Customized reviews for small user-databases using iterative SVD and content based filtering

Abstract

Recommender systems have proven to be a valuable tool for web companies like Amazon and Netflix for attracting and maintaining a large user base. However, in situations when user data is more scarce (e.g., for mid-sized companies, or an online retailer testing a new ratings system) algorithms tailored to smaller datasets can be used to further increase accuracy. This paper explores the potential of combining collaborative and content-based (using user comments) filtering algorithms using Yelp.com data from a single city. We present the method to combine two approaches, and find that the MSE for predicting a user's new rating can be reduced from a baseline MSE of 1.744 to 0.934 given just 2500 rated items in our real-world dataset.