Comments Mining And Application In E-commerce

User’s feedback comments in e-commerce system not only provide important reference for potential customers but also help manufactures to understand consumer opinions and market status. Unscrupulous business men hired some people deliberately to publish fake comments in order to increase their own interests. Fake reviews mislead consumers’ decision. It also let consumers loss confidence in e-commerce evaluation system. Detect spammers accurately and efficiently is more and more important.This thesis aims to solve these problems above. Based on human behavior theory, we analysis the features of human behavior in e-commerce in-depth. We extract users’ reviews behavioral characteristics, using pattern mining method to detect spammer and spammer group. The main contents are as below:1. We use detrended fluctuation analysis method to analysis the scaling law of human behavior in e-commerce system based on actual real data. Results show that human’s purchasing and browsing behavior has long-range correlations and is self-organized. They match the kinetic mechanism of deviation of Poisson.2. Single spammer detection. We proposed two algorithms to detect single spammer in e-commerce. One is rating behavior clustered algorithm. It use a two-stage method to give each user’s reputation after computing users’ herd clustering strength. The other detect algorithm is called the evolution of users’ experience detect method. Based on the evolution of users’ experience over time, we proposed two characteristics to describe reviewers’ experience value. One is stability and the other is called disorder. Experiments show that both detect algorithms have high accuracy and efficiency, better than other baseline algorithm, especially in the real system which users’ rating ratio is very sparse.3. Spammer group detect. Based on the assumption of maximize benefits, we mining spammer group by association rule. This paper proposed algorithm not only can find out the size of spammer group but also detect specified size of spammer group effectively.4. We implements the algorithms above and complete the spammer detect system. This system can import data, detect spammer and spammer group and it also can check and mark the result.The innovation of this thesis is that we use consumer rating behavior feature to detect spammer besides specific review contents. We propose two spammer detect methods based on rating behavior cluster and user experience evolution. Both algorithms improve the efficiency as well as high detection accuracy.