Multi-Behavior Euler Transformer for Click-Through Rate Prediction

Click-Through Rate (CTR) prediction is crucial for recommender systems, including e-commerce platforms. It estimates the likelihood of users clicking on recommended items, enabling more accurate and effective recommendations. Modeling user interests through behavior sequences has become the mainstream of CTR prediction task. In order to capture user interest preferences, models of the Transformer architecture have been widely used to model behavior data, and previous work has modeled semantic and positional differences in behavior data in different ways, inhibiting the expressive power of sequence modeling. User behaviors often show diversity, making it necessary to fuse multiple user behavioral data to provide more accurate recommendations. To address this challenge, we propose the Multi-Behavior Euler Transformer-based CTR prediction model (MBET). This model enhances multi-behavior interest modeling and adaptive fusion for better recommendation. Specifically: we introduce multi-behavior sequence data, enhance the representation of long-tailed distributed items using a contrast learning task. Utilize the Euler Transformer to transform representations into a complex vector space, unifying semantic and positional differences, and finally, design a multi-behavior adaptive fusion mechanism to dynamically adjust behavior weights based on different needs. Extensive experiments verify that MBET achieves the best performance on two real-world datasets.