Multimodal recommendation combines the user historical behaviors with the modal features of items to capture the tangible user preferences, presenting superior performance compared to the conventional ID-based recommender systems. However, existing methods still encounter two key problems in the representation learning of users and items, respectively: (1) the initialization of multimodal user representations is either agnostic to historical behaviors or contaminated by irrelevant modal noise, and (2) the widely used KNN-based item-item graph contains noisy edges with low similarities and lacks audience co-occurrence relationships. To address such issues, we propose MLLMRec, a novel preference reasoning paradigm with graph refinement for multimodal recommendation. Specifically, on the one hand, the item images are first converted into high-quality semantic descriptions using a multimodal large language model (MLLM), thereby bridging the semantic gap between visual and textual modalities. Then, we construct a behavioral description list for each user and feed it into the MLLM to reason about the purified user preference profiles that contain the latent interaction intents. On the other hand, we develop the threshold-controlled denoising and topology-aware enhancement strategies to refine the suboptimal item-item graph, thereby improving the accuracy of item representation learning. Extensive experiments on three publicly available datasets demonstrate that MLLMRec achieves the state-of-the-art performance with an average improvement of 21.48% over the optimal baselines. The source code is provided at https://github.com/Yuzhuo-Dang/MLLMRec.

Dynamic link prediction aims to predict the future links in dynamic graphs. Existing generative dynamic link prediction studies utilize the global degree distribution for mitigating the over-estimation problem, which can model the time-invariant features while neglecting the time-varying features, resulting in capturing inaccurate evolution patterns. However, such time related features are intrinsically coupled, which makes simultaneously and independently modeling both features infeasible. Motivated by these issues, we propose a Time-wise causal debiasing framework (Tide) for generative dynamic link prediction, which does not resort to any extra trainable modules. Instead, to obtain the time-invariant features, we first utilize a time-invariant deconfounded learning mechanism for decoupling the prediction score with the degree distribution. To leverage the time-varying features, we intervene in the model during the inference stage by a predicted future degree distribution, aiming to make the accurate predictions for dynamic graphs. Experiments conducted on four public datasets under both inductive and transductive settings present that our Tide enhanced models can outperform their corresponding vanilla versions by up to 21.42% and 27.73% in terms of NDCG and Jaccard, respectively.

The pervasive use of short-video applications has raised concerns about their potential negative effects on users, particularly addiction. Existing research often relies on psychological questionnaires, which lack real-world behavioral data, limiting scalability and analytical depth. To address this, we assess the addiction status of short-video platform users using a standardized psychometric questionnaire, combined with platform behavioral data and interview responses to uncover features associated with addiction. Using feature-based modeling, we scale to a dataset of 10,111 addiction-labeled users and identify key indicators of addiction, including prolonged daily watch time, especially at night, and excessive video consumption, while also revealing that higher watch frequency is not fully correlated with addiction. Additionally, we find that addicted users tend to consume a narrower range of content, suggesting a filter bubble effect. Our large-scale analysis provides valuable insights for platform designers, policymakers, and mental health professionals seeking to promote healthier engagement and mitigate the risks of short-video addiction.

Multi-modal recommenders (MRs) focus on leveraging the item modality features to facilitate user preferences modeling. Previous research mainly suffers from two limitations: (1) The pre-trained modality features are usually extracted by the encoders trained on general tasks (e.g., text classification), and thus inevitably contain the recommendation-irrelevant features. (2) Existing modality fusion mechanisms often diminish the contribution of features from weaker modalities, leading to biased fused representations. To address these challenges, we propose a novel Behavior Conditional Difussion model for Multi-Modal recommendation (BCDMM). Specifically, we first design a Behavior Multi-modal Diffusion (BMD) module to filter the recommendation-irrelevant noise within the pre-trained modality features. Then, we iteratively denoise the modality features with the guidance of user behavior signals to reconstruct the recommendation-related features. Next, we apply a Multi-modal Graph Fusion (MGF) module to explore the item modality latent structures. Moreover, we construct a modality fusion graph to capture the cross-modal complementary features for comprehensively modeling user preferences. Finally, a set of adversarial loss functions is used to balance the preservation of modality-specific and modality-shared features. Extensive experiments on three real-world datasets demonstrate the superiority of our method. We release our code at https://github.com/fanko79/BCDMM2025.

Dynamic link prediction aims to predict the future links on dynamic graphs, which can be applied to wide scenarios such as recommender systems and social networks on the World Wide Web. Existing methods mainly (1) focus on the in-graph learning, which cannot generalize to graphs unobserved during training; or (2) achieve the cross-graph predictions in a many-many mechanism by training on multiple graphs across various domains, which results in a large computational cost. In this paper, we propose a cross-graph dynamic link predictor named CrossDyG, which achieves the cross-graph transferability in a one-many mechanism which trains on one single source graph and test on different target graphs. Specifically, we provide causal and empirical analysis on the structural bias caused by the graph-specific structural characteristics in cross-graph predictions. Then, we conduct deconfounded training to learn the universal network evolution pattern from one single source graph during training. Finally, we apply the causal intervention to leverage the graph-specific structural characteristics of each target graph during inference. Extensive experiments conducted on three benchmark data of dynamic graphs demonstrate that CrossDyG outperforms the state-of-the-art baselines by up to 11.01% and 17.02% in terms of AP and AUC, respectively. In addition, the improvements are especially significant when training on small source graphs.

Dynamic graph link prediction is widely utilized in the complex web of the real world, such as social networks, citation networks, recommendation systems, etc. Recent Transformer-based link prediction methods on dynamic graphs not only fail to model the fine-grained structures such as triangles with the vanilla Transformers in the graph serialization process, but also amplify the imbalanced distribution of graphs because of their over-estimation of high-degree nodes. To tackle these issues, we propose a Topology-aware Transformer on Dynamic Graph (TopDyG) for link prediction, consisting of a topology injected Transformer (Ti-Transformer) and a mutual information learning (Mi-Learning). The Ti-Transformer explores the explicit structure of serialized graphs, capturing the topological features. The Mi-Learning mines the relationship between nodes by modeling the mutual information with a prior knowledge, alleviating the over-estimation of high-degree nodes when applying the Transformer-based models for the dynamic graph link prediction task. Extensive experiments on four public datasets containing both transductive and inductive settings present the superiority of our proposal. In particular, TopDyG presents an improvement of 43.27% and 28.75% against the state-of-the-art baselines in terms of NDCG and Jaccard, respectively. The advantages are especially obvious on the high-density graphs.

Dynamic graph learning on temporal networks aims to understand the continuous evolution pattern of networks, with an important application on forecasting the future temporal network. Existing methods mainly focus on modeling the structural and temporal features, with recent research interest shifting toward considering the structural correlations between nodes through their neighbor co-occurrences. Though satisfactory performance has been achieved, there still remain several limitations: (1) the deviation of investigated scenarios from real-world applications, since most previous researches concentrate on special cases of multigraphs with abundant repeat edges; (2) the insufficient computational efficiency of modeling the structural features, since the existing neighbor co-occurrence scheme fails to consider explicit structural correlations between nodes and suffers from a time-consuming pairwise encoding strategy; (3) the unsatisfying prediction accuracy due to inadequate modeling of temporal features, since each neighbor’s historical temporal features and the temporal domain shifting with network evolving are both neglected. To solve these issues, we first focus on the general scenarios of temporal networks without abundant repeat edges for approaching the actual applications and propose an efficient and effective dynamic graph learning method named LightDyG. Specifically, (1) on the one hand, to increase the computational efficiency, LightDyG decouples the structural correlations between nodes and their individual substructures for fast convergence based on the analysis of existing co-occurrence mechanism, and further designs an incremental strategy for efficient structural encoding; (2) on the other hand, to improve the prediction accuracy, the temporal characteristics are considered by including both the interaction and appearance timestamps of neighbors, and a time-invariant temporal encoding strategy is designed to eliminate the temporal bias introduced by the network evolution. Extensive experiments conducted on four public temporal networks demonstrate that LightDyG outperforms the best baselines by 4.54–11.39% and 6.06–16.24% in terms of AP and AUC on the temporal link prediction tasks, respectively. In addition, LightDyG reduces the time cost for training and test up to 45.91% and 63.94%, respectively, and also achieves a fast convergence speed during training. The implementation of our approach is available in https://github.com/nudtzpan/LightDyG .

Session-based recommendation aims to generate recommendations merely based on the ongoing session, which is a challenging task. Previous methods mainly focus on modeling the sequential signals or the transition relations between items in the current session using RNNs or GNNs to identify user’s intent for recommendation. Such models generally ignore the dynamic connections between the local and global item transition patterns, although the global information is taken into consideration by exploiting the global-level pair-wise item transitions. Moreover, existing methods that mainly adopt the cross-entropy loss with softmax generally face a serious over-fitting problem, harming the recommendation accuracy. Thus, in this article, we propose a Graph Co-Attentive Recommendation Machine (GCARM) for session-based recommendation. In detail, we first design a Graph Co-Attention Network (GCAT) to consider the dynamic correlations between the local and global neighbors of each node during the information propagation. Then, the item-level dynamic connections between the output of the local and global graphs are modeled to generate the final item representations. After that, we produce the prediction scores and design a Max Cross-Entropy (MCE) loss to prevent over-fitting. Extensive experiments are conducted on three benchmark datasets, i.e., Diginetica, Gowalla, and Yoochoose. The experimental results show that GCARM can achieve the state-of-the-art performance in terms of Recall and MRR, especially on boosting the ranking of the target item.

Session-based recommendation (SBR) , which mainly relies on a user’s limited interactions with items to generate recommendations, is a widely investigated task. Existing methods often apply RNNs or GNNs to model user’s sequential behavior or transition relationship between items to capture her current preference. For training such models, the supervision signals are merely generated from the sequential interactions inside a session, neglecting the correlations of different sessions, which we argue can provide additional supervisions for learning the item representations. Moreover, previous methods mainly adopt the cross-entropy loss for training, where the user’s ground truth preference distribution towards items is regarded as a one-hot vector of the target item, easily making the network over-confident and leading to a serious overfitting problem. Thus, in this article, we propose a Collaborative Graph Learning (CGL) approach for session-based recommendation. CGL first applies the Gated Graph Neural Networks (GGNNs) to learn item embeddings and then is trained by considering both the main supervision as well as the self-supervision signals simultaneously. The main supervisions are produced by the sequential order while the self-supervisions are derived from the global graph constructed by all sessions. In addition, to prevent overfitting, we propose a Target-aware Label Confusion (TLC) learning method in the main supervised component. Extensive experiments are conducted on three publicly available datasets, i.e., Retailrocket, Diginetica, and Gowalla. The experimental results show that CGL can outperform the state-of-the-art baselines in terms of Recall and MRR.

Session-based recommendation is a challenging task. Without access to a user's historical user-item interactions, the information available in an ongoing session may be very limited. Previous work on session-based recommendation has considered sequences of items that users have interacted with sequentially. Such item sequences may not fully capture complex transition relationship between items that go beyond inspection order. Thus graph neural network (GNN) based models have been proposed to capture the transition relationship between items. However, GNNs typically propagate information from adjacent items only, thus neglecting information from items without direct connections. Importantly, GNN-based approaches often face serious overfitting problems. We propose Star Graph Neural Networks with Highway Networks (SGNN-HN) for session-based recommendation. The proposed SGNN-HN applies a star graph neural network (SGNN) to model the complex transition relationship between items in an ongoing session. To avoid overfitting, we employ highway networks (HN) to adaptively select embeddings from item representations. Finally, we aggregate the item embeddings generated by the SGNN in an ongoing session to represent a user's final preference for item prediction. Experiments on two public benchmark datasets show that SGNN-HN can outperform state-of-the-art models in terms of [email protected] and [email protected] for session-based recommendation.

Session-based recommendation produces item predictions mainly based on anonymous sessions. Previous studies have leveraged collaborative information from neighbor sessions to boost the recommendation accuracy for a given ongoing session. Previous work often selects the most recent sessions as candidate neighbors, thereby failing to identify the most related neighbors to obtain an effective neighbor representation. In addition, few existing methods simultaneously consider the sequential signal and the most recent interest in an ongoing session. In this paper, we introduce an Intent-guided Collaborative Machine for Session-based Recommendation (ICM-SR). ICM-SR encodes an ongoing session by leveraging the prior sequential items and the last item to generate an accurate session representation, which is then used to produce initial item predictions as intent. After that, we design an intent-guided neighbor detector to locate the correct neighbor sessions. Finally, the representations of the current session and the neighbor sessions are adaptively combined by a gated fusion layer to produce the final item recommendations. Experiments conducted on two public benchmark datasets show that ICM-SR achieves a significant improvement in terms of Recall and MRR over the state-of-the-art baselines.

Session-based recommendation aims to predict a user's actions at the next timestamp based on anonymous sessions. Previous work mainly focuses on the transition relationship between items that the user interacted with during an ongoing session. They generally fail to pay enough attention to the importance of the items involved in these interactions in terms of their relevance to user's main intent. In this paper, we propose a Session-based Recommendation approach with an Importance Extraction Module, i.e., SR-IEM, that considers both a user's long-term and recent behavior in an ongoing session. We employ a modified self-attention mechanism to estimate item importance in a session, which is then used to predict user's long-term preference. Item recommendations are produced by combining the user's long-term preference and their current interest as conveyed by the last item they interacted with. Comprehensive experiments are conducted on two publicly available benchmark datasets. The proposed SR-IEM model outperforms start-of-the-art baselines in terms of Recall and MRR for the task of session-based recommendation. In addition, compared to state-of-the-art models, SR-IEM has a reduced computational complexity.