbalance situation.

Most current session-based recommendations model session sequences solely based on the user's target behavior, ignoring the user's hidden preferences in auxiliary behaviors. Additionally, they use ordinary graphs to model one-to-one item correlations in the current session and fail to leverage other sessions to learn richer higher-order item correlations. To address these issues, a multi-behavior hypergraph contrastive learning model for session-based recommendations is proposed. This model represents all the sessions as global hypergraphs according to two types of behavior sequences. It employs contrastive learning to obtain global item embeddings, which are further aggregated to generate a global session representation that captures higher-order correlations of items from all session perspectives. A novel local heterogeneous hypergraph is designed for the current session to capture higher-order correlations between items with different behaviors in the current session, thus enhancing the local session representation. Additionally, a novel self-supervised signal is created by constructing a multi-behavior line graph, enhancing the global session representation. Finally, the local session representation, global session representation, and global item embedding are used to learn the predicted interaction probability of each item. Extensive experiments are conducted on three real datasets, and the results demonstrate that the proposed model significantly improves recommendation accuracy.

Trajectory-outlier detection can be used to discover the fraudulent behaviour of taxi drivers during operations. Existing detection methods typically consider each trajectory as a whole, resulting in low accuracy and slow speed. In this study, a trajectory outlier detection method based on group division is proposed. First, the urban vector region is divided into a series of grids of fixed size, and the grid density is calculated based on the urban road network. Second, according to the grid density, the grids were divided into high- and low-density grids, and the code sequence for each trajectory was obtained using grid coding and density. Third, the trajectory dataset is divided into several groups based on the number of low-density grids through which each trajectory passes. Finally, based on the high-density grid sequences, a regular subtrajectory dataset was obtained within each trajectory group, which was used to calculate the trajectory deviation to detect outlying trajectories. Based on experimental results using real trajectory datasets, it has been found that the proposed method performs better at detecting abnormal trajectories than other similar methods.

To address the problem of data sparsity in recommendation systems, various studies have used knowledge graphs as auxiliary information. These studies have employed multitask learning (MTL) to enhance recommendation performance. However, the shared information between tasks is not fully explored when using an MTL strategy for training both recommendation and knowledge graph-related tasks. Moreover, most studies cannot effectively model the knowledge sharing, consequently affecting recommendation performance. In response to these problems, we proposed a novel knowledge graph-based personalized multitask enhanced recommendation model. To explore the shared information between tasks, a relation attention mechanism was proposed to distinguish the relative importance of neighborhood information to the central entity. Additionally, we utilized a lightweight graph convolutional network to more effectively aggregate high-order neighborhood information from the knowledge graph. This approach improves the accuracy of neighborhood feature and ensures that more suitable shared information is obtained. Furthermore, we developed a linear interaction component to model knowledge sharing between recommendation and knowledge graph embedding tasks. This component allows for detailed feature interaction learning between items and entities, enhancing the shared feature representation, generalization capabilities, and overall performance of the recommendation system. The experimental results on three public datasets indicate that our model outperforms other benchmark models in CTR prediction and top-$\boldsymbol{K}$recommendation.

Traditional recommendation systems require users to upload local data to the server to generate recommendation results. In this process, users’ privacy is easy to disclose. Federated recommendations can solve the problem of local data privacy leakage, but the intermediate computing results are not protected. The existing work mainly protects the information in this process through encryption or disturbance schemes, which will result in complex calculations or low accuracy. Also, the two schemes only protect the rating data and process parameter information, but the existence information is not protected. Aiming at the above problems, this article proposes a federated matrix factorization based on secret sharing (FMFSS) to protect users’ privacy. The parameters are randomly divided into pieces, and then, the secret sharing technology is used to transmit private information between user–user and user–server, which does not introduce additional encryption cost and ensures value privacy, process privacy, and existence privacy. In addition, this article introduces the user–item interaction value, which is transmitted to the server with gradient information. In this way, the real gradient of the user cannot be inferred from the information received by the server from all parties, but the aggregated final average parameter information is real. The experimental comparison with the existing work and the analysis of computation time show that the proposed method can ensure the accuracy of model privacy recommendations without introducing additional encryption operations.

In the current matrix factorization recommendation approaches, the item and the user latent factor vectors are with the same dimension. Thus, the linear dot product is used as the interactive function between the user and the item to predict the ratings. However, the relationship between real users and items is not entirely linear and the existing recommendation model of matrix factorization faces the challenge of data sparsity. To this end, we propose a kernelized deep neural network recommendation model in this article. First, we encode the explicit user-item rating matrix in the form of column vectors and project them to higher dimensions to facilitate the simulation of nonlinear user-item interaction for enhancing the connection between users and items. Second, the algorithm of association rules is used to mine the implicit relation between users and items, rather than simple feature extraction of users or items, for improving the recommendation performance when the datasets are sparse. Third, through the autoencoder and kernelized network processing, the implicit data are connected with the explicit data by the multilayer perceptron network for iterative training instead of doing simple linear weighted summation. Finally, the predicted rating is output through the hidden layer. Extensive experiments were conducted on four public datasets in comparison with several existing well-known methods. The experimental results indicated that our proposed method has obtained improved performance in data sparsity and prediction accuracy.

Sequential recommendation models analyze users’ historical interactions to predict the next item they will en gage with. In order to better capture users’ dynamic interest preferences, most existing sequential recommendation models that introduce heterogeneous time intervals lead to increased model complexity, which raises computational costs and training difficulty. This is particularly evident in long sequential data, where the model need to handle a large variety of different time intervals. Additionally, accurately modeling the impact of long time intervals on user behavior remains a significant challenge. To address these issues, we propose a lightweight sequential recommendation algorithm with time interval awareness augmen tation (TALSAN). This model introduces a novel uniform data augmentation operator to improve the distribution of original data samples and employs a time-aware self-attention layer to model user interactions, maintaining the continuity of the original sequence. By integrating temporal context with posi tional features, TALSAN constructs a streamlined self-attention network for predicting user behavior. Comparative testing on datasets such as ML-100K, ML-1M, Amazon Beauty, Amazon Toys, and Amazon Fashion demonstrates the model’s superiority over existing baselines. Our results confirm that TALSAN not only mitigates cold start issues but also enhances the ability to learn user preferences, leading to improved prediction accuracy.

In the postmodern era of tourism, tourists’ behavior has undergone a substantial change and the demand of customized experience dominates the tourism market. The traditional single‐objective travel route recommendation method fails to meet the multiobjective needs of users. To handle this problem, a multiobjective hybrid tabu search algorithm for urban travel route recommendations is proposed in this paper. First, the rating and level of attractions, as well as the corresponding information of hotels and restaurants within a certain radius, are considered. Then, based on this information, a crowdsensing scoring method is established. Second, the hybrid particle swarm genetic optimization algorithm is exploited to generate a single‐object route, and the fast nondominated Pareto sorting algorithm is exploited to find the optimized solution. Then, the hybrid tabu algorithm is used to optimize the personalized route chosen according to multiple objects set by users. This algorithm combines the global search ability of the genetic algorithm and the neighborhood search ability of the tabu algorithm to prevent convergence from falling into a local optimum. Finally, the experiments are conducted on the real‐world data collected from the Dianping and Ctrip web sites. The comparison with baseline algorithms indicates that the algorithm proposed in this paper provides accurate and reasonable route recommendations for users.

The recommendation system (RS) predicts user ratings by collecting user information, but the users’ private information may be exposed in this process. Thus, it is crucial to achieving a balance between recommendation performance and privacy-preserving of RSs. Aiming to solve the above problem, this article proposes a novel matrix factorization (MF) algorithm. The algorithm predicts the user rating through a linear weighting of global average rating, item average rating, user average rating, and MF, which improves the prediction accuracy. Then, based on the above algorithm, this article proposes a MF RS for preserving user privacy by using local differential privacy technology. In this algorithm, the rating data are normalized on the user side to reduce global sensitivity. Then, Laplace noise is added to sensitive data before it is sent to the aggregator. Finally, based on the disturbed data, rating prediction is realized by using the MF algorithm. The proposed method is compared with five well-known recommendation methods on four public datasets. The experimental results show that the proposed algorithm achieves better recommendation performance at the same privacy-preserving level.

Abstract Spatial query is an important supporting technology in the Internet of Things (IoT) and location‐based services (LBS). The k‐nearest neighbor query is widely used for spatial queries. However, user location privacy may be leaked in the query. In addition, some users are malicious or uncooperative. With the objective of overcoming these problems, a k‐nearest neighbor query method based on trust and location privacy protection is proposed. First, we employ a new K‐anonymity method based on cooperation to protect a query user's location privacy. In this method, the query user constructs an anonymous group by introducing a trust mechanism to incentivize cooperation among users. Then, according to the different radii of the selection area set by the query user, agent users with higher reputation values who send query requests for the query user are selected. Finally, the agent users obtain the query results from the LBS server and forward them to the query user, and the query user screens the results according to his or her real location. The experiments show that our method can effectively stimulate users to cooperate, better exclude malicious users, and improve the accuracy of the query results while protecting the privacy of the query user.

Abstract Aiming at overcoming problems such as malicious entities and trust crises in cloud environments, a trust management model based on mutual trust and a reward‐with‐punishment mechanism is proposed in this paper. First, according to reputation values of entities and trust relationships among entities, we calculate the comprehensive trust values of a service request user to several candidate service providers and probe these providers with high comprehensive trust values. Second, we select a trade provider according to mutual trust. Finally, after the trade, we update the reputation values and other data of the request user, trade provider, and recommenders based on the final trade result to reward or punish them to various degrees, thus ultimately reducing malicious or dishonest behavior. The experimental results show that our model can effectively identify malicious entities to increase the trade success rate.

In this poster, we propose a novel, effective tuple-based anonymization technique for categorical data over the Internet. By utilizing a new structure, called Candidate Encoding Sequence with Frequency, and a set of new rules for generating such a sequence for each domain value of the categorical data, we can effectively solve the key limitation of the existing methods. Our experimental results demonstrate the superiority of our method against the existing method in terms of the strength of privacy protection.

Summary Aiming at preventing the privacy disclosure of sensitive information, issues related to privacy protection in cloud computing have attracted the interest of researchers. To protect the privacy of users during clustering in a cloud computing environment, we present a privacy‐preserving density peak clustering (PPDPC) algorithm that neither discloses personal privacy information nor leaks the cluster centers. Our scheme contains two steps of density peak clustering: First, a cloud service provider calculates the cluster centers without knowing each participant's private data and without disclosing any cluster center information to the other participants, and second, participant allocation is secure and every participant is prevented from identifying the other members of the same cluster. Security analysis and comparison experiments show that the proposed PPDPC algorithm not only obtains good accuracy with respect to density peak clustering but also resists collusion attacks even if the cloud service provider is collaborating with all except one participant. Both theoretical analysis and experimental results confirm the security and accuracy of our method.