The Query Satisfaction Prediction Considering Time Sequence

Abstract

　In information seeking process, a user often submits multiple queries and repeats the search to achieve the purpose. Therefore, in order to discover the user's search intention more accurately, predicting the query level satisfaction from user’s search behavior is one of the important issues currently. In this research, we arrange the queries which submitted in a session by time sequence, and propose a method for predicting user’s query level satisfaction. The proposed method uses click data, interval time and query reformulation as indicators, which are often used to predict query satisfaction. In addition, we examine the effect of similarity between reformulated query and the search results of the previous query on query satisfaction. We predict query satisfaction by combining these four features. From the evaluation experiment, it shows that comparing with query reformulation only, the estimation effect has been improved 6 points approximately by using the similarity between the query and the document, but when using all of four features, the accuracy is 71.27% which was not the best result. In this paper, we discuss results with a failure analysis, and further improvement measures.