Selected work

Representative Papers

Improving Item Discoverability in e-Commerce Search via Related Intent Generation

Jul 29, 2026

This work addresses the limitations of traditional e-commerce search systems, which overly rely on exact matching and consequently suffer from insufficient recall of substitute, complementary, and thematically related items, thereby hindering user discovery and commercial conversion. To overcome this, the authors propose an intent-conditioned recall expansion mechanism comprising a two-stage hybrid architecture: first, a closed-source large language model (LLM) enhances discoverability for head queries; then, a small language model (SLM), fine-tuned via LoRA and trained through teacher–student distillation, generalizes this capability to long-tail queries. This approach maintains high relevance while increasing the coverage of discoverable queries from 60% to 80% and reducing inference costs to approximately 30% of the teacher model’s, significantly boosting exposure for long-tail and emerging products and demonstrating strong effectiveness and scalability in real-world deployment.

0 citationsRead paper

A Cascaded Generative Approach for e-Commerce Recommendations

May 11, 2026

This work addresses the limitations of traditional e-commerce recommendation systems, which suffer from fragmented components that hinder page-level personalization, semantic coherence, and flexible adaptation to dynamic operational objectives. To overcome these challenges, the authors propose a cascaded generative product display framework that decomposes homepage construction into two stages: layout-slot theme generation and constrained keyword generation. The approach integrates teacher-student distillation for computational efficiency and incorporates conventional ranking models to preserve the benefits of hybrid architectures. This design enables end-to-end dynamic content generation coupled with AI-driven quality filtering, achieving a balance between personalization, semantic consistency, and deployment safety. Online experiments demonstrate a 2.7% increase in add-to-cart actions per page view compared to strong baselines, with the fine-tuned model attaining performance comparable to that of closed-source large language models.

0 citationsRead paper

Subimage Overlap Prediction: Task-Aligned Self-Supervised Pretraining For Semantic Segmentation In Remote Sensing Imagery

Jan 05, 2026arXiv.org

This work addresses the inefficiency of existing self-supervised pre-training methods for remote sensing image semantic segmentation under data-scarce conditions, where reliance on large volumes of unlabeled data limits transferability. To overcome this limitation, the authors propose a novel self-supervised pre-training task—sub-image overlap position prediction—that aligns closely with the downstream segmentation objective by learning semantic representations through predicting the relative positions of cropped sub-images within their original context. This approach substantially reduces the required amount of pre-training data while enhancing segmentation performance when labeled samples are scarce. Experimental results demonstrate that the proposed method consistently achieves comparable or superior mean Intersection over Union (mIoU) across multiple remote sensing benchmarks and mainstream network architectures, using less data and converging faster than conventional approaches.

0 citationsRead paper

Efficient Federated Learning with Heterogeneous Data and Adaptive Dropout

Jul 14, 2025

To address two key challenges in federated learning—degraded model accuracy due to non-IID data distributions and slow convergence caused by resource-constrained edge devices—this paper proposes FedDHAD, a novel framework integrating edge-aware design. Methodologically, it introduces (i) a Dynamic Heterogeneous Aggregation mechanism (FedDH), which adaptively assigns client aggregation weights based on local data non-IIDness; and (ii) a neuron-level Adaptive Dropout mechanism (FedAD), enabling lightweight, on-the-fly pruning of redundant neurons during training to accelerate convergence and mitigate overfitting. Crucially, both components are designed to operate under strict edge constraints without incurring additional communication overhead. Extensive experiments on standard non-IID benchmarks demonstrate that FedDHAD achieves up to 6.7% higher accuracy, 2.02× faster training, and 15.0% lower computational cost compared to state-of-the-art methods.

0 citationsRead paper

Recent publications

Latest Papers

Improving Item Discoverability in e-Commerce Search via Related Intent Generation

Jul 29, 2026

This work addresses the limitations of traditional e-commerce search systems, which overly rely on exact matching and consequently suffer from insufficient recall of substitute, complementary, and thematically related items, thereby hindering user discovery and commercial conversion. To overcome this, the authors propose an intent-conditioned recall expansion mechanism comprising a two-stage hybrid architecture: first, a closed-source large language model (LLM) enhances discoverability for head queries; then, a small language model (SLM), fine-tuned via LoRA and trained through teacher–student distillation, generalizes this capability to long-tail queries. This approach maintains high relevance while increasing the coverage of discoverable queries from 60% to 80% and reducing inference costs to approximately 30% of the teacher model’s, significantly boosting exposure for long-tail and emerging products and demonstrating strong effectiveness and scalability in real-world deployment.

0 citationsRead paper

A Cascaded Generative Approach for e-Commerce Recommendations

May 11, 2026

This work addresses the limitations of traditional e-commerce recommendation systems, which suffer from fragmented components that hinder page-level personalization, semantic coherence, and flexible adaptation to dynamic operational objectives. To overcome these challenges, the authors propose a cascaded generative product display framework that decomposes homepage construction into two stages: layout-slot theme generation and constrained keyword generation. The approach integrates teacher-student distillation for computational efficiency and incorporates conventional ranking models to preserve the benefits of hybrid architectures. This design enables end-to-end dynamic content generation coupled with AI-driven quality filtering, achieving a balance between personalization, semantic consistency, and deployment safety. Online experiments demonstrate a 2.7% increase in add-to-cart actions per page view compared to strong baselines, with the fine-tuned model attaining performance comparable to that of closed-source large language models.

0 citationsRead paper

Subimage Overlap Prediction: Task-Aligned Self-Supervised Pretraining For Semantic Segmentation In Remote Sensing Imagery

Jan 05, 2026arXiv.org

This work addresses the inefficiency of existing self-supervised pre-training methods for remote sensing image semantic segmentation under data-scarce conditions, where reliance on large volumes of unlabeled data limits transferability. To overcome this limitation, the authors propose a novel self-supervised pre-training task—sub-image overlap position prediction—that aligns closely with the downstream segmentation objective by learning semantic representations through predicting the relative positions of cropped sub-images within their original context. This approach substantially reduces the required amount of pre-training data while enhancing segmentation performance when labeled samples are scarce. Experimental results demonstrate that the proposed method consistently achieves comparable or superior mean Intersection over Union (mIoU) across multiple remote sensing benchmarks and mainstream network architectures, using less data and converging faster than conventional approaches.

0 citationsRead paper

Efficient Federated Learning with Heterogeneous Data and Adaptive Dropout

Jul 14, 2025

To address two key challenges in federated learning—degraded model accuracy due to non-IID data distributions and slow convergence caused by resource-constrained edge devices—this paper proposes FedDHAD, a novel framework integrating edge-aware design. Methodologically, it introduces (i) a Dynamic Heterogeneous Aggregation mechanism (FedDH), which adaptively assigns client aggregation weights based on local data non-IIDness; and (ii) a neuron-level Adaptive Dropout mechanism (FedAD), enabling lightweight, on-the-fly pruning of redundant neurons during training to accelerate convergence and mitigate overfitting. Crucially, both components are designed to operate under strict edge constraints without incurring additional communication overhead. Extensive experiments on standard non-IID benchmarks demonstrate that FedDHAD achieves up to 6.7% higher accuracy, 2.02× faster training, and 15.0% lower computational cost compared to state-of-the-art methods.

0 citationsRead paper