Towards better ad experience: Click prediction leveraging sequential networks derived specifically from user search behaviors

We propose a sequential modeling approach to improve click prediction for search engine advertising. Unlike previous studies leveraging advertisement content and their relevance-to-query information, we employ only users' search behavioral features such as users' query texts and actual click records of both organic search results and advertisements. By leveraging long short-term memory (LSTM) networks, we successfully modeled users' sequential search behaviors and fully utilized them in click predictions. Through experiments conducted with large-scale search log data obtained from an actual commercial search engine, we demonstrated that our method combining users' current and previous search behaviors reaches better prediction performance than baseline methods.