Unsupervised Rating Prediction based on Local and Global Semantic Models.
Adrian Boteanu, Sonia Chernova · 2013
Current recommendation engines attempt to answer the same question: given a user with some activity in the system, which is the next entity, be it a restaurant, a book or a movie, that the user should visit or buy next. The presumption is that the user would favorably review the item being recommended. The goal of our project is to predict how a user would rate an item he/she never rated, which is a generalization of the task recommen-dation engines perform. Previous work successfully em-ploys machine learning techniques, particularly statisti-cal methods. However, there are some outlier situations which are more difficult to predict, such as new users. In this paper we present a rating prediction approach tar-geted for entities for which little prior information ex-ists in the database. We put forward and test a number of hypotheses, exploring recommendations based on near-est neighbor-like methods. We adapt existing common sense topic modeling methods to compute similarity measures between users and then use a relatively small set of key users to predict how the target user will rate a given business. We implemented and tested our system for recommending businesses using the Yelp Academic Dataset. We report initial results for topic-based rating predictions, which perform consistently across a broad range of parameters.