Affiliations: [a]
College of Computer Science and Engineering, North Minzu University, Yinchuan, China
| [b] The Key Laboratory of Images and Graphics Intelligent Processing of State Ethnic Affairs Commission, North Minzu University, Yinchuan, China
| [c]
Agricultural Information Institute, Chinese Academy of Agricultural Sciences, Beijing, China
Correspondence:
[*]
Corresponding author. Wang Yingfei, College of Computer Science and Engineering, North Minzu University, Yinchuan, 750021, China. [email protected].
Abstract: To better understand users’ behavior patterns in web search, numerous click models are proposed to extract the implicit interaction feedback. Most existing click models are heavily based on the implicit information to model user behaviors, ignoring the impact of explicit information between queries and documents in search sessions. In this paper, we fully consider the topic relevance between queries and documents in search sessions and propose a novel topic relevance-aware click model (TRA-CM) for web search. TRA-CM consists of a relevance estimator and an examination predictor. The relevance estimator consists of a topic relevance predictor and a click context encoder. In the topic relevance predictor, we utilize the pre-trained BERT model to model the content information of queries and documents in search sessions. Meanwhile, we use transformer to encode users’ click behaviors in the click context encoder. We further apply a two-stage fusion strategy to obtain the final relevance scores. The examination predictor estimates the examination probability of each document. We further utilize learnable filters to attenuate log noise and obtain purer input features in both relevance estimator and examination predictor, and investigate different combination functions to integrate relevance scores and examination probabilities into click prediction. Extensive experiment results on two real-world session datasets prove that TRA-CM outperforms existing click models in both click prediction and relevance estimation tasks.
Keywords: BERT, click model, click prediction, deep learning, web search
