Low-Rank Adaptation (LoRA) enables parameter-efficient fine-tuning of Large Language Models (LLMs), and recent Mixture-of-Experts (MoE) extensions further enhance flexibility by dynamically combining multiple LoRA experts.However, existing MoE-augmented LoRA methods assume that experts operate independently, often leading to unstable routing, expert dominance.In this paper, we propose TalkLoRA, a communication-aware MoELoRA framework that relaxes this independence assumption by introducing expert-level communication prior to routing.TalkLoRA equips low-rank experts with a lightweight Talking Module that enables controlled information exchange across expert subspaces, producing a more robust global signal for routing.Theoretically, we show that expert communication smooths routing dynamics by mitigating perturbation amplification while strictly generalizing existing MoELoRA architectures.Empirically, TalkLoRA consistently outperforms vanilla LoRA and MoELoRA across diverse language understanding and generation tasks, achieving higher parameter efficiency and more balanced expert routing under comparable parameter budgets.These results highlight structured expert communication as a principled and effective enhancement for MoE-based parameterefficient adaptation.Code is available at https://github.com/why0129/TalkLoRA.

Contrastive learning has demonstrated strong performance in attributed hypergraph clustering. Typically, existing methods based on contrastive learning first learn node embeddings and then apply clustering algorithms, such as k-means, to these embeddings to obtain the clustering results.However, these methods lack direct clustering supervision, risking the inclusion of clustering-irrelevant information in the learned graph. To this end, we propose a Contrastive learning approach for Attributed Hypergraph Clustering (CAHC), an end-to-end method that simultaneously learns node embeddings and obtains clustering results. CAHC consists of two main steps: representation learning and cluster assignment learning. The former employs a novel contrastive learning approach that incorporates both node-level and hyperedge-level objectives to generate node embeddings.The latter joint embedding and clustering optimization to refine these embeddings by clustering-oriented guidance and obtains clustering results simultaneously.Extensive experimental results demonstrate that CAHC outperforms baselines on eight datasets.

Existing community deception methods focus on perturbing edges to hide target communities, but they often overlook the unnoticeability of perturbations. Such neglect makes the attacks susceptible to anomaly detection. Therefore, we propose the “unnoticeable community deception” problem, which aims to hide the target community while requiring perturbations to be unnoticeable. To address this, we introduce local community deception based on comprehensive modularity (LCDC). Specifically, LCDC first constrains the perturbation scope to ensure newly added edges are indistinguishable from existing inter-community edges, thereby enhancing unnoticeability. LCDC uses only local information within the perturbation scope, without requiring access to the entire network. It then employs a perturbation cost to guide the selection of edges for addition or deletion. This cost is formulated as a weighted combination of comprehensive modularity and a node-level tendency to depart. Comprehensive modularity combines global and local components: global modularity ensures the overall network structure within the perturbation scope remains stable before and after perturbation, while local modularity aims to disperse nodes from the target community into other groups. LCDC only accesses the target community and its neighboring communities, not the entire network. Experimental results show that LCDC outperforms baseline methods in terms of unnoticeability and is competitive with them in terms of hiding effectiveness.

Accurately capturing user preferences across diverse domains is a fundamental challenge in cross-domain recommendation (CDR) systems. Recent literature has established that disentangling user preferences into global and domain-specific components significantly enhances recommendation performance. However, existing CDR systems are hindered by two critical challenges: 1) how to align user representations across domains to accommodate the discrepancy of user intents in different domains; and 2) how to mitigate data conflicts caused by cross-domain heterogeneity. To address these limitations, we propose intent-aware contrastive learning for cross-domain recommendation (ICCDR). This framework learns unified user interests across domains while mitigating data conflicts from heterogeneity. The ICCDR framework consists of the following core components: 1) a cross-intent module that dynamically identifies users’ consistent intents across domains to guide the transfer of knowledge; and 2) an information bottleneck-based aggregation module designed to mitigate data conflicts by filtering out task-irrelevant information, thereby retaining only the essential latent features required for effective knowledge transfer. Extensive experiments conducted on Amazon datasets demonstrate that the ICCDR framework outperforms traditional baseline methods.

Heterogeneous self-supervised graph learning has gained considerable attention in recommender systems for its ability to capture diverse semantic and structural relationships in real-world data. Contrastive learning enhances representation learning by maximizing agreement between positive pairs while distinguishing negative ones in cross-views. However, two key challenges remain: 1) noise, such as false negatives, that degrades representation quality; and 2) lack of cross-view alignment causes biased and inconsistent representations. To address these challenges, we propose heterogeneous neighborhood-enhanced graph contrastive learning for recommendation (HNGCL). HNGCL ensures cross-view consistency through alignment and uniformity losses, encouraging embeddings that are both well-aligned and uniformly distributed across views, thereby enhancing generalization and discriminative power. To mitigate noise, HNGCL introduces a neighborhood-enhanced strategy that integrates collaborative neighbors to generate high-quality positive pairs, reducing false negatives and suppressing noise propagation. By leveraging heterogeneous graph structures and cross-view contrastive learning, HNGCL effectively captures intricate semantic and structural patterns, producing robust feature representations. Extensive experiments on real-world datasets demonstrate that HNGCL significantly outperforms state-of-the-art methods in recall and normalized discounted cumulative gain (NDCG), showcasing its effectiveness in overcoming these challenges and advancing recommendation performance. Our code for the model implementation is available athttps://github.com/zhangchi107/HNGCL.

Low-rank adaptation (LoRA) has been developed as an efficient approach for adapting large language models (LLMs) by finetuning two low-rank matrices, thereby reducing the number of trainable parameters.However, prior research indicates that many of the weights in these matrices are redundant, leading to inefficiencies in parameter utilization.To address this limitation, we introduce Dense Low-Rank Adaptation (DenseLoRA), a novel approach that enhances parameter efficiency while achieving superior performance compared to LoRA.DenseLoRA builds upon the concept of representation fine-tuning, incorporating a single Encoder-Decoder to refine and compress hidden representations across all adaptation layers before applying adaptation.Instead of relying on two redundant low-rank matrices as in LoRA, DenseLoRA adapts LLMs through a dense low-rank matrix, improving parameter utilization and adaptation efficiency.We evaluate DenseLoRA on various benchmarks, showing that it achieves 83.8% accuracy with only 0.01% of trainable parameters, compared to LoRA's 80.8% accuracy with 0.70% of trainable parameters on LLaMA3-8B.Additionally, we conduct extensive experiments to systematically assess the impact of DenseLoRA's components on overall model performance.Code is available at https://github.com/mulin-ahu/DenseLoRA.

Event Extraction involves extracting event-related information such as event types and event arguments from context, which has long been tackled through well-designed neural networks or fine-tuned pre-trained language models. These approaches require substantial annotated data for tuning parameters and are resource-intensive. Recently, Prompting strategies with frozen parameters, such as Chain-of-Thought and Self-Consistency, have delivered success in NLP using LLMs by generating intermediate thought steps. However, they suffer from the challenge of error propagation and lack of interaction between different thoughts. In this paper, we proposeNeural Network-based Prompting(NetPrompt), a novel network-structured prompting strategy for event extraction. The core idea behind NetPrompt is to imitate the excellent information integration capabilities of neural network structures. Specifically, we first decompose the event extraction problem into diverse intermediate subtasks, and each subtask is represented as a node in different layers of the network, the output of the nodes in the preceding layer is fed into the subsequent layer. Secondly, we propose pruning strategies to adapt the reasoning overhead to different problems. Finally, we have conducted extensive experiments on two widely used event extraction benchmarks to evaluate NetPrompt. The results demonstrated that NetPrompt significantly improved the event extraction performance compared to previous methods.