CTU Journal of Innovation and Sustainable Development

SignKG: Spatially aware urban business knowledge graph construction from Vietnamese street signboards

Nguyen Hoang Pham — 2026-11-09

Street-level signboards contain valuable urban business information, yet conventional OCR-based systems typically return isolated text strings without identifying their semantic roles or relationships. This limitation is particularly challenging for Vietnamese signboards because of decorative fonts, complex diacritics, irregular layouts, and ambiguous text content. This paper proposes SignKG, a spatially aware framework that transforms Vietnamese street-signboard images into structured urban business knowledge graphs. SignKG integrates visual perception, spatially aware entity extraction, constrained relation inference, and graph optimisation. The framework jointly exploits lexical semantics, textbox positions, neighbourhood context, business rules, and symbolic compatibility constraints to identify business entities and construct semantically consistent relations. Experiments on a real-world Vietnamese signboard dataset show that SignKG achieves entity-, relation-, and graph-level F₁ scores of 0.910, 0.874, and 0.851, respectively. It also obtains a valid-edge ratio of 0.967 and a store-connectivity ratio of 0.951, outperforming rule-based, text-only, layout-aware, and LLM-based baselines. Ablation results confirm that spatial information is critical for entity disambiguation, while symbolic constraints improve graph validity and structural consistency.

Patent network analysis of core technology groups and convergence hubs in generative-AI-enabled digital transformation

Chitaranjan L. Singh — 2026-11-09

Generative AI (GenAI) is redefining the technological architecture of enterprise digital transformation, yet its domain structure and convergence topology remain uncharacterized at the patent level. This study applies a three-stage patent analytics pipeline- Cooperative Patent Classification domain mapping, keyword-rule topic assignment across 15 GenAI application archetypes, and weighted co-occurrence network analysis- to 9,692 Lens.org patent records from ten jurisdictions, of which 7,623 published from 2018 onward form the analytical corpus. Filings grew at a compound annual rate of 42% over 2018–2025. Domain mapping identifies Computing & Information Systems (69.5%), AI & Machine Learning (36.5%) and Business & Commerce (25.2%) as a three-layer infrastructure, model and application stack. Large language model development and retrieval-augmented generation together account for 64.5% of filings, while agentic AI and edge inference clusters remain under 1% each. The co-occurrence network (26 nodes, 151 edges, density 0.465) identifies Computing & Information Systems as the dominant convergence hub (composite score 0.759) and Healthcare Informatics as a structural broker (betweenness 0.138) disproportionate to its filing volume. Cross-domain patenting peaked at 2.34 domains per patent in 2021 and moderated to 1.65 by 2025. The findings extend general-purpose technology convergence theory to GenAI and provide a structural map for technology managers.

Hyperspectral image super-resolution using spatial-spectral attention and spectral transformer

Thanh-Van Nguyen — 2026-11-09

Hyperspectral image super-resolution (HSI-SR) aims to reconstruct high-resolution hyperspectral images (HR-HSI) from low-resolution observations (LR-HSI). Existing convolutional neural network (CNN)-based methods excel at extracting local spatial features but struggle to capture long-range spectral dependencies, while full spatial Transformer architectures incur high computational cost. We propose ESSA-SSAM-SpecTrans, a hybrid architecture that extends the Efficient Spectral-Spatial Attention (ESSA) baseline with two novel modules: (1) a Separated Spatial-Spectral Attention Module (SSAM) that independently models spatial and spectral features in 2D image space via dual-branch attention, and (2) a Spectral Transformer that applies multi-head self-attention along the spectral dimension, treating each spectral band as a token with complexity O(C²·HW) instead of O((HW)²·C). Experiments on two benchmark datasets with a scale factor of ×4 demonstrate the effectiveness of the proposed approach. On the CAVE dataset, the model achieves a PSNR of 37.78 dB, an SSIM of 0.960, and an ERGAS of 3.18, the best results among the compared methods on these three metrics. On the Pavia Centre dataset, it achieves the lowest SAM (4.80°) and ERGAS (5.51), indicating strong spectral fidelity on real-world remote-sensing data.

Benchmarking encoder-decoder architectures and input resolutions for rice lodging semantic segmentation using UAV RGB imagery

Loan Ngoc Thi Chau — 2026-11-09

Rice lodging reduces yield and hinders mechanised harvesting; consequently, its timely delineation is important for field monitoring and damage assessment. This study benchmarks semantic segmentation models using UAV RGB imagery acquired during three field surveys in An Giang Province, Vietnam. The dataset contains 715 annotated images with four classes: Background, Lodged Rice, Ponds/Ditches, and Field Edges/Paths. Fifteen encoder–decoder configurations were first benchmarked at an input resolution of 512×512 pixels. Based on validation-set mIoU, one configuration from each decoder was then selected for further evaluation at 640 × 640 and 768 × 768 pixels: U-Net with a MiT-B1 encoder, DeepLabV3+ with an EfficientNet-B1 encoder, U-Net++ with a MobileNetV2 encoder, and SegFormer with a MiT-B1 encoder. To address class imbalance, training used a hybrid loss that combined Dice loss and weighted cross-entropy loss. At 512 × 512 pixels, SegFormer with the MiT-B1 encoder achieved the highest mIoU (66.72%) and mF1 scores (79.01%). At 640 × 640 pixels, DeepLabV3+ with the EfficientNet-B1 encoder achieved the highest mIoU (67.48%) and mF₁-score (79.60%), whereas SegFormer with the MiT-B1 encoder obtained the highest mPA (91.23%). Increasing the input resolution to 768 × 768 pixels did not yield consistent improvements. Overall, 640 × 640 pixels provided the most favourable performance in terms of mIoU and mF₁-score among the evaluated settings.

How far can sub-3B open language models go in zero-shot essay scoring on an 8 GB consumer GPU?

Nguyen Dung Son — 2026-11-09

Zero-shot essay scoring with large language models is usually demonstrated with proprietary remote models, yet the classrooms that most need it are often barred from using them. This study asks what survives when the scorer must be a sub-3B open model run locally. Four models from two families, Qwen2.5 at 0.5B, 1.5B and 3B and SmolLM2-1.7B, were evaluated on all eight Automated Student Assessment Prize prompts on one 8 GB consumer graphics processing unit, with bootstrap intervals and Holm-corrected tests. Under batch min-max aggregation, rubric-decomposed prompting improved macro quadratic weighted kappa for every model, though Qwen2.5-3B fell significantly on one prompt, and, of the two strategies, only it made all four models track the per-prompt human ceiling (Spearman ρ +0.60 to +0.85, significant for one model only). Aggregation interacted with grader calibration: one model collapsed to 0.204 under naive averaging and recovered to 0.388 under Multi-Trait Specialization's min-max normalization. Signed error fell with essay length in eleven of twelve configurations, opposite to large judge models' verbosity bias. The best configuration reached 0.388, half of the human ceiling (0.769), and stayed below a length-only baseline (0.523, itself batch min-max normalized), so sub-3B local scoring suits only formative, human-supervised feedback.

Multi-agent training for Pommerman: Curriculum learning and population-based self-play approach

Nhat Minh Huynh — 2026-11-09

Pommerman is a multi-agent environment that has received considerable attention from researchers in recent years. This environment is an ideal benchmark for multi-agent training, providing a battleground for two teams with communication capabilities among allied agents. Pommerman poses significant challenges for model-free reinforcement learning due to delayed action effects, sparse rewards, and false positives, where opponents can lose because of their own mistakes. This study introduces a system designed to train multi-agent systems to play Pommerman using a combination of curriculum learning and population-based self-play. We also tackle two challenging problems when deploying the multi-agent training system for competitive games: sparse reward and a suitable matchmaking mechanism. Specifically, we propose an adaptive annealing factor, based on agents’ performance, to dynamically adjust the dense exploration reward during training. Additionally, we implement a matchmaking mechanism utilizing the Elo rating system to pair agents effectively. Our experimental results demonstrate that our trained agent can outperform top learning agents without requiring communication among allied agents.

3D-ReSE-VN: Advanced Deep Learning for Protein Ligand Affinity Prediction in Cancer-Associated Targets

Muhammad Hussain — 2026-11-09

Accurate protein-ligand binding affinity prediction is important for structure-based anticancer drug discovery. This study presents 3D-ReSE-VN, a three-dimensional residual attention network that combines voxelized docking-derived complexes with residual learning, squeeze-and-excitation channel attention, and spatial attention. The model was evaluated on 4,620 protein-ligand complexes using random five-fold cross-validation and a predefined train/validation/test split. In cross-validation, 3D-ReSE-VN achieved an RMSE of 1.9679 ± 0.8242, an MAE of 0.6952 ± 0.1079, a Pearson correlation of 0.8169 ± 0.1095, and an R² of 0.6724 ± 0.1755. On the fixed hold-out test set of 693 complexes, the model achieved an RMSE of 1.3251, an MAE of 0.6420, a Pearson correlation of 0.9098, and an R²of 0.8271. The model showed competitive performance against the reported architectures, although performance varied across data partitions and affinity ranges. These findings support further evaluation on larger, chemically diverse, and disjoint protein-ligand datasets.

Privacy-preserving blood cancer detection via vision transformers and federated learning

Nguyen Thanh Luan Huynh — 2026-11-09

Artificial Intelligence acts as a powerful tool to analyze medical images in smart healthcare. However, selecting a deep learning architecture is difficult in distributed computing environments due to privacy issues that prohibit centralized training. In this study, five architectures were considered for the evaluation of the performance of blood cancer image classification via federated learning (FL), where a baseline Convolutional Neural Network (CNN), ResNet50, DenseNet121, Xception, and a Vision Transformer (ViT) were selected. This setting allows the exploration of each architecture's performance in terms of image classification in a distributed, data-partitioned setting. These architectures were trained using the same federated learning framework. The experimental results show that such federated configurations offer competitive results compared with traditional methods used in centralized training of neural networks. According to our analysis, the best performance in accuracy, precision, recall, and F1-score was obtained with the ViT architecture, with a score of 0.98. Xception came second by obtaining accuracy and F1-score scores of 0.89, while ResNet50, CNN, and DenseNet121 achieved lower scores in the range of 0.82 - 0.86. The use of FL in conjunction with ViT yielded the highest score of accuracy and was more robust than the other FL methods.

FLAIR-XGT: Local visual patch feature aggregation with multi-graph fusion for image-based recommendation

Huynh Quoc Bao Le — 2026-11-09

Existing multimodal recommender systems discard fine-grained spatial information before recommendation training even starts by representing each product image as a single global vector, either a pooled CNN output or a Vision Transformer’s CLS token. We propose FLAIR-XGT, an image-based recommendation framework that uses three parallel graph networks (LightGCN, VisualGCN, and TextGCN) to jointly train representative local DINOv2 ViT-B/14 patch tokens (each covering a pixel region) end-to-end under a combined BPR + InfoNCE objective. Our primary scoring component, GlobalVis, combines the local patch tokens using a projector that is optimized in tandem with the recommender. In comparison to all fixed pre-extracted global-feature baselines, this joint training of local patch aggregation with multi-graph collaborative fusion produces NDCG@20 on Amazon Fashion ( over MMGCN) and NDCG@20 on Sports & Outdoors ( over MMGCN). We also study XGT, a user-conditioned cross-attention module that focuses spatial attention on user-relevant patches. A Graph Transformer block with a CF residual connection fuses the attended output. One important empirical finding is that XGT cross-attention performs worse than GlobalVis on both datasets, which reflects a cold-start limitation: user embeddings are too sparse to form discriminative attention queries, making learned mean-pooling of local patches the stronger configuration in practice. Sensitivity analysis verifies resilience across the critical hyperparameters, and ablation studies link each performance gain to particular design decisions.

Application of YOLO models for colorectal cancer diagnosis: A case study in Vietnam

Duc Luu Ngo — 2026-11-09

Early detection and treatment of colorectal cancer (CRC) are expected to increase the survival rate of patients. Deep learning methods for automatically diagnosing diseases based on CT images are increasingly developed, and good data preparation plays an important role in affecting the effectiveness of the training model. This study will introduce how we collect and process data for the object detection model, evaluate the effectiveness of the data based on the performance evaluation of the YOLOv8m, YOLOv9c, YOLOv10s models, and evaluate the effectiveness of disease sign detection on authentic images collected at hospitals that have never been used to train the model. We collected public data on the TCIA website (1558 slices showing CRC lesions, 121 patients) and actual data collected at Can Tho City Oncology Hospital, Vietnam (19093 JPEG CT images of 31 patients). As a result, the training and external evaluation models all achieved high results (YOLOv8m: The F1 score, recall, and mAP were all stable at 95% or higher; YOLOv9c: Precision = 1, F1 score = 97. 4%, recall = 95%, mAP = 94.9%; YOLOv10s: F1 score = 97. 4%, mAP = 94. 4%, Recall = 95%). When evaluated on an external dataset collected from Can Tho Oncology Hospital, the results showed that the trained models could detect signs of disease in 10/10 patients with a confidence level of 50% or higher.

GFlowRule: Detailed-balance GFlowNets for diverse symbolic rule selection in neuro-symbolic image classification

Truong Thanh Ma — 2026-11-09

We propose a neuro-symbolic image-classification framework that combines CNN representations, Random-Forest-derived symbolic rules, and a Detailed-Balance GFlowNet for diverse rule-set selection. The GFlowNet optimizes a multi-component reward based on macro accuracy, correct and incorrect coverage, and inter-rule conflict, while the selected rules regularize CNN fine-tuning through either a fixed rule set or stochastic marginalization over ???? = 32 sampled rule sets. On a 12 class Vietnamese cultural-heritage dataset with 10,674 images, four backbones, and five random seeds, GFlowNet consistently obtains the highest matched-budget rule-set reward compared with Top-???? Confidence, Greedy Coverage, and Random selection. Rule regularization generally matches or improves the CNN baseline, achieving its largest Macro-F1 gain of +0.73 percentage points on MobileNetV3-S. The learned policy also yields highly diverse rule sets, with a 100% unique-sample ratio over 5,000 draws. Overall, the results indicate that GFlowNet is effective at learning a diverse distribution of high-reward symbolic rule sets, while the downstream benefit of rule regularization remains dependent on the CNN backbone and regularization strategy.

A distributed multi-agent framework for adaptive personalized online learning with dynamic knowledge graphs and learner modeling

Khue Minh Tran Nguyen — 2026-11-09

A distributed multi-agent framework is developed to enhance adaptive personalized online learning. The framework integrates Large Language Models (LLMs), a Dynamic Educational Knowledge Graph (DEKG), Knowledge Tracing (KT), Reinforcement Learning (RL), and Explainable Artificial Intelligence (XAI) to support learner modeling, adaptive recommendation, and explanation. The architecture separates learner-facing interaction from distributed domain processing while enabling coordinated knowledge representation and adaptive decision-making. A controlled simulation-based feasibility study evaluates learner-state estimation and personalized recommendation. The results show that graph-enhanced KT provides a modest improvement over the independent KT baseline, while the graph-aware RL policy achieves the highest final mastery and mastery gain among the evaluated strategies, although its advantage does not extend to every evaluation metric. The findings also indicate that recommendation performance is influenced by reward design. Overall, the simulation provides preliminary evidence of the feasibility of the proposed framework. Further evaluation using real educational datasets and learner interactions is required to assess its effectiveness and generalizability.

Intelligent location-aware recommender system for urban micro-connections using general graph matching

Huong Thao Tran — 2026-11-09

The rise of conurbations exacerbates "urban loneliness," yet current connection platforms suffer from appearance bias, high cancellation rates, and inadequate safety constraints. To address these issues, our study proposes.SLOT, a real-time 1-on-1 social interaction system. It integrates anonymous recognition and a Location-Aware Recommender System to securely anchor meeting points in public spaces, optimizing bilateral converging distances. The core contribution is the multi-objective matching engine, solving the allocation problem on a generalized graph using Edmonds' Blossom algorithm. Combining time-window batching, elastic pruning, and symmetric credit scoring, our system empirically achieves a balanced multi-objective trade-off among liquidity, converging distance, and fairness. Compared to Weighted-sum Greedy at 3.8 and Distance-based Greedy at 13.0,.SLOT compresses credit deviation to 0.9 with a minor 1.8% trade-off in matching rate. Simulations on a synthetic dataset of 1,000 agents within a university campus micro-zone demonstrate that by spatially decomposing requests into micro-Anchor Stations, the system achieves sub-second latency, mitigating the O(V³) computational bottleneck of direct global batching.

Reversible data hiding in encrypted images based on overlapping block partitioning and shared-MSB encoding

Hoang Nam Tram — 2026-11-09

Reversible data hiding in encrypted images (RDHEI) allows secret data to be embedded into an encrypted image while enabling lossless recovery of the original image. Its practical payload, however, remains limited by the embedding room available in the encrypted domain. This paper proposes an RDHEI method based on overlapping sub-block partitioning and shared-MSB encoding to improve the net embedding capacity. Each encrypted image block is organized into four overlapping sub-blocks. A boundary pixel shared by two adjacent sub-blocks serves as the reference pixel of the next sub-block, thereby increasing pixel reuse. The shared most significant bits within each sub-block are then exploited to create embedding room. Experiments on six standard images show an average net embedding capacity of 408,730 bits, representing improvements of 5.61% and 17.88% compared with two existing methods. Additional experiments on 100 randomly selected images from each of the BOWS-2 and BOSSBase datasets also show higher average net embedding capacities. Moreover, the proposed method provides error-free data extraction, exact image recovery, and separable operation according to the available keys.

SymbiosisMatch: An AI-augmented decision-support tool for energy and carbon symbiosis in industrial parks

Zheng Grace Ma — 2026-11-09

Early-stage screening of industrial symbiosis is constrained by incomplete data and limited representation of energy storage. This study presents SymbiosisMatch, a browser-based decision-support framework for identifying energy and carbon symbiosis opportunities in industrial parks. It combines editable park profiles with deterministic, storage-aware multi-criteria scoring of 15 solution archetypes, separate robustness and confidence indicators, and physical-output safeguards. An optional large language model interprets results without modifying calculations. Evaluation uses an illustrative mixed-industrial-park profile, four controlled storage configurations, and a contrasting electricity-focused scenario. Battery and thermal storage contributions change compatible solution scores by 6–10 points, altering heat-recovery and electricity-flexibility priorities while leaving unrelated scores unchanged. Sensitivity analysis shows that stable rankings can coexist with low input confidence, underscoring the distinction between compatibility and supporting evidence. Unsupported impact estimates are withheld when necessary physical inputs are unavailable. SymbiosisMatch provides traceable candidate prioritisation and identifies information needs before detailed technical and economic feasibility assessment.

Vichat: summary-augmented contextual retrieval for Vietnamese intangible cultural heritage question answering

Minh Tan Tran — 2026-11-09

Retrieval-augmented generation over cultural heritage archives is undermined by semantic isolation; fixed-size chunking strips the topic, temporal, and geographic context that makes a passage interpretable, so a chunk reading recognized as national heritage in 2014 no longer identifies which festival or province it describes. Contextual Retrieval repairs this by having an LLM contextualize each chunk, but at one LLM call per chunk. We propose Summary-Augmented Contextual Retrieval (SACR), which summarizes each document once and prepends that single summary to all of its chunks before indexing, reducing cost to one call per document. We further release ViCH-QA, an expert-validated benchmark of 4,719 Vietnamese question-answer (QA) pairs over 159 documents spanning 18 intangible cultural heritage (ICH) topics of Southern Vietnam. Evaluating lexical, dense, and hybrid retrievers under two Vietnamese word segmenters and an unsegmented baseline, we find that SACR raises NDCG@10 in all six matched configurations tested, most sharply for unreranked BM25 (+0.0394), reaching 0.8840 at best. The improvement reflects an inherent trade-off: SACR concentrates relevant passages at the top of the ranking while slightly reducing Hit@10, which is favorable for downstream generation but less suited for broad recall-oriented retrieval. Across six open-source generator backbones, the optimal pipeline is invariably SACR with reranking and few-shot prompting, peaking at a token-F1 of 0.5560; since SACR’s cost is paid offline and few-shot prompting is faster than chain-of-thought, our most accurate configuration is also among the cheapest to serve.

MIRAGE-SR: Missing-modality Invariant Representation And Generative Enrichment for multimodal Session-based Recommendation

Thi Cam Nhung Mai — 2026-11-09

Multimodal session-based recommendation commonly assumes that ID, visual, and textual features are available for every item, but this assumption is difficult to satisfy in real-world e-commerce data. This paper proposes MIRAGE-SR, a session-graph neural network architecture designed to maintain recommendation quality when one or more modalities are missing. The method combines two-stage modality imputation over an item-item co-occurrence graph, modality-invariant and modality-specific representation learning, and a missing-modality-aware late-fusion gate trained with curriculum modality dropout. MIRAGE-SR is evaluated on two subsets of Amazon Reviews 2023 under two settings: simulated modality cold-start on Grocery and naturally missing modalities on Cellphones. In the cold-start scenario, the configurations combining modality imputation and curriculum modality dropout achieve higher P@20 and MRR@20 than the zero-imputation configuration without curriculum modality dropout. On Cellphones, the two-stage GMI configuration achieves a P@20 of 26.24% and an MRR@20 of 14.61%, improving upon zero imputation by 3.91 and 1.53 points, respectively. The results indicate that the configurations combining imputation and training-time modality masking improve performance over the zero-imputation configuration; the current evaluation does not isolate the contribution of each component. These findings provide initial evidence of the potential of MIRAGE-SR for incomplete multimodal sequential recommendation.