User Click Preference Prediction Using Attention-based LSTM Model

Abstract

Predicting user clicking preferences is a critical issue in the ranking process reflected by the returned results from search engines. A few studies have explored the extracted sequen- tial information in the queries while this underlying information can be focused on as a valuable source for predicting user click preferences. This paper proposes an attention-based LSTM model to predict user click preferences on a submitted input quey using the previous queries sequence and user click history. In this model, we use the topic distribution of user documents as attention to the LSTM model. The feature information of the content extracted is used as the attention information of the LSTM network during the training process. We compare the model with topic-based ranking models with data from an AOL search engine and Session TREC 2013,2014 to show its performance. The result reveals significant improvement in the attention-based LSTM model using topics in the Mean Reciprocal Rank by 13% compared to the baseline topic-based models.