基于多特征融合的智能客服用户评论行为预测及可解释性分析

Intelligent Customer Service User Comment Behavior Prediction and Explainability Analysis Based on Multi-Feature Fusion

摘要: 智能客服作为用户与平台交互的核心渠道，仍面临服务瓶颈与用户态度两极分化的问题，致使对用户行为预测变得至关重要。基于感知价值理论，本文借助BERTopic模型从评论数据中识别感知价值维度；并采用集成自注意力机制的双向长短期记忆(BiLSTM)模型识别感知价值评分与情感倾向。结合用户基本特征、感知价值特征与情感特征，通过常春藤算法优化BP神经网络(IVY-BP)对用户评论行为进行预测。结果表明，用户情感呈现显著的两极分化趋势；IVY-BP模型在准确率与F1分数上均优于PSO-BP与GA-BP模型，验证了多特征融合相较于单一特征的优越性；SHAP可解释分析显示，影响力、认证类型、功能价值与社会价值是评论行为的关键驱动因素。

Abstract: As the core channel for user-platform interaction, intelligent customer service systems still face service bottlenecks and user attitude polarization, making user behavior prediction critically important. Based on perceived value theory, this study utilizes the BERTopic model to identify perceived value dimensions from review data, and employs a bidirectional long short-term memory (BiLSTM) model with an integrated self-attention mechanism to analyze perceived value scores and emotional tendencies. By integrating user basic characteristics, perceived value features, and emotional traits, we optimize the backpropagation neural network (IVY-BP) using the Ivy algorithm to predict user review behaviors. Results demonstrate significant polarization in user emotions; the IVY-BP model outperforms PSO-BP and GA-BP models in both accuracy and F1 score, validating the superiority of multi-feature fusion over single-feature approaches. SHAP explanatory analysis reveals that influence level, certification type, functional value, and social value serve as key drivers of review behavior.