Selected work

Representative Papers

Time-Aware Diffusion based on Preference Disentanglement for Generative Recommendation

Jun 01, 2026

Existing diffusion-based generative recommender systems typically employ a uniform, static diffusion process over users’ historical interactions, failing to capture the non-stationary nature of preference evolution over time. To address this limitation, this work proposes the Temporal-aware Diffusion Preference Model (TDPM), which introduces, for the first time in generative recommendation, a time-aware diffusion mechanism. TDPM decouples user preferences into long-term stable periodic preferences and recent event-driven transient preferences via semantic index tokens, explicitly integrating both components into the diffusion process. Experimental results on three real-world datasets demonstrate that TDPM significantly outperforms state-of-the-art methods, achieving average improvements of 29.21% in HR@20 and 25.45% in NDCG@20.

0 citationsRead paper

Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommenders

May 16, 2026

Generative recommender systems exacerbate popularity bias due to suboptimal token-level optimization and undifferentiated item tokenization, leading to unfair and less diverse recommendations. To address this, this work proposes Ghost, a novel framework that jointly mitigates bias through asymmetric negative likelihood optimization and a skeleton-based semantic item tokenization mechanism, operating synergistically on both the optimization objective and representation structure. Experimental results demonstrate that Ghost significantly alleviates popularity bias across three benchmark datasets, achieving markedly improved fairness and diversity in recommendations while incurring only marginal degradation in overall recommendation performance.

0 citationsRead paper

Rel-MOSS: Towards Imbalanced Relational Deep Learning on Relational Databases

Mar 09, 2026

Entity classification in relational databases is often hindered by class imbalance, which compromises the performance on minority classes. This work presents the first systematic study of this issue and proposes a relation-centric oversampling framework based on graph neural networks. The approach models relational databases as heterogeneous graphs, introduces a relation-aware gating mechanism to adaptively aggregate neighbor information, and designs a relation-guided synthetic strategy for minority classes that preserves relational consistency in generated samples. Extensive experiments on 12 entity classification datasets demonstrate that the proposed method significantly outperforms existing techniques, achieving average improvements of 2.46% in balanced accuracy and 4.00% in G-Mean, thereby effectively enhancing the representation and classification capability for minority classes.

0 citationsRead paper

Detecting Non-Optimal Decisions of Embodied Agents via Diversity-Guided Metamorphic Testing

Dec 23, 2025

This work addresses the critical yet overlooked problem of “Non-optimal but Successful Planning” (NoD)—where embodied agents accomplish tasks successfully yet generate suboptimal plans—under resource-constrained settings, a flaw neglected by existing evaluation benchmarks. We formally define the NoD phenomenon and propose NoD-DGMT, a diversity-guided mutation testing framework. Methodologically, we introduce four novel mutation relations specifically designed to assess planning optimality, construct a behavioral invariance model to distinguish functional correctness from optimality violations, and integrate a diversity-driven test case selection strategy to balance coverage and detection efficiency. Evaluated on four state-of-the-art planning models in the AI2-THOR environment, NoD-DGMT achieves an average detection rate of 31.9%, outperforming the optimal baseline by 16.8 percentage points; the diversity-guidance mechanism further improves detection rate by 4.3 points and diversity score by 3.3 points.

0 citationsRead paper

Urban1960SatSeg: Unsupervised Semantic Segmentation of Mid-20$^{th}$ century Urban Landscapes with Satellite Imageries

Jun 11, 2025

Addressing severe degradation—including geometric distortion, spectral deficiency, and misregistration—as well as the lack of pixel-level annotations in mid-20th-century historical satellite imagery (e.g., Keyhole), this work introduces Urban1960SatBench, the first benchmark dataset for urban remote sensing from the 1960s. We further propose Urban1960SatUSM, an unsupervised semantic segmentation framework tailored to such challenging imagery. Its core innovations comprise: (i) a self-supervised architecture for joint geometric-spectral modeling of historical images; (ii) a confidence-aware spatial alignment mechanism to mitigate registration errors; and (iii) a focal confidence loss function designed to refine noisy pseudo-labels. Evaluated on the Urban1960SatSeg test set, our method achieves over 12% average accuracy improvement over state-of-the-art unsupervised approaches. This work establishes the first reproducible and robust technical foundation for quantitative visual analysis of long-term urban evolution using archival satellite data.

0 citationsRead paper

Recent publications

Latest Papers

Time-Aware Diffusion based on Preference Disentanglement for Generative Recommendation

Jun 01, 2026

Existing diffusion-based generative recommender systems typically employ a uniform, static diffusion process over users’ historical interactions, failing to capture the non-stationary nature of preference evolution over time. To address this limitation, this work proposes the Temporal-aware Diffusion Preference Model (TDPM), which introduces, for the first time in generative recommendation, a time-aware diffusion mechanism. TDPM decouples user preferences into long-term stable periodic preferences and recent event-driven transient preferences via semantic index tokens, explicitly integrating both components into the diffusion process. Experimental results on three real-world datasets demonstrate that TDPM significantly outperforms state-of-the-art methods, achieving average improvements of 29.21% in HR@20 and 25.45% in NDCG@20.

0 citationsRead paper

Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommenders

May 16, 2026

Generative recommender systems exacerbate popularity bias due to suboptimal token-level optimization and undifferentiated item tokenization, leading to unfair and less diverse recommendations. To address this, this work proposes Ghost, a novel framework that jointly mitigates bias through asymmetric negative likelihood optimization and a skeleton-based semantic item tokenization mechanism, operating synergistically on both the optimization objective and representation structure. Experimental results demonstrate that Ghost significantly alleviates popularity bias across three benchmark datasets, achieving markedly improved fairness and diversity in recommendations while incurring only marginal degradation in overall recommendation performance.

0 citationsRead paper

Rel-MOSS: Towards Imbalanced Relational Deep Learning on Relational Databases

Mar 09, 2026

Entity classification in relational databases is often hindered by class imbalance, which compromises the performance on minority classes. This work presents the first systematic study of this issue and proposes a relation-centric oversampling framework based on graph neural networks. The approach models relational databases as heterogeneous graphs, introduces a relation-aware gating mechanism to adaptively aggregate neighbor information, and designs a relation-guided synthetic strategy for minority classes that preserves relational consistency in generated samples. Extensive experiments on 12 entity classification datasets demonstrate that the proposed method significantly outperforms existing techniques, achieving average improvements of 2.46% in balanced accuracy and 4.00% in G-Mean, thereby effectively enhancing the representation and classification capability for minority classes.

0 citationsRead paper

Detecting Non-Optimal Decisions of Embodied Agents via Diversity-Guided Metamorphic Testing

Dec 23, 2025

This work addresses the critical yet overlooked problem of “Non-optimal but Successful Planning” (NoD)—where embodied agents accomplish tasks successfully yet generate suboptimal plans—under resource-constrained settings, a flaw neglected by existing evaluation benchmarks. We formally define the NoD phenomenon and propose NoD-DGMT, a diversity-guided mutation testing framework. Methodologically, we introduce four novel mutation relations specifically designed to assess planning optimality, construct a behavioral invariance model to distinguish functional correctness from optimality violations, and integrate a diversity-driven test case selection strategy to balance coverage and detection efficiency. Evaluated on four state-of-the-art planning models in the AI2-THOR environment, NoD-DGMT achieves an average detection rate of 31.9%, outperforming the optimal baseline by 16.8 percentage points; the diversity-guidance mechanism further improves detection rate by 4.3 points and diversity score by 3.3 points.

0 citationsRead paper

Urban1960SatSeg: Unsupervised Semantic Segmentation of Mid-20$^{th}$ century Urban Landscapes with Satellite Imageries

Jun 11, 2025

Addressing severe degradation—including geometric distortion, spectral deficiency, and misregistration—as well as the lack of pixel-level annotations in mid-20th-century historical satellite imagery (e.g., Keyhole), this work introduces Urban1960SatBench, the first benchmark dataset for urban remote sensing from the 1960s. We further propose Urban1960SatUSM, an unsupervised semantic segmentation framework tailored to such challenging imagery. Its core innovations comprise: (i) a self-supervised architecture for joint geometric-spectral modeling of historical images; (ii) a confidence-aware spatial alignment mechanism to mitigate registration errors; and (iii) a focal confidence loss function designed to refine noisy pseudo-labels. Evaluated on the Urban1960SatSeg test set, our method achieves over 12% average accuracy improvement over state-of-the-art unsupervised approaches. This work establishes the first reproducible and robust technical foundation for quantitative visual analysis of long-term urban evolution using archival satellite data.

0 citationsRead paper