Multi-Kernel-FM: A Multi-Embedding & Kernelization Factorization Machine Framework for CTR Prediction

Click-Through Rate (CTR) Prediction is one of the most critical components in recommender systems, where the task is to estimate the probability that a user clicks an item. In CTR models, embedding methods are widely used in feature representation to map categorical features into lower dimensional vectors, and thus those representations can be further exploited by various machine learning algorithms such as Factorization Machines (FMs) for CTR prediction. However, in the literature, most existing embedding models can only extract one latent vector for each individual feature as they calculate the feature interaction based on simple element product or inner product, limiting its ability to model user-item interactions in a high-dimensional space. It may miss some deep and complex interacted latent features, and therefore lead to a less proper representation as well as an inaccurate prediction. Motivated by the status quo, in this paper, we therefore propose a novel Multi-Kernel-FM (MKFM) framework for the task of CTR prediction. First of all, an embedding-based approach called Multi-FM (MFM) is proposed. It uses multiple embedding strategy and considers multiple representation sub-spaces for representing user-item features. After that, we construct a MKFM framework which combines kernel function and MFM to capture non-linear feature interactions. Then, the concept of kernel function is introduced and employed for capturing more high-dimensional feature interactions to further improve prediction accuracy. The results of our experiments on four public datasets demonstrate the superiorities of the proposed framework to some existing methods with respect to both prediction accuracy and training cost.