计算机应用研究2026，Vol.43Issue(8)：2270-2277,8.DOI:10.19734/j.issn.1001-3695.2025.12.0507
ALU-TransSHAP:基于主动学习的无偏TransSHAP可解释模型
ALU-TransSHAP:unbiased TransSHAP explainable model based on active learning
摘要
Abstract
With the widespread application of deep learning models in sentiment analysis,the"black-box"nature induces problems like opaque prediction processes and difficult tracing of key decision-making factors,which seriously impairs model credibility and practical deployment.This study proposed an active learning-based unbiased TransSHAP explanation model(ALU-TransSHAP).The model screened high-entropy samples via an active learning background selection module to mitigate the distribution bias of background data,optimized the TransSHAP adaptation layer to construct a bidirectional mapping be-tween short sentences and subwords for aligning semantic units with explanation units,and designed an unbiased Shapley value calculation engine integrated with paired sampling to enhance the accuracy and stability of feature attribution.Experiments on the Weibo hot major review dataset and SENTI_RATIONALE dataset demonstrate that ALU-TransSHAP outperforms all base-line models significantly in fidelity(0.89/0.87),stability(0.92/0.88)and prediction change(0.83/0.92),while main-taining acceptable sparsity(0.209/0.211)and computational efficiency(2.5/2.7).The proposed ALU-TransSHAP effec-tively addresses the interpretation challenges of Transformer models in Chinese scenarios,fully preserves semantic logic,pro-vides reliable explanatory support for sentiment analysis,and remarkably improves the transparency and credibility of model decision-making.关键词
模型可解释性/SHAP/情感分析/深度学习Key words
model interpretability/SHAP/sentiment analysis/deep learning分类
信息技术与安全科学引用本文复制引用
刘彤,杨雅萱,倪维健..ALU-TransSHAP:基于主动学习的无偏TransSHAP可解释模型[J].计算机应用研究,2026,43(8):2270-2277,8.基金项目
山东省自然科学基金资助项目(ZR2022MF319) （ZR2022MF319）
科技创新2030—"新一代人工智能"重大项目(2022ZD0119502-07) （2022ZD0119502-07）
新一代人工智能国家科技重大专项(2022ZD0119501) （2022ZD0119501）
