Suhua Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Suhua Wang has a total of 40 co-authors. Here are the top co-authors by number of shared publications:
- Xiaoxin Sun
School of Information Science and Technology, Northeast Normal University, Changchun 130117, China
4 shared publications - Hongjie Ji
School of Information Science and Technology, Northeast Normal University, Changchun 130117, China
2 shared publications Rui Ma
Changchun Humanities and Sciences College
2 shared publications- Xin Li
College of Optical Science and Engineering, Zhejiang University, Hangzhou, P. R. China
2 shared publications Xiuzhuo Wei
Changchun Humanities and Sciences College (China)
2 shared publications- Yun Zhang
College of Computer Science, Changchun Humanities and Sciences College, Changchun 130117, China
2 shared publications
Latest Publications
- 14 July 2026
- Published by Springer Nature
- p. 51-61
ManchuTTS: Towards High-Quality Manchu Speech Synthesis via Flow Matching and Hierarchical Text Representation
- 14 July 2026
- Published by Springer Nature
- p. 51-61
ManchuTTS: Towards High-Quality Manchu Speech Synthesis via Flow Matching and Hierarchical Text Representation
- 9 May 2026
- Published by World Scientific Pub Co Pte Ltd in International Journal on Artificial Intelligence Tools
- Vol. 35 (03)
Transformer–Based Models for Low–Resource Language Processing in Specialized Technical Domains
Natural language processing encounters major difficulties when processing low-resource languages because technical fields require specialized terminology and lack sufficient annotated data. The performance of transformer-based multilingual models depends on their ability to operate in different domains and their available resources for training. The researchers developed a transformer-based framework that uses XLM-RoBERTa (XLM-R) to process languages that have limited resources for their needs. The proposed methodology includes systematic data preprocessing methods that create low-resource data through data augmentation and implement a two-stage domain adaptation approach that begins with pretraining on unlabeled technical text before moving to task-specific classification fine-tuning. The FLORES-200 multilingual dataset provides a structured benchmark assessment that tests low-resource conditions through its multiple testing methods. Model performance assessment uses standard evaluation metrics, which include accuracy and precision and recall and [ ]1-score, together with confusion matrix analysis and evaluation curve measurements. The experimental results show that the proposed approach achieved 0.98889 accuracy, 0.98925 precision, 0.98889 recall, and 0.98889 [ ]1-score, which demonstrates strong accurate classification abilities. The study shows that multilingual transformer models, which use domain adaptation and data augmentation methods, deliver effective solutions to data scarcity challenges while improving performance in technical fields.
- 9 May 2026
- Published by World Scientific Pub Co Pte Ltd in International Journal on Artificial Intelligence Tools
- Vol. 35 (03)
Transformer–Based Models for Low–Resource Language Processing in Specialized Technical Domains
Natural language processing encounters major difficulties when processing low-resource languages because technical fields require specialized terminology and lack sufficient annotated data. The performance of transformer-based multilingual models depends on their ability to operate in different domains and their available resources for training. The researchers developed a transformer-based framework that uses XLM-RoBERTa (XLM-R) to process languages that have limited resources for their needs. The proposed methodology includes systematic data preprocessing methods that create low-resource data through data augmentation and implement a two-stage domain adaptation approach that begins with pretraining on unlabeled technical text before moving to task-specific classification fine-tuning. The FLORES-200 multilingual dataset provides a structured benchmark assessment that tests low-resource conditions through its multiple testing methods. Model performance assessment uses standard evaluation metrics, which include accuracy and precision and recall and [ ]1-score, together with confusion matrix analysis and evaluation curve measurements. The experimental results show that the proposed approach achieved 0.98889 accuracy, 0.98925 precision, 0.98889 recall, and 0.98889 [ ]1-score, which demonstrates strong accurate classification abilities. The study shows that multilingual transformer models, which use domain adaptation and data augmentation methods, deliver effective solutions to data scarcity challenges while improving performance in technical fields.
- Open Access
- 12 May 2025
- Published by Springer Nature in Journal of King Saud University Computer and Information Sciences
- Vol. 37 (3) , 1-18
CASEMark: A hybrid model for robust anatomical landmark detection in multi-structure X-rays
Anatomical landmark detection is crucial in medical image analysis, facilitating accurate diagnosis, surgical planning, and treatment evaluation. However, existing methods often struggle to simultaneously capture global context and local details while exhibiting limited generalization across diverse datasets and imaging modalities. To relieve this, we propose a hybrid model that leverages convolutional operations to capture local information and a Swin Transformer to enhance global context. Specifically, we introduce a novel U-shaped architecture, termed Convolutional Attention Swin Enhanced Landmark Detection Network (CASEMark). CASEMark integrates three key innovations: (1) a Convolutional Attention Swin Transformer module (CAST) that integrates transformer-based global context modeling with convolutional operations for local feature extraction, (2) an Enhanced Skip Attention Module (ESAM) enabling adaptive feature fusion between encoder and decoder pathways, and (3) a multi-resolution heatmap learning strategy that aggregates information across scales. This approach effectively balances global-local feature extraction with robust cross-modality generalization. Extensive experiments on four public datasets demonstrate the superiority of CASEMark. The code and datasets will be made publicly available.
- 12 May 2025
- Published by Springer Nature in Journal of King Saud University Computer and Information Sciences
- Vol. 37 (3) , 1-18
Open AccessCASEMark: A hybrid model for robust anatomical landmark detection in multi-structure X-rays
Anatomical landmark detection is crucial in medical image analysis, facilitating accurate diagnosis, surgical planning, and treatment evaluation. However, existing methods often struggle to simultaneously capture global context and local details while exhibiting limited generalization across diverse datasets and imaging modalities. To relieve this, we propose a hybrid model that leverages convolutional operations to capture local information and a Swin Transformer to enhance global context. Specifically, we introduce a novel U-shaped architecture, termed Convolutional Attention Swin Enhanced Landmark Detection Network (CASEMark). CASEMark integrates three key innovations: (1) a Convolutional Attention Swin Transformer module (CAST) that integrates transformer-based global context modeling with convolutional operations for local feature extraction, (2) an Enhanced Skip Attention Module (ESAM) enabling adaptive feature fusion between encoder and decoder pathways, and (3) a multi-resolution heatmap learning strategy that aggregates information across scales. This approach effectively balances global-local feature extraction with robust cross-modality generalization. Extensive experiments on four public datasets demonstrate the superiority of CASEMark. The code and datasets will be made publicly available.
- Open Access
- 11 September 2023
- Published by SPIE-Intl Soc Optical Eng
- Vol. 12779, 58-689
Graph neural network based entity augmented representation for recommendation system
A lot of rich auxiliary information can be integrated into the scoring and show good performance, which plays a good role in many recommended methods. In this paper, we use the rich semantic of knowledge graph to explore user preferences and improve the algorithm performance by using item attributes. A large number of experiments have been carried out on two data sets to prove the validity of the graph convolution recommendation model.
- 11 September 2023
- Published by SPIE-Intl Soc Optical Eng
- Vol. 12779, 58-689
Open AccessGraph neural network based entity augmented representation for recommendation system
A lot of rich auxiliary information can be integrated into the scoring and show good performance, which plays a good role in many recommended methods. In this paper, we use the rich semantic of knowledge graph to explore user preferences and improve the algorithm performance by using item attributes. A large number of experiments have been carried out on two data sets to prove the validity of the graph convolution recommendation model.
- Open Access
- 1 January 2023
- Published by American Institute of Mathematical Sciences (AIMS) in Electronic Research Archive
- Vol. 31 (2) , 793-811
An adaptive preference retention collaborative filtering algorithm based on graph convolutional method
Collaborative filtering is one of the most widely used methods in recommender systems. In recent years, Graph Neural Networks (GNN) were naturally applied to collaborative filtering methods to model users' preference representation. However, empirical research has ignored the effects of different items on user representation, which prevented them from capturing fine-grained users' preferences. Besides, due to the problem of data sparsity in collaborative filtering, most GNN-based models conduct a large number of graph convolution operations in the user-item graph, resulting in an over-smoothing effect. To tackle these problems, Adaptive Preference Retention Graph Convolutional Collaborative Filtering Method (APR-GCCF) was proposed to distinguish the difference among the items and capture the fine-grained users' preferences. Specifically, the graph convolutional method was applied to model the high-order relationship on the user-item graph and an adaptive preference retention mechanism was used to capture the difference between items adaptively. To obtain a unified users' preferences representation and alleviate the over-smoothing effect, we employed a residual preference prediction mechanism to concatenate the representation of users' preferences generated by each layer of the graph neural network. Extensive experiments were conducted based on three real datasets and the experimental results demonstrate the effectiveness of the model.
- 1 January 2023
- Published by American Institute of Mathematical Sciences (AIMS) in Electronic Research Archive
- Vol. 31 (2) , 793-811
Open AccessAn adaptive preference retention collaborative filtering algorithm based on graph convolutional method
Collaborative filtering is one of the most widely used methods in recommender systems. In recent years, Graph Neural Networks (GNN) were naturally applied to collaborative filtering methods to model users' preference representation. However, empirical research has ignored the effects of different items on user representation, which prevented them from capturing fine-grained users' preferences. Besides, due to the problem of data sparsity in collaborative filtering, most GNN-based models conduct a large number of graph convolution operations in the user-item graph, resulting in an over-smoothing effect. To tackle these problems, Adaptive Preference Retention Graph Convolutional Collaborative Filtering Method (APR-GCCF) was proposed to distinguish the difference among the items and capture the fine-grained users' preferences. Specifically, the graph convolutional method was applied to model the high-order relationship on the user-item graph and an adaptive preference retention mechanism was used to capture the difference between items adaptively. To obtain a unified users' preferences representation and alleviate the over-smoothing effect, we employed a residual preference prediction mechanism to concatenate the representation of users' preferences generated by each layer of the graph neural network. Extensive experiments were conducted based on three real datasets and the experimental results demonstrate the effectiveness of the model.
- 24 August 2022
- Published by Springer Nature
- p. 57-71
Using Meta Path Information to Adjust Embedded Recommendation System
- 24 August 2022
- Published by Springer Nature
- p. 57-71
Using Meta Path Information to Adjust Embedded Recommendation System
- 20 June 2022
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 78 (17) , 19322-19345
Leveraging side information as adjusting embedding to improve user representation for recommendations
- 20 June 2022
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 78 (17) , 19322-19345
Leveraging side information as adjusting embedding to improve user representation for recommendations
- Open Access
- 31 May 2022
- Published by MDPI AG in Mathematics
- Vol. 10 (11) , 1881
Enhanced Graph Learning for Recommendation via Causal Inference
The goal of the recommender system is to learn the user’s preferences from the entity (user–item) historical interaction data, so as to predict the user’s ratings on new items or recommend new item sequences to users. There are two major challenges: (1) Datasets are usually sparse. The item side is often accompanied by some auxiliary information, such as attributes or context; it can help to slightly improve its representation. However, the user side is usually presented in the form of ID due to personal privacy. (2) Due to the influences of confounding factors, such as the popularity of items, users’ ratings on items often have bias that cannot be recognized by the traditional recommendation methods. In order to solve these two problems, in this paper, (1) we explore the use of a graph model to fuse the interactions between users and common rating items, that is, incorporating the “neighbor” information into the target user to enrich user representations; (2) the operator is used to deduce the causality after removing the influences of confounding factors, rather than the correlation of the data surface fitted by traditional machine learning. We propose the EGCI model, i.e., enhanced graph learning for recommendation via causal inference. The model embeds user relationships and item attributes into the latent semantic space to obtain high-quality user and item representations. In addition, the mixed bias implied in the rating process is calibrated by considering the popularity of items. Experimental results on three real-world datasets show that EGCI is significantly better than the baselines.- 31 May 2022
- Published by MDPI AG in Mathematics
- Vol. 10 (11) , 1881
Open AccessEnhanced Graph Learning for Recommendation via Causal Inference
The goal of the recommender system is to learn the user’s preferences from the entity (user–item) historical interaction data, so as to predict the user’s ratings on new items or recommend new item sequences to users. There are two major challenges: (1) Datasets are usually sparse. The item side is often accompanied by some auxiliary information, such as attributes or context; it can help to slightly improve its representation. However, the user side is usually presented in the form of ID due to personal privacy. (2) Due to the influences of confounding factors, such as the popularity of items, users’ ratings on items often have bias that cannot be recognized by the traditional recommendation methods. In order to solve these two problems, in this paper, (1) we explore the use of a graph model to fuse the interactions between users and common rating items, that is, incorporating the “neighbor” information into the target user to enrich user representations; (2) the operator is used to deduce the causality after removing the influences of confounding factors, rather than the correlation of the data surface fitted by traditional machine learning. We propose the EGCI model, i.e., enhanced graph learning for recommendation via causal inference. The model embeds user relationships and item attributes into the latent semantic space to obtain high-quality user and item representations. In addition, the mixed bias implied in the rating process is calibrated by considering the popularity of items. Experimental results on three real-world datasets show that EGCI is significantly better than the baselines. - Open Access
- 4 February 2022
- Published by MDPI AG in Mathematics
- Vol. 10 (3) , 503
Neural Metric Factorization for Recommendation
All current recommendation algorithms, when modeling user–item interactions, basically use dot product. This dot product calculation is derived from matrix factorization. We argue that an inherent drawback of matrix factorization is that latent semantic vectors of users or items sometimes do not satisfy triangular inequalities, which may affect the performance of the recommendation. Recently, metric factorization was proposed to replace matrix factorization and has achieved some improvements in terms of recommendation accuracy. However, similar to matrix factorization, metric factorization still uses a simple, linear fashion. In this paper, we explore leveraging nonlinear deep neural networks to realize Euclidean distance interaction between users and items. We propose a generic Neural Metric Factorization Framework (NMetricF), which learns representations for users and items by incorporating Euclidean metric factorization into deep neural networks. Extensive experiments on six real-world datasets show that, compared to the previous recommendation algorithms based purely on rating data, NMetricF achieves the best performance.
- 4 February 2022
- Published by MDPI AG in Mathematics
- Vol. 10 (3) , 503
Open AccessNeural Metric Factorization for Recommendation
All current recommendation algorithms, when modeling user–item interactions, basically use dot product. This dot product calculation is derived from matrix factorization. We argue that an inherent drawback of matrix factorization is that latent semantic vectors of users or items sometimes do not satisfy triangular inequalities, which may affect the performance of the recommendation. Recently, metric factorization was proposed to replace matrix factorization and has achieved some improvements in terms of recommendation accuracy. However, similar to matrix factorization, metric factorization still uses a simple, linear fashion. In this paper, we explore leveraging nonlinear deep neural networks to realize Euclidean distance interaction between users and items. We propose a generic Neural Metric Factorization Framework (NMetricF), which learns representations for users and items by incorporating Euclidean metric factorization into deep neural networks. Extensive experiments on six real-world datasets show that, compared to the previous recommendation algorithms based purely on rating data, NMetricF achieves the best performance.
- Open Access
- 1 January 2022
- Published by American Institute of Mathematical Sciences (AIMS) in STEM Education
- Vol. 2 (2) , 157
Personalized exercise recommendation method based on causal deep learning: Experiments and implications
The COVID-19 pandemic has accelerated innovations for supporting learning and teaching online. However, online learning also means a reduction of opportunities in direct communication between teachers and students. Given the inevitable diversity in learning progress and achievements for individual online learners, it is difficult for teachers to give personalized guidance to a large number of students. The personalized guidance may cover many aspects, including recommending tailored exercises to a specific student according to the student′s knowledge gaps on a subject. In this paper, we propose a personalized exercise recommendation method named causal deep learning (CDL) based on the combination of causal inference and deep learning. Deep learning is used to train and generate initial feature representations for the students and the exercises, and intervention algorithms based on causal inference are then applied to further tune these feature representations. Afterwards, deep learning is again used to predict individual students′ score ratings on exercises, from which the Top-N ranked exercises are recommended to similar students who likely need enhancing of skills and understanding of the subject areas indicated by the chosen exercises. Experiments of CDL and four baseline methods on two real-world datasets demonstrate that CDL is superior to the existing methods in terms of capturing students′ knowledge gaps in learning and more accurately recommending appropriate exercises to individual students to help bridge their knowledge gaps.
- 1 January 2022
- Published by American Institute of Mathematical Sciences (AIMS) in STEM Education
- Vol. 2 (2) , 157
Open AccessPersonalized exercise recommendation method based on causal deep learning: Experiments and implications
The COVID-19 pandemic has accelerated innovations for supporting learning and teaching online. However, online learning also means a reduction of opportunities in direct communication between teachers and students. Given the inevitable diversity in learning progress and achievements for individual online learners, it is difficult for teachers to give personalized guidance to a large number of students. The personalized guidance may cover many aspects, including recommending tailored exercises to a specific student according to the student′s knowledge gaps on a subject. In this paper, we propose a personalized exercise recommendation method named causal deep learning (CDL) based on the combination of causal inference and deep learning. Deep learning is used to train and generate initial feature representations for the students and the exercises, and intervention algorithms based on causal inference are then applied to further tune these feature representations. Afterwards, deep learning is again used to predict individual students′ score ratings on exercises, from which the Top-N ranked exercises are recommended to similar students who likely need enhancing of skills and understanding of the subject areas indicated by the chosen exercises. Experiments of CDL and four baseline methods on two real-world datasets demonstrate that CDL is superior to the existing methods in terms of capturing students′ knowledge gaps in learning and more accurately recommending appropriate exercises to individual students to help bridge their knowledge gaps.