Click-through rate prediction with multi-behavior sequences and shared interest learning

Click-through rate (CTR) prediction is a vital technique for assessing the effectiveness of online advertising. CTR prediction hinges on effectively modeling consumers’ product interests based on their historical behavior. However, extant CTR studies have not fully integrated consumer purchase history with other behavior histories such as browsing, adding to cart, and adding to favorites. How should we fully leverage the rich behavior history to extract personal interest and that of similar others? When extracting similar others’ interests, how should we maintain the relevance to personal interests while enjoying the novelty of others’ interests? To address these issues, we conceptualize two types of consumer interest: manifested interests and potentially shared interests when building their interest profile. We propose the Augmented Deep Multi-Behavior Interest Network (ADMBIN) to extract, integrate, and balance the two types of interest from the multi-behavior sequences. Experimental results demonstrate that the ADMBIN outperforms benchmark models in CTR prediction and the contribution of our design components. The proposed model can boost advertising revenue and improve consumer experience.