Institution profile

Ton Duc Thang University

Academic institutionasia · vn
Official website
Research library14linked papers
Opportunities0open roles
Selected work

Representative Papers

Combining Social Relations and Interaction Data in Recommender System With Graph Convolution Collaborative Filtering

Jun 03, 2025IEEE Access

To address the challenges of data sparsity, noise interference, and ineffective fusion of social influence with collaborative signals in social recommendation, this paper proposes a Robust Graph Convolutional Collaborative Filtering framework (R-GCCF). R-GCCF jointly models the user-item interaction graph and the social relation graph. It introduces, for the first time, an adaptive input denoising mechanism to suppress noise arising from sparse interactions, and enables dynamic, weighted integration of social influence and collaborative similarity within a unified GCN architecture. Additionally, it incorporates social regularization and confidence-weighted interaction modeling. Extensive experiments on multiple public benchmarks demonstrate that R-GCCF consistently outperforms state-of-the-art baselines—including NGCF, LightGCN, and SocialLGN—achieving absolute improvements of 12.7% in Recall@20 and 9.3% in NDCG@20, thereby validating its effectiveness and robustness.

2 citationsRead paper

Improvement Graph Convolution Collaborative Filtering with Weighted Addition Input

Mar 27, 2025Asian Conference on Intelligent Information and Database Systems

To address the problem that conventional graph neural network–based recommender systems neglect interaction weights between users and items—leading to insufficiently discriminative representations—this paper proposes a learnable weighted input mechanism. Specifically, it constructs an auxiliary weight graph to explicitly model heterogeneous influences of different users on the same item. This approach achieves, for the first time in GCN-based collaborative filtering, end-to-end learnable weighting of the adjacency matrix without relying on explicit social relations; instead, it infers implicit user associations solely from co-occurrence patterns in user-item interactions. Integrated into LightGCN and NGCF architectures, the proposed method consistently outperforms state-of-the-art baselines—including GCMC, NGCF, and LightGCN—across multiple public benchmarks, achieving up to a 12.6% improvement in Recall@20. The implementation is publicly available.

2 citationsRead paper

CombiGCN: An Effective GCN Model for Recommender System

Mar 27, 2025International Conference on Computational Social Networks

To address the limited graph signal modeling capability in GNN-based collaborative filtering, this paper proposes a dual-graph co-propagation framework. Item embeddings propagate solely on the user-item bipartite graph, while user embeddings propagate in parallel on both the interaction graph and an explicitly constructed weighted user–user similarity graph. Crucially, user similarity is modeled as a learnable weighted graph, and a Light Graph Convolution operator coupled with a weighted embedding fusion mechanism is designed to decouple user and item propagation paths, thereby enhancing collaborative signal representation. Experiments on three public benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance, improving Recall@20 by up to 12.6% over existing approaches—validating the effectiveness of dual-graph co-propagation in boosting recommendation accuracy.

1 citationsRead paper
Recent publications

Latest Papers

A Revisit to Point Estimation Through the Empirical Bayes Method: The Case of Binomial Distribution with Beta Prior and Extension to Poisson Distribution

Aug 13, 2026

This study reevaluates the efficacy of empirical Bayes methods for parameter estimation in binomial (with Beta priors) and Poisson models. Through theoretical analysis and extensive numerical experiments, it specifically investigates Type-II maximum likelihood (ML-II) under general two-parameter Beta priors and extends the examination to the Gamma–Poisson setting. The findings reveal that the ML-II procedure fails under a general two-parameter Beta prior; even when restricted to a symmetric one-parameter Beta prior, the resulting estimator does not substantially outperform the maximum likelihood estimator under quadratic loss. These results challenge the commonly presumed superiority of empirical Bayes approaches in point estimation and provide a critical counterexample grounded in rigorous empirical evidence.

0 citationsRead paper

Robust Transformer-Based One-Step Stock Index Forecasting via Shifted Data Augmentation

Jun 14, 2026

This study addresses the limitations of standard Transformers in financial time series forecasting, where performance is hindered by noise interference, short-term memory dynamics, and distribution shifts. To overcome these challenges, the authors propose an enhanced Transformer architecture that integrates cosine annealing with warm restarts for learning rate scheduling and introduces a novel Shifted Data Augmentation (SDA) technique. SDA substantially reduces prediction errors and training instability, demonstrating superior robustness compared to merely increasing model complexity. Empirical evaluations on the VN30 and S&P 500 datasets show that the proposed method achieves state-of-the-art performance in terms of prediction accuracy, training stability, and robustness to hyperparameter variations.

0 citationsRead paper

C2L-Net: A Data-Driven Model for State-of-Charge Estimation of Lithium-Ion Batteries During Discharge

May 08, 2026

This work addresses the high computational cost and initial padding bias inherent in existing data-driven state-of-charge (SOC) estimation methods, which typically rely on long historical sequences. To overcome these limitations, the authors propose C2L-Net, a novel framework that decouples contextual modeling from recent observations. By leveraging an ultra-short window of merely 20 seconds, C2L-Net efficiently encodes historical context through chunked feature extraction, Theta Attention Pooling, Fourier seasonal bases, and a causal cosine attention mechanism, while a recursive filtering-style decoder fuses the latest measurements. Evaluated on multiple public lithium-ion battery datasets, the method achieves state-of-the-art or competitive accuracy, reduces model parameters substantially, accelerates inference speed by up to 60×, and demonstrates strong robustness under unseen driving conditions.

0 citationsRead paper