講演情報
[5Yin-A-38]Optimizing High-Dimensional Review Features: Via LLM Semantic Extraction and Lightweight mRMR Feature Selection Network Architecture
〇Hengshuo Yang1, Tengfei Shao1, Masayuki Goto1 (1. Waseda University)
キーワード：
Interpretable AI、Large Language Models、ML
Understanding how specific service dimensions drive customer ratings is essential for effective restaurant management, yet most existing review-based models rely on opaque text embeddings with limited interpretability. This study proposes an interpretable review rating prediction framework that integrates domain knowledge with large language models (LLMs).We construct a human-guided modular ontology grounded in service quality theories and employ an LLM to extract fine-grained, interpretable experience factors from review texts. To mitigate redundancy and noise in the extracted features, we introduce an incremental marginal-gain-based mRMR selector with conditional redundancy penalization and data-driven early stopping. The selected factors are incorporated into gradient boosting models (LightGBM and CatBoost) for rating prediction. Experiments on the Yelp dataset demonstrate that the proposed framework achieves strong predictive performance with a compact feature set, while enabling transparent attribution of ratings to concrete service attributes.