Two-Stage Approach to Item Recommendation from User Sessions

Abstract

We present our solution to the 2015 RecSys Challenge [1]. This challenge was based on a large scale dataset of over 9.2 million user-item click sessions from an online e-commerce retailer. The goal was to use this data to predict which items (if any) were bought in the 2.3 million test sessions. Our solution to this problem was two-staged, we first predicted if a given session contained a buy event and then predicted which items were bought. Both stages were fully automated and used classifiers trained on large sets of extracted features. The prediction rules were further optimized to the target objective using a greedy procedure developed specifically for this problem. Our best submission, which was a blend of several different models, achieved a score of 60,265 and placed 4'th out of 567 teams. All approaches presented in this work are general and can be applied to any problem of this type.