文献类型：期刊文献

英文题名：An attention-weighted Bayesian network learning approach for categorical representation of mixed data

作者：Li, Qiude Yang, Tingting Ji, Shengfen Yu, Yang Chen, Sen Hu, Zuquan Zeng, Zhu

第一作者：Li, Qiude

通信作者：Zeng, Z[1];Ji, SF[2]

机构：[1]Guizhou Med Univ, Sch Biol & Engn, Guiyang 561113, Guizhou, Peoples R China;[2]Guizhou Inst Technol, Sch Foreign Language, Guiyang 550003, Guizhou, Peoples R China

第一机构：Guizhou Med Univ, Sch Biol & Engn, Guiyang 561113, Guizhou, Peoples R China

通信机构：corresponding author), Guizhou Med Univ, Sch Biol & Engn, Guiyang 561113, Guizhou, Peoples R China;corresponding author), Guizhou Inst Technol, Sch Foreign Language, Guiyang 550003, Guizhou, Peoples R China.|贵州理工学院;

年份：2026

卷号：664

外文期刊名：NEUROCOMPUTING

收录：;WOS:【SCI-EXPANDED(收录号:WOS:001622661200004)】；

基金：We thank anonymous reviewers for their valuable comments and suggestions. This work was supported by the National Natural Science Foundation of China (Grant No. 62166009, 12132006) , Guizhou Provincial Basic Research Program (Natural Science) (Grant No. MS [2025] 553) , QKHJC- [2024] Youth 226) , Cultivation Project of National Natural Science Foundation of China at Guizhou Medical University (No. 25NSFCP20) , Science and Technology Foundation of Guizhou Provincial Health Commission (Grant No. gzwkj2023-258, 2025GZWJKJXM0914) , Youth Science and Technology Project of Guizhou Provincial Education Department (No. QJJ- [2024] 101) , and Medical Research Union Foundation for High-quality Health Development of Guizhou Province (No. 2024GZYXKYJJXM0119) .

语种：英文

外文关键词：Representation learning; Mixed data; Bayesian networks; Attribute weighting; Attention mechanism

摘要：It is an important but challenging task to build a representation learning method that can transform the categorical attribute values in mixed data into high-quality numerical vectors, by mining the complicated dependencies between categorical attributes and class labels. Many existing representation learning methods rely on the attribute conditional independence assumption, resulting in unsatisfactory performance when there is a strong dependency between attributes in a dataset or their instance size is insufficient. Therefore, this paper explores a way to relax this assumption by using Bayesian network method, and proposes an Attention-Weighted Bayesian Network learning approach (AWBN) for the categorical representation of mixed data. Specifically, we construct a high-dependency Bayesian network to relax this assumption, and subsequently represent each categorical value as a numerical vector according to this network topology. Then, we employ a fine-grained attribute weighting method to explore the differences in categorical values across various instances, drawing inspiration from the channel attention mechanism in deep learning to design an attention-weighted method based on a dual-attention mechanism (combining SENet (Squeeze-and-Excitation Networks) and dual-pooling SENet). Extensive experimental results on 32 mixed datasets show that AWBN significantly enhances its performance, compared with its state-of-the-art competitors.