Recommender systems may provide lower-quality recommendations to users with sparse interaction histories while overexposing head items. Existing two-sided fairness methods often share parameters across user- and item-side objectives, causing cross-side interference and unstable accuracy– fairness trade-offs. We propose Dual2Fair, a backbone-compatible in-processing framework with separate representation-calibration paths for activity-based user utility fairness and relevance-aware item exposure fairness. The user path transfers preference structure from interest prototypes learned from advantaged users to disadvantaged users through entropy-regularized optimal transport (OT). The item path calibrates long-tail representations toward a relevance-aware target combining rank-based merit with uniform smoothing and uses adaptive Nyström approximation to reduce transport cost. Both paths convert transport correspondences into confidence-aware barycentric corrections. An accuracy-first hierarchical alternating strategy preserves the native recommendation objective and restricts fairness corrections to the corresponding calibration mappings. Experiments on three public datasets and three backbones demonstrate favorable trade-offs among NDCG, user utility disparity, and item exposure disparity, together with stable optimization and practical computational savings from low-rank approximation. Our code is available at https://github.com/grw-zju/Dual2Fair.
Fake news on social media increasingly combines deceptive narratives with misleading images, creating verification needs that vary across samples. Most multimodal detectors nevertheless apply the same inference pipeline to every input, over-processing routine cases while providing insufficient reasoning for ambiguous claims. We propose DS-FND, a dual-system framework for multimodal fake news detection inspired by human System 1–System 2 decision-making. System 1 provides fast surface-level perception using textual, visual, image manipulation, cross-modal alignment, and text–image similarity features extracted with frozen pretrained encoders. System 2 fine-tunes the encoders and performs deeper semantic reasoning and evidence-aware verification through knowledge-enhanced scene graph interaction and dual-evidence verification. Two dynamic scheduling strategies, based on prediction confidence and task complexity, coordinate these pathways and selectively invoke the more costly verification process. Experiments on Twitter and Weibo show competitive detection performance and a favorable accuracy–efficiency trade-off. Calibration, fallback, ablation, and visualization analyses establish the complementary roles of the two systems, while inference measurements with pre-fetched evidence quantify the savings from selective verification.
Session-based recommendation (SBR) infers user interest from limited interaction sequences, yet current models show a consistent performance drop when longer histories are available. This degradation is associated with decoder designs that overemphasize recent items while underutilizing distant, yet semantically relevant, interactions, leading to poor modeling of long-range dependencies and multi-interest behaviors. To address this challenge, we introduce MIDA (Multi-Interest Decoding with Adaptive Aggregation for Session-based Recommendation), a lightweight and encoder-agnostic framework that provides a reformulation of the decoding stage. The central design principle is a two-space decoupling: clustering operates on raw item embeddings so that semantic groupings are not distorted by encoder-learned relational biases; encoder outputs are reserved for aggregation, where sequential context remains intact. Concretely, MIDA (i) employs differentiable multi-interest clustering on raw embeddings, preserving semantic structure that encoder representations would otherwise obscure, and (ii) fuses stability and plasticity branches through a gate conditioned on average cluster size, so that the inferred clustering structure directly governs aggregation behavior through a structural coupling across the two stages. Extensive experiments on multiple benchmark datasets demonstrate that MIDA consistently outperforms state-of-the-art methods, with meaningful accuracy improvements while remaining computationally practical.
With the success of large language models (LLMs) driving progress towards general-purpose AI, there has been a growing focus on extending these models to multi-modal domains, giving rise to multimodal large language models (MLLMs). Unlike existing reviews that focus on specific model frameworks or scenarios, this survey summarizes and provides insights into the current research on MLLMs from the more general perspective of input-output space extension. Particularly, we discuss the following questions: (i) How to construct multi-modal inputoutput spaces with discretely or continuously encoded modality signals? (ii) How to design model architectures and corresponding training strategies to align the constructed multi-modal representation space? (iii) How to evaluate MLLMs quantitatively based on the expanded input-output space? We hope to provide an intuitive and comprehensive overview and inspire future work.
Dynamic anchor-based subspace clustering with tensor norm regularization constitutes a leading approach to multi-view clustering. Dynamic anchors enhance the flexibility of structural descriptions, while tensor norms capture high-order correlations among views. However, from a representation learning perspective, dynamic anchors cannot correct erroneous positions using only subspace constraints, and from a multi-view fusion perspective, existing tensor norms overlook prior information related to the clustering task. To address these challenges, we first propose a spatial distance enhanced anchor strategy (SDE) that leverages the structural consistency inherent in both distance and subspace representations to guide anchors toward more reliable locations, thereby improving the reliability of the final representation. Subsequently, we construct a Cluster Information Fused Dynamic Tensor (IFDT) norm to adaptively incorporate prior-aware knowledge, achieving a consistent structure across views that adapts to the task's characteristics. By integrating these two modules in a mutually reinforcing manner, we propose a novel self-consistent multi-view subspace clustering algorithm that ensures intra-view representation consistency and inter-view structural consistency, named Consistency-driven Anchor Correction and Fusion Clustering (CACFC). Furthermore, theoretical analysis guarantees the convergence and efficiency of CACFC, and extensive experiments demonstrate superior accuracy and stability compared to 14 SOTA methods (Compared to the current best baseline, the accuracy increases by an average of two percent). Additional experiments explore the operational modes and scalability of SDE and IFDT.
