Large Language Models (LLMs) have shown strong potential for text-attributed graph (TAG) learning, yet effectively integrating LLM semantics with graph structural information remains challenging.Embeddings obtained from frozen LLMs lack topology awareness, while fine-tuning LLMs is often computationally expensive.Moreover, LLM embeddings are highdimensional, and naively reducing dimensionality tends to destroy semantics.To address these issues, we propose GASE, a framework for learning Graph-Aware Semantic Embeddings using frozen LLMs.GASE consists of two key stages: First, we introduce a Training-Free Structure-Aware Semantic Extraction (TSSE) module.Through inter-layer semantic feedback and progressive masked attention, it efficiently compresses and propagates semantic context from neighboring nodes without updating LLM parameters.Second, we propose a Subspace Decomposition and Structural Injection (SDSI) strategy.Embeddings obtained from TSSE are decomposed into a semantic-rich subspace and a structural injection subspace, and structural signals are injected into the latter, which preserves original semantics while integrating graph information.Experiments demonstrate that GASE outperforms state-of-the-art baselines on node classification and achieves a 5× speedup over fine-tuning-based methods.

Multi-step reasoning remains challenging for language models with limited capacity.While recent reasoning distillation approaches transfer chain-of-thought supervision from large teacher models, they typically apply uniform supervision across all Transformer components, overlooking the fact that different modules contribute unequally to reasoning.We propose Module-Aware Reasoning Distillation, a parameter-efficient framework that explicitly targets key Transformer components for effective reasoning transfer.Through systematic analysis, we identify the feed-forward network projections and the output projection of self-attention as primary bottlenecks for reasoning.Based on these findings, we introduce lightweight adapter modules at these components while freezing the backbone parameters, enabling focused and efficient distillation.Our approach adopts an offline distillation setting, where a strong teacher model provides reasoning trajectories in advance, and incorporates an adaptive supervision strategy that adjusts the strength of reasoning-related losses according to problem difficulty.Experiments on mathematical reasoning benchmarks demonstrate consistent improvements over strong baselines, and ablation studies confirm the importance of both module-aware placement and adaptive supervision.

Recommender systems are critical for mitigating information overload, assisting users in uncovering their latent interests, and enhancing their overall experience. Sequential recommendation leverages users' historical interaction sequences to predict dynamic interests more effectively than traditional rec-ommendation approaches. However, existing models-including RNN-based and Transformer-based methods-face significant limitations. RNNs struggle with vanishing gradients and long-term dependency capture, while Transformers, though effective for long-range relationships, suffer from computational inefficiency due to their quadratic attention complexity. Recent advancements have employed contrastive learning for sequential recommendation, aiming to enhance the consistency between augmented views and improve self-supervised learning signals. Despite their promise, these methods often lack diversity in data augmentation strategies, which restricts their capacity for bias mitigation, resulting in augmented data that still retains inherent biases. To address these challenges, we propose MDEC, a novel sequential modeling framework that leverages State Space Models (SSM) combined with unbiased contrastive learning. MDEC utilizes Mamba to efficiently model user preferences as an alternative to Transformer-based models. Additionally, it integrates graph-based information, including item transition and co-interaction data, to improve data augmentation comprehensively. Finally, we introduce adaptive anchor-enhanced contrastive learning, which adaptively utilizes augmented samples to improve representation quality and bias mitigation. Extensive experiments on multiple datasets demonstrate that MDEC significantly out-performs existing models, showcasing improved efficiency, better mitigation of biases, and enhanced recommendation quality. Code is available at https://github.com/Echohuangyan/CSLP.

The goal of Emotion Recognition in Conversations (ERC) is to accurately identify the emotions expressed in each utterance within a dialogue. Despite advancements made by current ERC methods, particularly those using RNN-based and GCN-based models to capture emotional dynamics and model speaker relationships, there remain two primary limitations: first, an insufficient integration of multiple feature representations and commonsense knowledge, which hampers the model's ability for deep emotional understanding; and second, the reliance on a single cross-entropy loss for classification optimization, which restricts the accuracy and robustness of emotion recognition. To address these issues, we propose a method for ERC using Multilayer Feature Fusion and Joint Loss Optimization (MFFJL). This approach combines contextual information, speaker dependency, and commonsense knowledge features by extracting feature vectors through RoBERTa and COMET, utilizing bidirectional LSTM and attention mechanisms to capture conversational context, and applying a cross-fusion module to deeply integrate various features, thus enhancing comprehension of complex emotional expressions. Additionally, the feature classification module incorporates joint cross-entropy and KL divergence optimization, further improving classification accuracy and consistency. Experimental results demonstrate the effectiveness of our method, as evidenced by superior performance on the IEMOCAP and MELD datasets. Our code is available at https://anonymous.4open.science/r/MFFJL-D9D8.

Multimodal dialogue emotion recognition integrates data from multiple modalities to accurately identify emotional states in conversations. However, differences in expression and information density across modalities complicate the fusion of features. Traditional methods may introduce redundant information from other utterances, reducing the accuracy of emotion recognition. Existing one-hot labels often fail to capture the full range of emotional expressions, leading to biased results. To address these issues, we propose a model that fuses different modalities within the same utterance to avoid redundancy. It employs a progressive classification process, refining emotion recognition from coarse to fine granularity. Additionally, we use emotion polarity probabilities as weights for fine-grained classification and introduce a multimodal information-rich label that considers both the data and their interactions. Experiments on IEMOCAP and MELD datasets demonstrate the model’s effectiveness, significantly improving dialog emotion recognition accuracy. Our code is available at https://anonymous.4open.science/r/LOCG-188E.

Emotion recognition in conversation (ERC) aims at accurately identifying emotional states expressed in conversational content. Existing ERC methods, although relying on semantic understanding, often encounter challenges when confronted with incomplete or misleading semantic information. In addition, when dealing with the interaction between emotional and semantic information, existing methods are often difficult to effectively distinguish the complex relationship between the two, which affects the accuracy of emotion recognition. To address the problems of semantic misdirection and emotional cross-talk encountered by traditional models when confronted with complex conversational data, we propose a semantic and emotional dual channel (SEDC) strategy for emotion recognition in conversations to process emotional and semantic information independently. Under this strategy, emotion information provides an auxiliary recognition function when the semantics are unclear or lacking, enhancing the accuracy of the model. Our model consists of two modules: the emotion processing module accurately captures the emotional features of each utterance through contrastive learning, and then constructs a dialogue emotion propagation map to simulate the emotional information conveyed in the dialogue; the semantic processing module combines an external knowledge base to enhance the semantic expression of the dialogue through knowledge enhancement strategies. This divide-and-conquer approach allows us to more deeply analyze the emotional and semantic dimensions of complex dialogues. Experimental results on the IEMOCAP, EmoryNLP, MELD, and DailyDialog datasets show that our approach significantly outperforms existing techniques and effectively improves the accuracy of dialogue emotion recognition.

The objective of sequential recommendation is to predict user preferences for items based on historical interaction sequences. This process often leads to a phenomenon known as popularity bias, where popular items are excessively recommended. Conformity, the tendency of users to follow popular items, is a significant factor contributing to this issue. Previous methods have not adequately disentangled conformity and interest, failing to accurately model users’ true intent. To address this, we propose a novel Disentangled Interest and Conformity Sequential Recommendation method (DICSRec) to mitigate the popularity bias. Specifically, we first design an Intent Encoding Module (IEM), which includes two independent encoders for conformity and interest to model their representations. To better disentangle these two factors, we design a disentangling task with proxy-based self-supervised learning and orthogonal regularization. Furthermore, to provide the Intent Encoding Module with more global information, we design a Global Conformity-aware Module (GCM), which supplies item popularity information and aids in enhancing user conformity representation. Lastly, recognizing the varying significance of user conformity and interest, we propose an adaptive Fusion Prediction Module (FPM) that adaptively aggregates user conformity and interest representations for final prediction. Experiments on four real-world datasets consistently demonstrate the superiority of our method over advanced sequential recommendation models. Code implementation is available at: https://github.com/lyra0611/DICSRec.

Sequential Recommender Systems (SRS), leveraging the temporal information from users' behaviors, have noticeably improved user experience against traditional systems. However, these behaviors often follow long-tail distribution, making the systems biased towards popular items (i.e., popularity bias). Moreover, popularity bias would amplify the neglect of long-tail recommendations, thereby sharpening the long-tail problem. Previous researches usually address these challenges independently, focusing on reducing the over-recommendation of popular items or enhancing the representation quality of tail items. Indeed, it is possible to incorporate their merits to achieve the best of both worlds. Thus, we propose a novel and unified framework, named Collaborative Solution to Long tailed problem and Popularity bias (CSLP), to tackle both the long-tail problem and popularity bias simultaneously. To achieve this, we first introduce a representation enhancement module featuring dual generators to enhance user and item representations, particularly for those in the tail. On the other hand, a debiasing module incorporating an Inverse Propensity Score (IPS) with a clipping strategy is introduced to further alleviate the popularity bias. Specifically, this clipping strategy demonstrates a clear decrease in the original IPS method's variance, effectively improving the recommendation for stability and accuracy. Experiments on three widely-used datasets show CSLP's effectiveness in solving both issues. CSLP surpasses all baselines (traditional, popularity bias, and long-tail problem) in overall performance, significantly enhancing recommendation accuracy for both tail users and items, and achieving a more balanced ratio of recommendations between popular and tail items. Code is available at https://github.com/Echohuangyan/CSLP.

Personalized News Recommendation (PNR) can enhance user experience by alleviating information overload. Traditional news recommendation methods consider all clicking behaviors as user interests, resulting in biased user modeling that fails to accurately capture user interests. In addition, although there are methods to reduce the impact of low-quality news at the representation level by simply filtering it through the attention mechanism. However, this only implicitly models the news in the interaction sequence without specifically considering the quality of each news, and thus has very limited effect in identifying noise. To address these issues, this paper proposes News Recommendation method based on News Quality-aware modeling (NQNR). We attempt to explicitly model the news in the click sequence and candidate ranking one by one to visually assess the quality of each news. Specifically, we design a detection module to detect whether the input news is low-quality news. Then, by reducing the influence of low-quality news in user modeling and candidate ranking, user interests are modeled more accurately, while recommendations of such news are reduced for users. In addition, to capture the similarity of vectors more accurately, we also design a similarity computation method based on the multiple attention mechanism in the detection module. Experiments on a large real-world Microsoft News Dataset (MIND) show that our model significantly outperforms previous models. Our code is posted at the following URL: https://github.com/xxbbjj/NQNR-.