DMBIN: A Dual Multi-behavior Interest Network for Click-Through Rate Prediction via Contrastive Learning
Open Access
- 18 July 2023
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 1366-1375
Abstract
Click-through rate (CTR) prediction plays a critical role in various online applications, aiming to estimate the user's click probability. User interest modeling from various interactive behaviors(e.g., click, add-to-cart, order) is becoming a mainstream approach to CTR prediction. We argue that the various user behaviors contain two important intrinsic characteristics: 1) The discrepancy in various behaviors reveals different aspects of user's behavior-specific interests. For example, one may click out of need but pay more attention to the rating when purchasing. 2) The consistency of various behaviors contains user's behavior-invariant interest. For example, the user prefers interacted items rather than other items. Therefore, it is necessary to disentangle the discrepancy and consistency signals from the massive behavior information. Unfortunately, previous methods have yet to study this phenomenon well, which limits the recommendation performance. To tackle this challenge, we propose a novel Dual Multi-Behavior Interest Network (DMBIN for short) to disentangle behavior-specific and behavioral-invariant interests from various behaviors for a better recommendation. Specifically, DMBIN formalizethe discrepancy and consistency characteristics among various behaviors. Extensive experiments and empirical analysis on two real-world datasets demonstrate that DMBIN significantly outperforms the state-of-the-art methods. Moreover, DMBIN is also deployed in the online sponsored search advertising system in Meituan and achieves 2.11% and 2.76% improvement on CTR and CPM, respectively.s the dismantlement task as two contrastive learning tasks of multi-behavior interests extracted through the Multi-behavior Interest Module: Multi-behavior Interest Contrast(MIC) task and Multi-behavior Interest Alignment(MIA) task. These two tasks focus on extractingKeywords
This publication has 35 references indexed in Scilit:
- Real-time Attention Based Look-alike Model for Recommender System2019
- Lifelong Sequential Modeling with Personalized Memorization for User Response Prediction2019
- Neural Multi-task Recommendation from Multi-behavior Data2019
- A Simple Convolutional Generative Network for Next Item Recommendation2019
- Deep Interest Network for Click-Through Rate Prediction2018
- Squeeze-and-Excitation Networks2018
- Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding2018
- Wide & Deep Learning for Recommender Systems2016
- Deep Neural Networks for YouTube Recommendations2016
- FaceNet: A unified embedding for face recognition and clustering2015