Retrieval-Based Factorization Machines for CTR Prediction

Click-through rate (CTR) prediction is a crucial task for personalized services such as online advertising and recommender system. Many methods including Factorization Machines (FM) and complex deep neural models have been proposed to predict CTR and achieve good results. However, they usually optimize the parameters through a global objective function such as minimizing logloss and mean square error for all training samples. Obviously they intend to capture global knowledge of user click behavior, but ignore local information. Therefore, we propose a novel approach of Retrieval-based Factorization Machines (RFM) for CTR prediction, which enhances FM by the neighbor-based local information. During online testing, we also leverage the K-Means clustering technique to partition the large training set to multiple small regions for efficient retrieval of neighbors. We evaluate our RFM model on three public datasets. The experimental results show that RFM performs better than existing models including FM and deep neural models, and is efficient because of the small number of model parameters.