Recommender systems heavily rely on modeling accurate user profiles and preferences. For place recommender systems, the authors present the opinion-based user profile model. A user’s rating (like or dislike) groups the reviews of other users into positive and negative subsets, and reviews in each subset are used to represent the user’s “positive profile” and “negative profile.” Four representation models of the positive (negative) profiles are used: (1) full reviews (FR) using bag of words; (2) selective term-based reviews (SR) using most frequent terms; (3) nouns-only reviews (NR); and (4) concise review summaries (RS). Similarly, the candidate places are represented with the positive (negative) reviews. A linear combination of similarity measures between the positive (negative) user profiles and positive (negative) candidate places are proposed for ranking candidate places. The recommendation summary can be customized referring to the positive profile features that are common in the candidate place model. Experimental results show that the opinion-based suggestion performs better than the category- or description-based profiling baselines. The noun-based review model (NR) for user profiling outperforms the other representations. This approach is also robust even when there are few reviews available.

The current user profiling is based on splitting the reviews based on the positive and negative ratings, but one can devise finer profiling not only by the opinions but by categories of places. Related work should cover the literature on collaborative filtering and other opinion-based recommender systems.