Diagnosing Electric Vehicle Failures With Long-Tailed Multilabel Consumer Reviews: A Bilateral-Branch Deep Learning Approach
针对消费者评论中多故障共现和长尾分布问题，提出双边分支深度网络实现细粒度多标签分类，帮助车企精准识别故障类别。
Consumer reviews on social media is providing valuable opportunities for electric vehicle companies to investigate product failure information. By noticing the fact that consumers often report multiple issues about electric vehicle failures within a single comment and failure classes usually exhibit a long-tailed distribution, we first introduce a bilateral-branch deep learning network to achieve a fine-grained multilabel classification of online reviews, which is different from the existing research with coarse-grained binary classification. In this case, considering the long-tailed characteristic of electric vehicle failure classes, each branch of the deep learning network is designed to focus on representation learning and classification learning with corresponding sampling methods, respectively. A novel multilabel contrastive learning method with label-description embedding is proposed to enhance the model's efficiency. We also adopt a progressive learning approach with a parabolic decay strategy to facilitate knowledge transfer between the two branches. Finally, our proposed method has been demonstrated to be effective via comparative experiments with other deep learning methods and ablation studies. Our proposed method dissects electric vehicle consumer reviews with fine granularity, which efficaciously supports enterprises in identifying specific electric vehicle failure categories.