Shenbao Yu

As user behavior sequences grow increasingly diverse, sequential recommender systems (SRS), especially multimodal SRS, are highly sought after. Distinct from traditional item identifier (ID)-based SRS, multimodal SRS exploits multiple-modality data (e.g., text and images) to capture fine-grained user–item interactions, thereby enhancing recommendation efficacy. A common approach to training a multimodal SRS is to employ a pre-trained model for item representations in terms of text (or image) features. However, the existing pretrained models extract text and image features separately, which could result in visual-textual mismatches. In addition, the multimodal training objectives often suffer from a seesaw phenomenon (i.e., improving one objective’s performance may degrade others’ performance). Recognizing these issues, in this article, we propose a fine-tuned contrastive language-image with multi-gated expert sequential recommendation ( f CLI-MSR) framework, of which the main novelties lie in (a) an enhanced item representation by seamlessly integrating image–text matching information and (b) multimodal learning balance via carefully designed multi-gated expert networks and loss functions. We demonstrate the performance of the new framework over four public multimodal benchmark datasets, yielding improved item prediction performance over sequential baselines and mitigates the seesaw phenomenon.

In many advanced recommender systems (e.g., GNN-based), local smoothing mechanisms would distill collaborative signals but could inadvertently amplify targeted data poisoning threats. Existing defenses predominantly relied on rigid binary filtration strategies. However, such structural deletion of suspicious nodes often compromises graph connectivity, leading to a severe robustness-fidelity dilemma. To fill this gap, we propose SiPo (Silencing the Poison), a novel framework that shifts the paradigm from node removal to gradient silencing. SiPo leverages unsupervised granular ball computing to adaptively capture adversarial topological micro-clusters within the feature space. Subsequently, it geometrically collapses high-risk nodes towards a non-informative singularity via a global gravity gating mechanism. This operation blocks the backpropagation path of malicious gradients, effectively suppressing attacks while preserving the collaborative signals of normal users. Theoretical analysis confirms that SiPo enforces a compression of the local Lipschitz constant. Furthermore, extensive experiments on four public datasets demonstrate that SiPo not only significantly reduces attack success rates but also consistently improves recommendation performance. The implementation code is available at: https://github.com/KREML0/SiPo.

With the increasing complexity of networks across various domains driven by the advancement of informatization, community detection is no longer confined to disjoint communities but also extends to overlapping communities. Nowadays, efficient algorithms for overlapping community detection (OCD) are highly sought after. Classical Bayesian theory-based nonnegative matrix factorization methods for OCD have laid a solid foundation, but they still have limitations in adaptability to networks with varying overlap degrees. This article presents an adaptive Bayesian nonnegative matrix tri-factorization (ABNMTF) model for OCD. This model can adjust both the node-community interaction strength and the community influence, providing strong adaptability for detecting overlapping communities in networks with varying degrees of overlap. Moreover, iterative and efficient matrix updating rules are introduced to solve the ABNMTF probabilistic model. The superiority of ABNMTF lies in its intuitive architecture, efficient computation, high-precision soft classification, and stable detection results. Experimental results, from both real-world and Lancichinetti-Fortunato-Radicchi (LFR) synthetic networks, illustrate that ABNMTF generally outperforms eleven state-of-the-art OCD algorithms, thereby highlighting its powerful capabilities and stability in detecting overlapping communities.

Predicting student performance (PSP) is critical to intelligent tutoring systems in online education services. Accurate predictions enable data-driven decision making and facilitate the implementation of timely educational interventions. While previous approaches, such as cognitive diagnosis models and data mining techniques, have demonstrated effectiveness with small-scale and static datasets, they face significant challenges in large-scale online learning environments. In these settings, vast volumes of learning responses are continuously generated as students engage with exercises. Those responses are characterized by high dimensionality and incompleteness (HDI) and often manifest as streaming data, which limits the applicability of most existing prediction methods, therefore posing a new challenge to traditional PSP tasks. To remedy the void of PSP in the HDI and incremental scenario, we propose an incremental nonlinear co-latent factor analysis (IN-CoLFA) model, which enhances latent factor analysis – an element-wise learning framework – by incorporating a co-factorization technique and integrating a neural network-inspired structure. To facilitate incremental learning, we develop momentum-accelerated stochastic gradient-based algorithms, which enable the model to perform offline training on historical data and continuously refine its predictions as new student performance becomes available. Experiments on several real-world datasets demonstrate the efficacy and efficiency of our approach in both stationary and incremental PSP tasks under HDI conditions.