Speaker: Dr. Shengxin Zhu

Time: 11:15-12:15, 9 March 2022 (Wednesday) (Beijing time)

Venue: C208, Lijiao Building, BNU at Zhuhai


Abstract

In this talk, I shall give some reflections on feature interactions, generation and model acceleration for CTR prediction, based on our recent work xDeepFig: an eXtreme deep model with Feature Interactions and Generation. xDeepFig is a new deep model designed by four undergraduates at UIC. Numerical results on two benchmark datasets for CTR demonstrates such feature fusion can bring some advantages and the xDeepFIG outperforms recent baseline models by Huawei, Google and Microsoft.

We shall give some intuition on how to design such models and why they work from the perspective of kernel approximation theory.


Reference:

[1] Bokai Xu, Shihan Bu, Xinyue Li, Yanzhi Lin, and Shengxin Zhu. 2021. XDeepFIG: An eXtreme Deep Model with Feature Interactions and Generation for CTR Prediction. In 2021 3rd International Conference on Big-data Service and Intelligent Computation (BDSIC 2021). ACM, New York, NY, USA, 42–51. DOI: https://doi.org/10.1145/3502300.3502306