Abstract
Modeling feature interactions is of crucial importance to predict click-through rate (CTR) in industrial recommender systems. Because of great performance and efficiency, the factorization machine (FM) has been a popular approach to learn feature interaction. Recently, several variants of FM are proposed to improve its performance, and they have proven the field information to play an important role. However, feature-length in a field is usually small; we observe that when there are multiple nonzero features within a field, the interaction between fields is not enough to represent the feature interaction between different fields due to the problem of short feature-length. In this work, we propose a novel neural CTR model named DeepFIM by introducing Field-aware Interaction Machine (FIM), which provides a layered structure form to describe intrafield and interfield feature interaction, to solve the short-expression problem caused by the short feature-length in the field. Experiments show that our model achieves comparable and even materially better results than the state-of-the-art methods.
Highlights
Among plenty of prediction algorithms, the factorization machine (FM) [1] is one of the most popular ones due to its excellent performance, and many variants have been derived. e key of FM is the vectorization of features and the use of an inner product of two vectors to model the effect of pairwise feature interactions
FATDeepFFM [6] adopts the field expression of features and estimates the importance of different feature interactions between fields; they argue that it is important to estimate the importance of features before they are crossed, but the existing models [7,8,9,10] are after the feature is crossed
The above models have achieved promising results, the challenge in real-world online advertising or recommender systems is that there may be multiple dense nonzero features in a field of an instance sample. In a model such as FAT-DeepFFM [6], each column vector of the field corresponds to the influence of other fields, and one column corresponds to itself. en, the expression ability of feature interaction in the field may be limited by short featurelength. is means that the short featurelength makes it easy to produce “oversimilarity” between feature interactions that should not be so similar, reducing model performance
Summary
Among plenty of prediction algorithms, the factorization machine (FM) [1] is one of the most popular ones due to its excellent performance, and many variants have been derived. e key of FM is the vectorization of features and the use of an inner product of two vectors to model the effect of pairwise feature interactions. AUFM [12] and FNFM [13] propose to combine FFM with SA (stacked autoencoder) and DNNs (deep neural networks) to enhance the ability of high-order feature expression. The above models have achieved promising results, the challenge in real-world online advertising or recommender systems is that there may be multiple dense nonzero features in a field of an instance sample. We propose a new feature interaction expression based on field identifier, which is called “hierarchical-structure expression.” On this basis, we design a cross-interaction layer to identify the intrafield and interfield interaction and use an attention mechanism to distinguish the importance of different features. (i) We propose a new model named DeepFIM to exploit the field information of features and better learn the interaction by addressing the problem of short-expression
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
