Time-aware Attentive Click Sequence Network for Click-Through Rate Prediction
Shoujian Yu, Changhui Yang, Zhenchi Jie, Xiujin Shi · 2022
User's behavior sequence contains user's interest, which plays an important role in predicting the user's future behavior. The sequence of user actions is usually represented as a collection of items that users have interacted with sorted in chronological order and trained with an RNN-based model, but such a representation ignores a large amount of important temporal information. Second, direct use of user's behavior sequences has limitations in predicting users' uninteracted items. By studying the Transformer sequence model, the paper utilizes relative time information inside the sequence and designs the relative time-aware Transformer module to model the effect of time intervals on the relationship between behaviors, thus improving the generalized expression of behavioral sequences. Experiments show that our proposed model has better improvement in all evaluation metrics compared with other click-through rate (CTR) models.