Institution profile

South China Agricultural University

Academic institutionasia · cn
Official website
Research library47linked papers
Opportunities0open roles
Selected work

Representative Papers

Freq-RemoteVAR: Next-Frequency Autoregressive Modeling for Remote Sensing Change Detection

Jul 28, 2026

This work addresses the limitations of existing remote sensing change detection methods, which typically rely on one-shot dense prediction and overlook the frequency characteristics of changes, leading to poor robustness under complex appearance variations and noise. To overcome this, the paper reformulates change detection as a structured generative task in the frequency domain and introduces a next-frequency autoregressive modeling paradigm that progressively generates change maps from coarse to fine scales. Key innovations include a frequency-aware masked tokenization strategy, a scale-aligned RoPE cross-attention module, and a change quality control mechanism. Built upon Fourier transform and quantized decomposition, the proposed Frequency VAR Transformer integrates dynamic normalization with spatial-frequency alignment. Extensive experiments on CDD, GZ-CD, and LEVIR-CD benchmarks demonstrate significant performance gains over state-of-the-art methods, particularly exhibiting superior robustness in complex scenarios.

0 citationsRead paper

Mitigating Matthew Effect: Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation

Jul 20, 2026

This work addresses the exacerbated Matthew effect in conversational recommender systems caused by dynamic user-item interactions. To mitigate popularity bias and enhance long-tail item coverage, the authors propose a novel multi-channel hypergraph architecture that jointly models items, knowledge graph entities, and words to construct a multi-granular representation of user interests. By integrating a multi-interest self-supervised learning mechanism, the approach effectively balances recommendation accuracy with fairness. Extensive experiments on four standard conversational recommendation benchmarks demonstrate that the proposed method consistently achieves state-of-the-art performance, significantly improving both recommendation fairness and effectiveness for long-tail items.

0 citationsRead paper

HyCoRec: Hypergraph-Enhanced Multi-Preference Learning for Alleviating Matthew Effect in Conversational Recommendation

Jul 19, 2026

This work addresses the Matthew effect in conversational recommender systems—where frequent interactions lead to overexposure of popular items and neglect of niche ones—by proposing a multi-preference representation learning framework that integrates five dimensions: items, entities, words, reviews, and knowledge. For the first time, hypergraph neural networks are introduced to model high-order correlations among these heterogeneous signals, enabling joint optimization of dialogue generation and recommendation prediction. Evaluated on two benchmark datasets, the proposed method significantly outperforms existing approaches, achieving superior recommendation accuracy while effectively mitigating the Matthew effect and balancing fairness with predictive performance.

0 citationsRead paper
Recent publications

Latest Papers

Freq-RemoteVAR: Next-Frequency Autoregressive Modeling for Remote Sensing Change Detection

Jul 28, 2026

This work addresses the limitations of existing remote sensing change detection methods, which typically rely on one-shot dense prediction and overlook the frequency characteristics of changes, leading to poor robustness under complex appearance variations and noise. To overcome this, the paper reformulates change detection as a structured generative task in the frequency domain and introduces a next-frequency autoregressive modeling paradigm that progressively generates change maps from coarse to fine scales. Key innovations include a frequency-aware masked tokenization strategy, a scale-aligned RoPE cross-attention module, and a change quality control mechanism. Built upon Fourier transform and quantized decomposition, the proposed Frequency VAR Transformer integrates dynamic normalization with spatial-frequency alignment. Extensive experiments on CDD, GZ-CD, and LEVIR-CD benchmarks demonstrate significant performance gains over state-of-the-art methods, particularly exhibiting superior robustness in complex scenarios.

0 citationsRead paper

Mitigating Matthew Effect: Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation

Jul 20, 2026

This work addresses the exacerbated Matthew effect in conversational recommender systems caused by dynamic user-item interactions. To mitigate popularity bias and enhance long-tail item coverage, the authors propose a novel multi-channel hypergraph architecture that jointly models items, knowledge graph entities, and words to construct a multi-granular representation of user interests. By integrating a multi-interest self-supervised learning mechanism, the approach effectively balances recommendation accuracy with fairness. Extensive experiments on four standard conversational recommendation benchmarks demonstrate that the proposed method consistently achieves state-of-the-art performance, significantly improving both recommendation fairness and effectiveness for long-tail items.

0 citationsRead paper

HyCoRec: Hypergraph-Enhanced Multi-Preference Learning for Alleviating Matthew Effect in Conversational Recommendation

Jul 19, 2026

This work addresses the Matthew effect in conversational recommender systems—where frequent interactions lead to overexposure of popular items and neglect of niche ones—by proposing a multi-preference representation learning framework that integrates five dimensions: items, entities, words, reviews, and knowledge. For the first time, hypergraph neural networks are introduced to model high-order correlations among these heterogeneous signals, enabling joint optimization of dialogue generation and recommendation prediction. Evaluated on two benchmark datasets, the proposed method significantly outperforms existing approaches, achieving superior recommendation accuracy while effectively mitigating the Matthew effect and balancing fairness with predictive performance.

0 citationsRead paper