Hai Bang Truong
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Hai Bang Truong has a total of 31 co-authors. Here are the top co-authors by number of shared publications:
Van Du Nguyen
Faculty of Information Technology, Nong Lam University, Ho Chi Minh City, Vietnam
3 shared publications- Adrianna Kozierkiewicz
Wroclaw University of Science and Technology
2 shared publications - Bogumiła Hnatkowska
Wroclaw University of Science and Technology
2 shared publications - Marcin Pietranik
Wroclaw University of Science and Technology
2 shared publications - Trung Van Nguyen
Faculty of Information Technology, Nguyen Tat Thanh University, 300A Nguyen Tat Thanh Street, Xom Chieu Ward, Ho Chi Minh City 70000, Vietnam
2 shared publications Van Cuong Tran
Faculty of Engineering and Information Technology, Quang Binh University, Dong Hoi, Vietnam
2 shared publications
Latest Publications
- Open Access
- 20 October 2026
- Published by Elsevier in SSRN Electronic Journal
When does fuzzy control make swarm search adaptive? A conditional analysis and controlled study of fuzzy-controlled PSO and GWO for text feature selection
Fuzzy controllers are often added to swarm optimizers to adapt their parameters, and swarm wrappers are used to select features for fake news detection; we audit both. For a Mamdani controller of binary particle swarm optimization (AF-PSO), normalising swarm diversity by the diagonal of the search box keeps this input near a constant, and traces show the improvement input also stays low; normalising both inputs by the run history makes the controller adapt. On continuous benchmarks the adaptive controller beat scheduled PSO on four of six functions, but a linear, non-fuzzy mapping of the same inputs did as well or better, and the advantage depended on swarm size and budget. On text, adaptation did not yield better feature subsets with 1,000 or 5,000 evaluations. For a fuzzy weighting of the leaders of grey wolf optimization (FW-FGWO), min–max normalisation fixes two of the three weights, and constant weights matched to its average recover most of its gain over GWO. Over ten paired partitions of de-duplicated LIAR and ISOT we found no statistically significant improvement of any wrapper over the full feature set or a χ² filter, and on the official LIAR split 39 of 40 wrapper runs scored below the full feature set; on LIAR the wrappers matched random subsets of the same size, and gains of one macro-F1 point or more were rejected. These findings hold for TF-IDF features, linear classifiers, budgets up to 5,000 evaluations and two English corpora; the theoretical results are conditional bounds, not convergence proofs.
- 20 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessWhen does fuzzy control make swarm search adaptive? A conditional analysis and controlled study of fuzzy-controlled PSO and GWO for text feature selection
Fuzzy controllers are often added to swarm optimizers to adapt their parameters, and swarm wrappers are used to select features for fake news detection; we audit both. For a Mamdani controller of binary particle swarm optimization (AF-PSO), normalising swarm diversity by the diagonal of the search box keeps this input near a constant, and traces show the improvement input also stays low; normalising both inputs by the run history makes the controller adapt. On continuous benchmarks the adaptive controller beat scheduled PSO on four of six functions, but a linear, non-fuzzy mapping of the same inputs did as well or better, and the advantage depended on swarm size and budget. On text, adaptation did not yield better feature subsets with 1,000 or 5,000 evaluations. For a fuzzy weighting of the leaders of grey wolf optimization (FW-FGWO), min–max normalisation fixes two of the three weights, and constant weights matched to its average recover most of its gain over GWO. Over ten paired partitions of de-duplicated LIAR and ISOT we found no statistically significant improvement of any wrapper over the full feature set or a χ² filter, and on the official LIAR split 39 of 40 wrapper runs scored below the full feature set; on LIAR the wrappers matched random subsets of the same size, and gains of one macro-F1 point or more were rejected. These findings hold for TF-IDF features, linear classifiers, budgets up to 5,000 evaluations and two English corpora; the theoretical results are conditional bounds, not convergence proofs.
- Open Access
- 27 August 2026
- Published by Elsevier in SSRN Electronic Journal
FKG-PPO: A Layered Fuzzy Expert System for Multi-Objective Personalized Learning Path Planning via Pareto-Optimal Planning and Self-Adaptive Curriculum Learning
Background. Personalized learning-path recommendation is a central challenge for adaptive online tutoring, requiring simultaneous modelling of latent learner knowledge states, prerequisite constraints among knowledge components (KCs), and the pedagogical rationale that justifies each recommended step. Existing knowledge-tracing (KT) methods and knowledge-graph (KG) recommenders produce competitive predictive accuracy but remain opaque, treat mastery as a binary or point-valued signal, and rarely emit natural-language explanations that stakeholders can audit.Objective. I introduce Fuzzy GraphRAG, a graph-retrieval-augmented generation framework that unifies (i) a fuzzy knowledge graph whose nodes carry fuzzy mastery memberships and whose edges carry fuzzy prerequisite strengths, (ii) a fuzzy Leiden community-detection stage yielding hierarchical concept clusters robust to noisy prerequisite links, (iii) a dual-level fuzzy retrieval mechanism composing local-neighbourhood and community-summary evidence, and (iv) an LLM-augmented generation stage producing both the next KC to practise and a grounded, learner-facing rationale.Methods. The framework is instantiated on the four public ASSISTments releases (2009, 2012, 2015, 2017), with a cross-domain protocol prepared for an additional non-ASSISTments corpus. Fuzzy learner-state updates combine an attention-based encoder with Bayesian-Knowledge-Tracing-style transitions and are calibrated against observed correctness. All hyperparameters are given a theoretical grounding and an automatic-calibration procedure. Next-response prediction is measured with AUC, ACC and RMSE; path quality with Prerequisite Satisfaction Rate (PSR) and Path Coherence (PC); and explanation quality with automated (BLEU-4, ROUGE-L, BERTScore) and expert-rated (relevance, faithfulness, actionability) metrics, together with a computation-versus-benefit analysis.Results. Fuzzy GraphRAG matches or exceeds nine strong baselines — BKT, PFA, DKT, DKVMN, SAKT, AKT, GKT, LPKT, SAINT — and a plain GraphRAG variant, achieving an average AUC of 0.836 on ASSISTments 2009 and 0.821 on ASSISTments 2017, together with 88.4% prerequisite satisfaction and expert-rated explanation faithfulness of 4.31/5. Ablations show that the largest, and complementary, gains are in path quality, explanation faithfulness and cold-start robustness rather than in raw AUC alone.Conclusion. By explicitly encoding epistemic and aleatoric uncertainty as fuzzy memberships and grounding LLM generation on retrieved subgraphs, Fuzzy GraphRAG delivers accurate, prerequisite-consistent and human-auditable learning paths — a deployable route to trustworthy adaptive tutoring.
- 27 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessFKG-PPO: A Layered Fuzzy Expert System for Multi-Objective Personalized Learning Path Planning via Pareto-Optimal Planning and Self-Adaptive Curriculum Learning
Background. Personalized learning-path recommendation is a central challenge for adaptive online tutoring, requiring simultaneous modelling of latent learner knowledge states, prerequisite constraints among knowledge components (KCs), and the pedagogical rationale that justifies each recommended step. Existing knowledge-tracing (KT) methods and knowledge-graph (KG) recommenders produce competitive predictive accuracy but remain opaque, treat mastery as a binary or point-valued signal, and rarely emit natural-language explanations that stakeholders can audit.Objective. I introduce Fuzzy GraphRAG, a graph-retrieval-augmented generation framework that unifies (i) a fuzzy knowledge graph whose nodes carry fuzzy mastery memberships and whose edges carry fuzzy prerequisite strengths, (ii) a fuzzy Leiden community-detection stage yielding hierarchical concept clusters robust to noisy prerequisite links, (iii) a dual-level fuzzy retrieval mechanism composing local-neighbourhood and community-summary evidence, and (iv) an LLM-augmented generation stage producing both the next KC to practise and a grounded, learner-facing rationale.Methods. The framework is instantiated on the four public ASSISTments releases (2009, 2012, 2015, 2017), with a cross-domain protocol prepared for an additional non-ASSISTments corpus. Fuzzy learner-state updates combine an attention-based encoder with Bayesian-Knowledge-Tracing-style transitions and are calibrated against observed correctness. All hyperparameters are given a theoretical grounding and an automatic-calibration procedure. Next-response prediction is measured with AUC, ACC and RMSE; path quality with Prerequisite Satisfaction Rate (PSR) and Path Coherence (PC); and explanation quality with automated (BLEU-4, ROUGE-L, BERTScore) and expert-rated (relevance, faithfulness, actionability) metrics, together with a computation-versus-benefit analysis.Results. Fuzzy GraphRAG matches or exceeds nine strong baselines — BKT, PFA, DKT, DKVMN, SAKT, AKT, GKT, LPKT, SAINT — and a plain GraphRAG variant, achieving an average AUC of 0.836 on ASSISTments 2009 and 0.821 on ASSISTments 2017, together with 88.4% prerequisite satisfaction and expert-rated explanation faithfulness of 4.31/5. Ablations show that the largest, and complementary, gains are in path quality, explanation faithfulness and cold-start robustness rather than in raw AUC alone.Conclusion. By explicitly encoding epistemic and aleatoric uncertainty as fuzzy memberships and grounding LLM generation on retrieved subgraphs, Fuzzy GraphRAG delivers accurate, prerequisite-consistent and human-auditable learning paths — a deployable route to trustworthy adaptive tutoring.
- Open Access
- 18 August 2026
- Published by Springer Nature in Research Square
A Transparent Fuzzy-Inference Layer for Explainable Knowledge Tracing: Why Prerequisite Graphs Must Be Expert-Supplied, and Which T-Norm to Use
Deep knowledge tracing (KT) predicts student difficulty accurately but is opaque about the prerequisite concepts a teacher must address. A popular neuro-symbolic remedy attaches a fuzzy prerequisite graph to a KT backbone, assuming the graph can be discovered from interaction logs. We test that assumption and report two principal results. First, a systematic negative result: on the ASSISTments 2009 benchmark, three log-only induction families — transfer entropy, confound-controlled mastery-lift, and semi-supervised expert-seeded classification — all fail to recover expert prerequisites at a usable level; the confound-controlled estimator ranks the 16 expert edges at a median of 3,622 out of 7,144 candidate pairs (chance ≈ 3,572). We trace this failure to an ability/difficulty confound that renders almost all skill pairs positively associated, and we show that the confound replicates on two further datasets (ASSISTments 2012 and Junyi, 25.9 million interactions). This finding concerns log-only induction and is orthogonal to content-based (NLP/LLM) methods. Second, we introduce a reusable, rater-free localisation-audit framework with model-defined ground truth: it measures whether a trained model encodes each expert prerequisite and whether an explanation localises it. The audit shows that the model encodes most prerequisites (12 of 16 on DKT, 11 of 16 on SAKT), whereas post-hoc attribution localises them unreliably (top-3 hit-rate 66% for Integrated Gradients and 61% for occlusion-SHAP on DKT), with occlusion-SHAP collapsing to 27% on a self-attentive model — a graph-grounded fuzzy layer localises reliably regardless of architecture. Our two central contributions are therefore the log-signal limit and the reusable, architecture-independent localisation-audit framework; we deliberately keep the fuzzy exploitation layer (D-FIS) fixed and parameter-free so that it is transparent and auditable by construction, and we compare four fuzzy conjunction operators (product, minimum, Łukasiewicz, drastic), finding that the product t-norm aligns best with the model's true prerequisite effect (positive in 97.6% of bootstrap resamples; permutation p = 0.026). All code and anonymised data are released.
- 18 August 2026
- Published by Springer Nature in Research Square
Open AccessA Transparent Fuzzy-Inference Layer for Explainable Knowledge Tracing: Why Prerequisite Graphs Must Be Expert-Supplied, and Which T-Norm to Use
Deep knowledge tracing (KT) predicts student difficulty accurately but is opaque about the prerequisite concepts a teacher must address. A popular neuro-symbolic remedy attaches a fuzzy prerequisite graph to a KT backbone, assuming the graph can be discovered from interaction logs. We test that assumption and report two principal results. First, a systematic negative result: on the ASSISTments 2009 benchmark, three log-only induction families — transfer entropy, confound-controlled mastery-lift, and semi-supervised expert-seeded classification — all fail to recover expert prerequisites at a usable level; the confound-controlled estimator ranks the 16 expert edges at a median of 3,622 out of 7,144 candidate pairs (chance ≈ 3,572). We trace this failure to an ability/difficulty confound that renders almost all skill pairs positively associated, and we show that the confound replicates on two further datasets (ASSISTments 2012 and Junyi, 25.9 million interactions). This finding concerns log-only induction and is orthogonal to content-based (NLP/LLM) methods. Second, we introduce a reusable, rater-free localisation-audit framework with model-defined ground truth: it measures whether a trained model encodes each expert prerequisite and whether an explanation localises it. The audit shows that the model encodes most prerequisites (12 of 16 on DKT, 11 of 16 on SAKT), whereas post-hoc attribution localises them unreliably (top-3 hit-rate 66% for Integrated Gradients and 61% for occlusion-SHAP on DKT), with occlusion-SHAP collapsing to 27% on a self-attentive model — a graph-grounded fuzzy layer localises reliably regardless of architecture. Our two central contributions are therefore the log-signal limit and the reusable, architecture-independent localisation-audit framework; we deliberately keep the fuzzy exploitation layer (D-FIS) fixed and parameter-free so that it is transparent and auditable by construction, and we compare four fuzzy conjunction operators (product, minimum, Łukasiewicz, drastic), finding that the product t-norm aligns best with the model's true prerequisite effect (positive in 97.6% of bootstrap resamples; permutation p = 0.026). All code and anonymised data are released.
- Open Access
- 12 August 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (16) , 8023
A Collective Intelligence Framework for Fake News Detection on Social Networks
The rapid diffusion of user-generated content on social networks has amplified the reach of fake news, creating an urgent need for detection methods that combine scalability with epistemic robustness. This paper proposes a Collective Intelligence (CI) framework for fake news detection that formalizes the crowd assessing a news item as an intelligent collective characterized by diversity, independence, decentralization, and aggregation. A directed weighted graph is used to represent the collective, where vertices denote users, edge weights encode reputation-derived influence, and each user contributes a veracity judgment together with a set of stance-bearing content and context features. We instantiate collective independence through a reputation-based influence measure adapted from prior work and integrate the resulting independence-aware weights into a two-stage aggregation pipeline: (i) a supervised classifier that produces machine-generated veracity scores from news content, and (ii) a consensus operator that fuses machine scores with independence-weighted crowd signals. The framework is evaluated on the GossipCop split of the FakeNewsNet corpus, treating the tweet propagation graph associated with each news item as the collective. Experimental results show that the CI-based model outperforms content-only and unweighted crowd baselines in accuracy, precision, recall, and F1-score, and that independence-aware aggregation contributes the largest incremental gain among the four CI principles. These findings support the view that treating social-media crowds as structured collectives, rather than as bags of independent votes, yields measurable robustness against coordinated misinformation.
- 12 August 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (16) , 8023
Open AccessA Collective Intelligence Framework for Fake News Detection on Social Networks
The rapid diffusion of user-generated content on social networks has amplified the reach of fake news, creating an urgent need for detection methods that combine scalability with epistemic robustness. This paper proposes a Collective Intelligence (CI) framework for fake news detection that formalizes the crowd assessing a news item as an intelligent collective characterized by diversity, independence, decentralization, and aggregation. A directed weighted graph is used to represent the collective, where vertices denote users, edge weights encode reputation-derived influence, and each user contributes a veracity judgment together with a set of stance-bearing content and context features. We instantiate collective independence through a reputation-based influence measure adapted from prior work and integrate the resulting independence-aware weights into a two-stage aggregation pipeline: (i) a supervised classifier that produces machine-generated veracity scores from news content, and (ii) a consensus operator that fuses machine scores with independence-weighted crowd signals. The framework is evaluated on the GossipCop split of the FakeNewsNet corpus, treating the tweet propagation graph associated with each news item as the collective. Experimental results show that the CI-based model outperforms content-only and unweighted crowd baselines in accuracy, precision, recall, and F1-score, and that independence-aware aggregation contributes the largest incremental gain among the four CI principles. These findings support the view that treating social-media crowds as structured collectives, rather than as bags of independent votes, yields measurable robustness against coordinated misinformation.
- 3 June 2026
- Published by Springer Nature
- p. 109-119
Detecting Fake News
- 24 July 2024
- Published by Elsevier in Food Chemistry
- Vol. 460 (Pt 1) , 140479
Toxic metals in rice among Asian countries: A review of occurrence and potential human health risks
- 24 July 2024
- Published by Elsevier in Food Chemistry
- Vol. 460 (Pt 1) , 140479
Toxic metals in rice among Asian countries: A review of occurrence and potential human health risks
- Open Access
- 22 March 2024
- Published by World Scientific Pub Co Pte Ltd in Vietnam Journal of Computer Science
- Vol. 12 (01) , 1-19
Community Detection Methods Based on Exploiting Attributes and Interactions on Social Networks: A Survey and Future Directions
In recent years, community detection in social network analysis has been a hot research topic that has attracted much more attention in the scientific community and a large number of articles have already been published in the literature. Community detection problems focus on determining the subgraphs of the network that have dense edges and whose nodes are expected to have similar features. Community detection has significant applications in different domains of social network analysis such as sociology, business, criminal detection and recommendation systems. The computational complexity is hampered by the huge size and the dynamic nature of social networks. This paper introduces a comprehensive survey of community detection methods in social networks based on mining attributes and interactions of users. We divide the community detection approaches into three basic categories, namely (1) node content-based approaches, (2) edge content-based approaches, and (3) node- and edge content-based approaches. We also discuss in detail the main idea of each method that is suitable for social network data. Furthermore, to promote the study of community detection, we introduce the benchmark datasets from available sources. Last, we present further research directions as well as some open challenges for the community detection problem in social networks.
- 22 March 2024
- Published by World Scientific Pub Co Pte Ltd in Vietnam Journal of Computer Science
- Vol. 12 (01) , 1-19
Open AccessCommunity Detection Methods Based on Exploiting Attributes and Interactions on Social Networks: A Survey and Future Directions
In recent years, community detection in social network analysis has been a hot research topic that has attracted much more attention in the scientific community and a large number of articles have already been published in the literature. Community detection problems focus on determining the subgraphs of the network that have dense edges and whose nodes are expected to have similar features. Community detection has significant applications in different domains of social network analysis such as sociology, business, criminal detection and recommendation systems. The computational complexity is hampered by the huge size and the dynamic nature of social networks. This paper introduces a comprehensive survey of community detection methods in social networks based on mining attributes and interactions of users. We divide the community detection approaches into three basic categories, namely (1) node content-based approaches, (2) edge content-based approaches, and (3) node- and edge content-based approaches. We also discuss in detail the main idea of each method that is suitable for social network data. Furthermore, to promote the study of community detection, we introduce the benchmark datasets from available sources. Last, we present further research directions as well as some open challenges for the community detection problem in social networks.
- Open Access
- 13 October 2023
- Published by Springer Nature in Research Square
Vietnam Spam Review Detection using Graph Reinforcement Learning
With the proliferation of social media platforms, especially commercial platforms today, and the ease of access to internet shopping, it is difficult to determine which products are good for clients based on reviews from users. Numerous successful innovations in the field of spam product review detection have been made. This study focuses on the creation of an efficient method for Vietnam spam product review detection using cutting-edge language models in addition to Vietnam-related work. For feature extraction, we utilize PhoBERT, a cutting-edge pretrained model for Vietnamese language processing. The downstream task model will be followed by the Graph Reinforcement Learning architecture, which has had many successes recently. Our proposed GRL model tries to capture text reviews into graph representations and utilize passing mechanisms for leveraging the product description as a global message passing to the review. By combining with the RL framework, we can construct the modulator, which controls the amount of information passing from the product description and gives out the best outcomes in the long term and outperformance compared to traditional methods.
- 13 October 2023
- Published by Springer Nature in Research Square
Open AccessVietnam Spam Review Detection using Graph Reinforcement Learning
With the proliferation of social media platforms, especially commercial platforms today, and the ease of access to internet shopping, it is difficult to determine which products are good for clients based on reviews from users. Numerous successful innovations in the field of spam product review detection have been made. This study focuses on the creation of an efficient method for Vietnam spam product review detection using cutting-edge language models in addition to Vietnam-related work. For feature extraction, we utilize PhoBERT, a cutting-edge pretrained model for Vietnamese language processing. The downstream task model will be followed by the Graph Reinforcement Learning architecture, which has had many successes recently. Our proposed GRL model tries to capture text reviews into graph representations and utilize passing mechanisms for leveraging the product description as a global message passing to the review. By combining with the RL framework, we can construct the modulator, which controls the amount of information passing from the product description and gives out the best outcomes in the long term and outperformance compared to traditional methods.
- 13 September 2023
- Published by Springer Nature
- p. 3-15
Hybrid Genetic Algorithms to Determine 2-Optimality Consensus for a Collective of Ordered Partitions
- 13 September 2023
- Published by Springer Nature
- p. 3-15
Hybrid Genetic Algorithms to Determine 2-Optimality Consensus for a Collective of Ordered Partitions