EulerNet: Adaptive Feature Interaction Learning via Euler's Formula for CTR Prediction

Zhen Tian,Ting Bai,Wayne Xin Zhao,Ji-Rong Wen,Zhao Cao

DOI: https://doi.org/10.1145/3539618.3591681

2023-09-12

Abstract:Learning effective high-order feature interactions is very crucial in the CTR prediction task. However, it is very time-consuming to calculate high-order feature interactions with massive features in online e-commerce platforms. Most existing methods manually design a maximal order and further filter out the useless interactions from them. Although they reduce the high computational costs caused by the exponential growth of high-order feature combinations, they still suffer from the degradation of model capability due to the suboptimal learning of the restricted feature orders. The solution to maintain the model capability and meanwhile keep it efficient is a technical challenge, which has not been adequately addressed. To address this issue, we propose an adaptive feature interaction learning model, named as EulerNet, in which the feature interactions are learned in a complex vector space by conducting space mapping according to Euler's formula. EulerNet converts the exponential powers of feature interactions into simple linear combinations of the modulus and phase of the complex features, making it possible to adaptively learn the high-order feature interactions in an efficient way. Furthermore, EulerNet incorporates the implicit and explicit feature interactions into a unified architecture, which achieves the mutual enhancement and largely boosts the model capabilities. Such a network can be fully learned from data, with no need of pre-designed form or order for feature interactions. Extensive experiments conducted on three public datasets have demonstrated the effectiveness and efficiency of our approach. Our code is available at: <a class="link-external link-https" href="https://github.com/RUCAIBox/EulerNet" rel="external noopener nofollow">this https URL</a>.

Information Retrieval

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the effective learning of high - order feature interactions in the click - through rate (CTR) prediction task. Specifically, existing methods are very time - consuming when calculating high - order feature interactions when dealing with large - scale features, and most methods reduce the computational cost by manually designing the maximum order and further filtering out useless interactions. Although these methods reduce the high computational cost caused by the exponential growth of high - order feature combinations, the model capacity is reduced due to the limited feature order. Therefore, how to improve computational efficiency while maintaining model capacity is a technical challenge, and this challenge has not been fully addressed.

To solve the above problems, the author proposes an adaptive feature interaction learning model - EulerNet. EulerNet learns feature interactions in the complex vector space by performing spatial mapping according to Euler's formula, thereby transforming the exponential operation of high - order feature interactions into a simple linear combination of modulus length and phase, making it possible to adaptively learn high - order feature interactions in an efficient manner. In addition, EulerNet also integrates explicit and implicit feature interactions into a unified architecture, achieving mutual enhancement and greatly enhancing the model capacity. This network can be fully learned from data without preset feature interaction forms or orders.