Novel Recommender Systems Using Personalized Sentiment Mining

Abstract

Websites like Netflix, Amazon, Yelp have lot of reviews and ratings. Ratings are of usually on a scale of 1-5 points or stars. Reviews are free-form text consisting of a few sentences. Text sentiment analysis classification has occupied a pivotal role in sentiment analysis research as it offers important opinion mining options. Using these reviews and ratings of a person, we can recommend him new products, movies, restaurants. Usually Recommender systems match user patterns by finding similar users and recommendations are developed. We solve the problem of taking into account user’s personal sentiments and judgments making the recommendations more directed and useful to him. In this paper we provide an example using movies from the MovieLens dataset. Making recommendations using sentiment tags along with usual recommendations proves to be a novel and more intuitive way from a user’s likability of the movie and preferences standpoint. We analyze various methods like Unigrams, Bigrams, Support vector machines, Bernoulli Naive Bayes, Random Forests on popular datasets like Yelp, MovieLens to get the best method for sentiment generation. Then we introduce a novel recommendation systems combining the Alternate Least Square (ALS) method and sentiment generation recommending movies to users which proved to give lesser RMSE than traditional models on the MovieLens Dataset.