E-Commerce Item Recommendation Based on Field-aware Factorization Machine

The RecSys 2015 contest [1] seeks the best solution to a top-N e-commerce item recommendation problem. This paper describes the team Random Walker's approach to this challenge, which won the 3rd place in the contest. Our solution consists of the following components. Firstly, we cast the top-N recommendation task into a binary classification problem and extract original features from the raw data. Secondly, we learn derived features using field-aware factorization machines (FFM) and gradient boosting decision tree (GBDT). Lastly, we train 2 FFM models with different feature sets and combine them by a non-linear weighted blending. This solution is the result of numerous tests and the scheme turns out to be effective. Our final solution achieved a score of 61075.2, ranking in the third place on the public leaderboard.