Detecting aspect-based review spam based on fuzzy reasoning

Recently, there appears aspect-based review spam as respect to different using aspects of merchandises nominated by e-commerce platforms. Most existing solutions cannot be transferred to manage this new problem because they were trained upon aspect-free texts. The primary challenge is twofold, e.g., the weak supervision problem and the sentiment distribution mining across many an aspect. In this paper, we first collect public reviews and label them via a clustering technique and recruit several volunteers to forge spammed reviews, and then apply sentiment analysis algorithms to mine part-of-speech differences between positive and negative reviews, and finally propose a fuzzy ellipsoid reasoning algorithm to identify review spam. Simulative experiments show that our approach outperforms state-of-the-art solutions upon precision (1.5 points), recall (1.6 points), F1 score (1.0 points) and AUC (1.0 points) with small margins, whether in detecting volunteer-written spam or LLM-generated spam. As a symbolic reasoning method, our method is born to have advantages on interpretability, thus it can also be improved backwards for subsequent research.