Recommendation based on Mining Product Reviewers' Preference Similarity Network
Feng Wang, Li Chen · 2012
Most products in e-commerce are with high cost (e.g., digital cameras, computers) and hence less likely experienced by users. Thus, the traditional recommender techniques (such as user-based collaborative ltering and content-based methods) are not applicable, because they largely assume that the users have prior experiences with the items. The ew user is hence a typical phenomenon and challenging issue that recommender systems face in this environment. In this paper, we have particularly proposed to build product reviewers’ preference similarity network to solve this problem. Specically, with product reviews, we have attempted to recover each reviewer’s weight preferences over features, based on which all reviewers can be connected in an implicit network with the edge denoting their pairwise preference similarity. We additionally adopted the Latent Class Regression Model (LCRM) to identify the sub-communities in this network, where each sub-community corresponds to a cluster of like-minded reviewers. The new user’s stated feature preferences can be then matched to the cluster with most relevant reviewers and their reviewed products can be taken as recommendation candidates. The experimental results reveal that our novel method outperforms others that either did not consider reviews, or did not attempt to reconstruct reviewers’ inherent feature preferences (from their written reviews) for beneting the recommendation process. The LCRM-based clustering method was also proven with higher accuracy than related ones (like k-Means based clustering), especially when the new user’s stated preferences were less complete.