Haize Hu
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
Research Subjects
Research Interests
Haize Hu has a total of 42 co-authors. Here are the top co-authors by number of shared publications:
Mengge Fang
Guangxi Normal University
8 shared publicationsJianxun Liu
School of Computer Science and Engineering Hunan University of Science and Technology Xiangtan China
6 shared publications- Jingli Wu
Guangxi Normal University
5 shared publications Xiangping Zhang
Hunan University of Science and Technology
5 shared publications- Jiafei Liu
Guangxi Normal University
4 shared publications - Gaoshi Li
Guangxi Normal University
2 shared publications
Latest Publications
- 1 September 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Biology and Bioinformatics
- Vol. 23 (5) , 2026-2037
Identifying Cancer Driver Genes Based on Vision Transformer and Hierarchical Feature Fusion Module
Cancer is a global public health problem that poses a huge threat to human life and health. The identification of cancer driver genes helps to discover the intrinsic mechanisms of cancer occurrence and development. Existing deep learning-based methods for identifying cancer driver genes are mainly biased towards local structure reconstruction, limiting the global learning. In order to increase the identification capacity of cancer driver genes, in this study, a new cancer driver genes prediction method VGDriver using Vision Transformer(ViT) and hierarchical feature fusion module is proposed. Its steps are as follows: firstly, the gene expression, copy number aberrations, and single nucleotide variations data are used to construct a gene property feature matrix; secondly, a new hierarchical feature fusion module is designed, which can perform a two-pipeline fusion of multi-omics features to obtain a multi-source feature matrix by co-designing the graph convolutional networks block and the multilayer perceptron block; finally, the multi-source feature matrix is input into ViT to train and perform the task of predicting cancer driver genes. The results of comparative experiments on five different protein-protein interaction datasets for pan-cancer and eight single cancer types show that the VGDriver method is better compared to seven representative methods in the evaluation metrics AUC and AUPR.
- 1 September 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Biology and Bioinformatics
- Vol. 23 (5) , 2026-2037
Identifying Cancer Driver Genes Based on Vision Transformer and Hierarchical Feature Fusion Module
Cancer is a global public health problem that poses a huge threat to human life and health. The identification of cancer driver genes helps to discover the intrinsic mechanisms of cancer occurrence and development. Existing deep learning-based methods for identifying cancer driver genes are mainly biased towards local structure reconstruction, limiting the global learning. In order to increase the identification capacity of cancer driver genes, in this study, a new cancer driver genes prediction method VGDriver using Vision Transformer(ViT) and hierarchical feature fusion module is proposed. Its steps are as follows: firstly, the gene expression, copy number aberrations, and single nucleotide variations data are used to construct a gene property feature matrix; secondly, a new hierarchical feature fusion module is designed, which can perform a two-pipeline fusion of multi-omics features to obtain a multi-source feature matrix by co-designing the graph convolutional networks block and the multilayer perceptron block; finally, the multi-source feature matrix is input into ViT to train and perform the task of predicting cancer driver genes. The results of comparative experiments on five different protein-protein interaction datasets for pan-cancer and eight single cancer types show that the VGDriver method is better compared to seven representative methods in the evaluation metrics AUC and AUPR.
- 1 August 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 22 (8) , 6743-6754
HFGCS: Industrial Code Search With Sample-Aware Hierarchical Fusion and Hub-Centric Heterogeneous Graph Reasoning for Reliable CPS Software Maintenance
- 1 August 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 22 (8) , 6743-6754
HFGCS: Industrial Code Search With Sample-Aware Hierarchical Fusion and Hub-Centric Heterogeneous Graph Reasoning for Reliable CPS Software Maintenance
- Open Access
- 24 July 2026
- Published by Elsevier in SSRN Electronic Journal
DCLM-impute: A Dual Contrastive Learning Framework with Multi-constraint for Imputing Single-cell RNA Sequencing Data
In single-cell RNA sequencing data analyses, massive dropout events markedly degrade the accuracy of downstream studies. Although various imputation methods have been developed, they suffer from two major limitations: an inability to balance local and global information, and a tendency to overlook the inherent complexity of biological data, often falling into local optima. To improve imputation performance, we propose DCLM-impute, a novel framework that integrates a dual contrastive learning architecture with a multi-constraint optimization strategy. The dual contrastive architecture is designed to capture and balance both local and global cellular features, while the multi-constraint strategy guides the model to capture the complex nature of biological systems and avoid local optima. DCLM-impute is comprehensively compared to 13 representative methods on five real datasets and one simulated dataset. The results demonstrate that DCLM-impute consistently outperforms all baselines across a wide range of downstream experiments, including the identification of dropout and real biological zeros, gene expression recovery, cell clustering, differential expression analysis, and trajectory inference. For example, under a 40% dropout rate on the Zeisel dataset, DCLM-impute achieves Pearson correlation coefficient improvements of 3.7%–31.0% and L1-distance improvementsof 2.1%–35.4% in gene expression recovery. Furthermore, ablation studies validate that both the dual contrastive learning architecture and multi-constraint optimization are essential for achieving the stateof-the-art performance.
- 24 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDCLM-impute: A Dual Contrastive Learning Framework with Multi-constraint for Imputing Single-cell RNA Sequencing Data
In single-cell RNA sequencing data analyses, massive dropout events markedly degrade the accuracy of downstream studies. Although various imputation methods have been developed, they suffer from two major limitations: an inability to balance local and global information, and a tendency to overlook the inherent complexity of biological data, often falling into local optima. To improve imputation performance, we propose DCLM-impute, a novel framework that integrates a dual contrastive learning architecture with a multi-constraint optimization strategy. The dual contrastive architecture is designed to capture and balance both local and global cellular features, while the multi-constraint strategy guides the model to capture the complex nature of biological systems and avoid local optima. DCLM-impute is comprehensively compared to 13 representative methods on five real datasets and one simulated dataset. The results demonstrate that DCLM-impute consistently outperforms all baselines across a wide range of downstream experiments, including the identification of dropout and real biological zeros, gene expression recovery, cell clustering, differential expression analysis, and trajectory inference. For example, under a 40% dropout rate on the Zeisel dataset, DCLM-impute achieves Pearson correlation coefficient improvements of 3.7%–31.0% and L1-distance improvementsof 2.1%–35.4% in gene expression recovery. Furthermore, ablation studies validate that both the dual contrastive learning architecture and multi-constraint optimization are essential for achieving the stateof-the-art performance.
- 5 January 2026
- Published by Wiley in Computational Intelligence
- Vol. 42 (1)
AI ‐Smart Classroom for English Translation: An Adaptive HMM ‐Based Framework
This study develops an AI‐enhanced smart classroom framework to overcome the limitations of traditional English translation instruction, particularly in addressing the theory‐practice gap. We designed a three‐phase intelligent teaching system incorporating Hidden Markov Models (HMM) for: (1) adaptive pre‐class preparation, (2) immersive virtual translation scenarios, and (3) automated post‐class assessment. A 6‐week controlled experiment ( N = 100) compared this approach with traditional instruction using quantitative metrics including engagement levels, translation accuracy, proficiency rates, and satisfaction surveys. The experimental group showed statistically significant improvements ( p < 0.05): 12.7% higher engagement ( d = 1.21), 8.3% better culture‐specific translation accuracy, 6.9% faster proficiency attainment, and 89.2% satisfaction rate (vs. 82.1% control). HMM analysis effectively tracked learning progression and identified competency gaps. The study demonstrates HMM's effectiveness for modeling translation competence development and validates AI‐enhanced instruction as a viable solution for translation education. The implemented framework offers a replicable model for intelligent language learning systems.
- 5 January 2026
- Published by Wiley in Computational Intelligence
- Vol. 42 (1)
AI ‐Smart Classroom for English Translation: An Adaptive HMM ‐Based Framework
This study develops an AI‐enhanced smart classroom framework to overcome the limitations of traditional English translation instruction, particularly in addressing the theory‐practice gap. We designed a three‐phase intelligent teaching system incorporating Hidden Markov Models (HMM) for: (1) adaptive pre‐class preparation, (2) immersive virtual translation scenarios, and (3) automated post‐class assessment. A 6‐week controlled experiment ( N = 100) compared this approach with traditional instruction using quantitative metrics including engagement levels, translation accuracy, proficiency rates, and satisfaction surveys. The experimental group showed statistically significant improvements ( p < 0.05): 12.7% higher engagement ( d = 1.21), 8.3% better culture‐specific translation accuracy, 6.9% faster proficiency attainment, and 89.2% satisfaction rate (vs. 82.1% control). HMM analysis effectively tracked learning progression and identified competency gaps. The study demonstrates HMM's effectiveness for modeling translation competence development and validates AI‐enhanced instruction as a viable solution for translation education. The implemented framework offers a replicable model for intelligent language learning systems.
- 9 November 2025
- Published by Elsevier in Pattern Recognition
- Vol. 172
DEGAN-CS: An efficient code search model based on dataenhanced optimization of generative adversarial networks
- 9 November 2025
- Published by Elsevier in Pattern Recognition
- Vol. 172
DEGAN-CS: An efficient code search model based on dataenhanced optimization of generative adversarial networks
- 16 October 2025
- Published by Elsevier in Neurocomputing
- Vol. 659
FNCD-CS: A neurocomputing-optimized hybrid neural network for cross-domain code search
- 16 October 2025
- Published by Elsevier in Neurocomputing
- Vol. 659
FNCD-CS: A neurocomputing-optimized hybrid neural network for cross-domain code search
- 10 September 2025
- Published by Wiley in Concurrency and Computation: Practice and Experience
- Vol. 37 (23-24)
A Survey of the Full Process of Code Search Based on Deep Learning
As a pivotal technology for enhancing software development efficiency, research on code search based on deep learning has emerged as a current hotspot. This review systematically deconstructs the entire code search process into four core stages: dataset construction, code preprocessing, heterogeneous representation model construction, and query expansion, while conducting an in‐depth analysis of the application status and challenges of deep learning technologies. In dataset construction, the Q and A pairs and C and D pairs relied on by deep learning models suffer from a lack of standardization. For example, CodeSearchNet exhibits insufficient cross‐lingual versatility, and CoDesc has incomplete noise filtering. During the code preprocessing stage, bottlenecks such as AST granularity selection and sequence information redundancy restrict the efficiency of feature extraction. Although the introduction of transformer and graph neural networks has optimized structural representation, a unified evaluation mechanism is lacking. In the research of heterogeneous representation models, while LSTM, CNN, and pretrained models (such as CodeBERT) effectively narrow the semantic gap, their cross‐domain search accuracy is insufficient. In terms of query expansion, deep learning‐based keyword expansion and intent completion methods struggle to capture users' real needs due to low semantic alignment accuracy. This review proposes, for the first time, a standardized dataset construction framework integrating multimodal data, a syntax‐semantic dual‐layer preprocessing evaluation mechanism, a cross‐domain transfer representation model, and a large language model‐driven intent dynamic expansion scheme. These contributions lay a theoretical foundation for the systematic development of code search technologies and provide cross‐task methodological references for related fields such as code clone detection.
- 10 September 2025
- Published by Wiley in Concurrency and Computation: Practice and Experience
- Vol. 37 (23-24)
A Survey of the Full Process of Code Search Based on Deep Learning
As a pivotal technology for enhancing software development efficiency, research on code search based on deep learning has emerged as a current hotspot. This review systematically deconstructs the entire code search process into four core stages: dataset construction, code preprocessing, heterogeneous representation model construction, and query expansion, while conducting an in‐depth analysis of the application status and challenges of deep learning technologies. In dataset construction, the Q and A pairs and C and D pairs relied on by deep learning models suffer from a lack of standardization. For example, CodeSearchNet exhibits insufficient cross‐lingual versatility, and CoDesc has incomplete noise filtering. During the code preprocessing stage, bottlenecks such as AST granularity selection and sequence information redundancy restrict the efficiency of feature extraction. Although the introduction of transformer and graph neural networks has optimized structural representation, a unified evaluation mechanism is lacking. In the research of heterogeneous representation models, while LSTM, CNN, and pretrained models (such as CodeBERT) effectively narrow the semantic gap, their cross‐domain search accuracy is insufficient. In terms of query expansion, deep learning‐based keyword expansion and intent completion methods struggle to capture users' real needs due to low semantic alignment accuracy. This review proposes, for the first time, a standardized dataset construction framework integrating multimodal data, a syntax‐semantic dual‐layer preprocessing evaluation mechanism, a cross‐domain transfer representation model, and a large language model‐driven intent dynamic expansion scheme. These contributions lay a theoretical foundation for the systematic development of code search technologies and provide cross‐task methodological references for related fields such as code clone detection.