Reproduce “xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems” on OpenResearch

About this paper

Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, Guangzhong Sun

The xDeepFM model introduces a Compressed Interaction Network (CIN) to explicitly learn vector-wise high-order feature interactions, combining this with implicit bit-wise interactions from a deep neural network and a linear component. This integrated architecture consistently outperforms existing state-of-the-art recommender systems in AUC and Logloss on multiple real-world datasets.

Key insights

• Explicitly modeling high-order feature interactions in a vector-wise manner (via CIN) proves more effective than implicit bit-wise approaches or constrained explicit designs like CrossNet.
• The combined architecture of xDeepFM, leveraging both explicit (CIN) and implicit (DNN) high-order interactions alongside a linear component, comprehensively captures feature relationships.
• Using an identity activation function within CIN layers yields optimal performance, implying that linear combinations effectively preserve the nature of explicit, vector-wise interactions.

Summary by alphaXiv.