Abstract

Recommender systems provide personalized suggestions to users and are critical to the success of many e-commerce sites, such as Netflix and Amazon. Outside of e-commerce, recommender systems can be deployed in fields such as intelligence analysis, for recommending high-quality information source to analysts for further examination. In this work, we present the square counting method for rating predictions in recommender systems. Our method is based on analyzing the bipartite rating network with score-labeled edges representing user nodes’ ratings to item nodes. Edges are denoted as an I-love-it or I-hate-it edge based on whether the rating score on the edge is above or below a threshold. For a target user-item pair, we count the number for each configuration of love-hate squares that involve the target pair, where the sequence of I-love-it or I-hate-it edges determine the particular configuration. The counts are used as features in a supervised machine learning framework for training and rating prediction. The method is implemented and empirically evaluated on a large-scale Yahoo! music user-item rating dataset. Results show that the square counting method is fast, simple to parallelize, scalable to massive datasets and makes highly accurate predictions. Finally, we report an interesting empirical finding that configurations with consecutive I-hate-it edges seem to provide the most powerful signal in predicting a user’s love for an item. (Approved for public release by Northrop Grumman Information Systems, ISHQ-2011-0042. The work was entirely performed when Kyle Teague was at Northrop Grumman.)