Input Enhanced Logarithmic Factorization Network for CTR Prediction

Abstract

Factorization-based methods, which can automatically model second-order or higher-order cross features, have been the benchmark models for click-through rate (CTR) prediction. In general, they enumerate all cross features with a predetermined order and filter out useless interactions through model training. However, two significant challenges remain. First, a maximum order for feature interactions needs to be defined in advance, imposing a trade-off between computational cost and the expression ability of models; Second, enumerating all feature interactions may introduce unwanted noise. In this work, we propose a novel model called Input Enhanced Logarithmic Factorization Network (ILFN), which can effectively learn arbitrary-order feature interactions and identify useful cross features. More importantly, ILFN can take full advantage of the original feature distribution in a discriminative way.

The core of ILFN is the Input Enhanced Component (IEC), which can represent the impact of each feature in a feature interaction as a trainable coefficient to learn arbitrary-order cross features. Moreover, IEC can reduce the demand for logarithmic neurons by exploiting the essential raw information and does not need to incorporate the deep neural network (DNN) to model high-order interactions. Therefore, ILFN is more effective and efficient and can converge to satisfactory results faster. Extensive experiments on four real-world datasets demonstrate that our ILFN model can outperform start-of-the-art methods. The effectiveness of each proposed component is also verified by hyper-parameter and ablation studies.