Time-aware Multi-layer Interest Extraction Network for Click-Through Rate Prediction

Abstract

Click-through rate (CTR) prediction, which is used to estimate the probability of a user clicking on a candidate item, acts as a core task in recommender system. Previous researchers model user's historical behaviors as a sequence and apply sequential models to extract user interests. However, user behaviors result from multiple factors, including not only their interests but also the time, especially in time-sensitive scenes. While a few researchers have considered behavior time in sequential modeling, the target predicted time is still ignored. In this paper, we propose a novel network for CTR prediction dubbed Time-aware Multi-layer Interest Extraction network (TMIE), which considers the influence imposed by user behavior time and the target predicted time along-side with modeling user interests. Specifically, we design and employ time-aware GRU as low-layer interest extractor to capture primary interests. Then simplified transformer is applied as high-layer extractor to further explore the mutual relevance among user's interests. We perform abundant comparative experiments on both public and industrial datasets and the excellent results demonstrate the rationality and effectiveness of our methods. Notably, our heuristic work is an exciting attempt to catch up the synergistic impact of behavior time and multi-layer user interests in CTR prediction.