Social recommendation systems leverage both user-user (u-u) social relations and user-item (u-i) collaborative interactions to improve recommendation quality. Despite their effectiveness, existing models often struggle with task-irrelevant information and misalignment between social and collaborative signals and the downstream recommendation task, leading to suboptimal performance. To address these limitations, we propose a novel framework for Effective and Consistent Information Extraction for Social Recommendation (ECSR). Our approach focuses on two key modules: (1) a task-irrelevant information discarding module that filters out noisy signals from both social relations and user-item interactions, and (2) a task-relevant information alignment module that captures both shared and view-specific task-relevant information, ensuring alignment with the recommendation objective. By integrating them into a unified form, our method extracts minimal and sufficient statistics, which significantly enhance the model's ability to predict user preferences. We validate ECSR on three real-world social recommendation datasets, demonstrating that it consistently outperforms state-of-the-art baselines.

Designing effective negative sampling strategies is crucial for training Multi-Modal Recommendation (MMRec) models, as it helps address the issues of sparse user-item interactions and facilitates the learning of high-dimensional modality features. However, most existing methods randomly sample non-interacted items as negative ones, which frequently result in easy negatives. They limit the model's ability to accurately capture user preferences. In this paper, we propose to Generate Difficulty-aware Negative Samples via conditional diffusion for MMRec (denoted as GDNSM). Leveraging the rich semantic and contextual information from multi-modal features, our method generates hard negative samples with varying difficulty levels, tailored to user preferences. They force the model to learn finer-grained distinctions between positive and negative samples, enhancing its ability to inferring user preferences. To avoid unstable training, we design a dynamic difficulty scheduling mechanism that schedules the negative samples from easy to hard for model training, ensuring both stability and effectiveness. Extensive experiments on three real-world datasets demonstrate that the effectiveness of our models.

The integration of social information into recommender systems (RSs) has gained significant popularity for enhancing recommendation performance and user experience. However, this practice introduces substantial privacy risks, particularly concerning the leakage of sensitive social relationships. While prior research has primarily focused on user-level and interaction-level privacy risks, the social relation-level privacy risks remain largely unexplored. To fill this gap, we investigate social privacy risks through membership inference attacks (MIA) and propose a Social relation-level MIA (SMIA) framework. Two key challenges arise: (1) the adversary can only access the recommended item IDs, which provide indirect and limited information about social relationships, and (2) extracting socially relevant preferences from recommendation results is inherently difficult. To tackle the first challenge, we leverage shadow models to transform sparse item IDs into dense features, enabling adversaries to effectively utilize recommendation outputs. For the second challenge, SMIA employs a dual-branch learning approach that disentangles social and behavioral preferences. Therefore, we can extract socially relevant signals from the disentangled preferences.Extensive experiments on real-world datasets demonstrate that both social and general RSs are highly vulnerable to such attacks, highlighting the urgent need for robust privacy protection mechanisms. To defend against these attacks, we introduce a Socially Adversarial Learning (SAL) defense mechanism that selectively obscures sensitive social information in user representations during training, effectively reducing privacy leakage. We further evaluate the effectiveness of our defense and discuss future directions for developing privacy-preserving mechanisms in social RSs.

User cold-start recommendation presents a significant challenge for recommender systems, affecting their overall effectiveness. Meta-learning-based methods have been introduced to address this issue. These methods treat the user cold-start recommendation problem as a few-shot learning task, where each user represents a unique task. The objective is to acquire shared initialization parameters that can be effectively applied across all cold-start users. Subsequently, these shared parameters are fine-tuned into personalized parameters using individual interaction data. Recent studies argue that shared parameters are unsuitable for all users with an implicit grouping distribution of user preference. Therefore, they propose adaptive-initialization-based methods, which first differentiate tasks based on user preferences and then generate task-adaptive initialization parameters using task representations. However, both the meta-learning and adaptive-initialization-based manners ignore discovering the adaptive capability of update strategies in the process of transferring initialization parameters to personalized parameters. Instead, they rely on task-shared optimization strategies, leading the model to fall into an overfitting or underfitting situation. In response to this, we propose a dual-adaptive update strategies-enhanced meta-optimization framework (DAUS) for user cold-start recommendation. First, we integrate dual-adaptive update strategies to enhance the adaptive capability of transferring initialization parameters. This involves incorporating both task-adaptive optimization hyperparameters and objectives. Second, we design a multifaceted task encoder , which can provide diverse task information to differentiate between tasks, including explicit task features (task relevance, training signals) and other implicit task information. Extensive experiments based on three real-world datasets demonstrate that our DAUS outperforms the state-of-the-art methods. The source code is available at https://github.com/XuHao-bit/DAUS .

Graph-based social recommendation improves the prediction accuracy of recommendation by leveraging high-order neighboring information contained in social relations. However, most of them ignore the problem that social relations can be noisy for recommendation. Several studies attempt to tackle this problem by performing social graph denoising, but they suffer from 1) adaptability issues for other graph-based social recommendation models and 2) insufficiency issues for user social representation learning. To address the limitations, we propose a model-agnostic graph denoising module (denoted as MADM) which works as a plug-and-play module to provide refined social structure for base models. Meanwhile, to propel user social representations to be minimal and sufficient for recommendation, MADM further employs mutual information maximization (MIM) between user social representations and the interaction graph and realizes two ways of MIM: contrastive learning and forward predictive learning. We provide theoretical insights and guarantees from the perspectives of Information Theory and Multi-view Learning to explain its rationality. Extensive experiments on three real-world datasets demonstrate the effectiveness of MADM.

Next Point-of-Interest (POI) recommendation is an essential part of the flourishing location-based applications, where the demands of users are not only conditioned by their recent check-in behaviors but also by the critical influence stemming from geographical dependencies among POIs. Existing methods leverage Graph Neural Networks with the aid of pre-defined POI graphs to capture such indispensable correlations for modeling user preferences, assuming that the appropriate geographical dependencies among POIs could be pre-determined. However, the pre-defined graph structures are always far from the optimal graph topology due to noise and adaptability issues, which may decrease the expressivity of learned POI representations as well as the credibility of modeling user preferences. In this paper, we propose a novel Adaptive Graph Representation-enhanced Attention Network (AGRAN) for next POI recommendation, which explores the utilization of graph structure learning to replace the pre-defined static graphs for learning more expressive representations of POIs. In particular, we develop an adaptive POI graph matrix and learn it via similarity learning with POI embeddings, automatically capturing the underlying geographical dependencies for representation learning. Afterward, we incorporate the learned representations of POIs and personalized spatial-temporal information with an extension to the self-attention mechanism for capturing dynamic user preferences. Extensive experiments conducted on two real-world datasets validate the superior performance of our proposed method over state-of-the-art baselines.

Sequential recommendation which aims to predict a user's next interaction based on his/her previous behaviors, has attracted great attention. Recent studies mainly employ deep recurrent neural networks or self-attention networks to capture dynamic user preferences. However, existing methods merely focus on modeling users' clear interests in interacted items. We argue that for an interaction, the user may also have ambiguous interests in items that are semantically related to the interacted one. For comprehensively capturing user preferences, it is beneficial to discover potential interests from historical interactions at a broader itemset level. Therefore, in this paper, we propose a knowledge graph enhanced sequential recommendation model namely KGIE, which focuses on enhancing user interest modeling with knowledge-enriched itemsets by incorporating the knowledge graph. Specifically, in addition to item-level interest modeling with interacted items, we further construct knowledge-enriched itemsets that are extracted via high-order knowledge associations with the interacted items. For capturing personalized itemset-level interests, we design an attentive aggregation unit to combine item embeddings considering both inherent and contextual personalization signals. Furthermore, to balance the contributions of both two levels of interest modeling, we adaptively learn high-level preference representations with a gating fusion unit. Extensive experiments on three real-world datasets demonstrate the superior performance beyond state-of-the-art methods and recommendation interpretability of our model.