Dynamic Query Intent Prediction from a Search Log Stream

Problem Definition

This problem is defined formally as follows. Let s = (a₁, a₂, a₃… a_n) be a query log stream of a user (search stream of a single user as a document) [Hsin et al. 2011]. [Di et al.2014]. But the user query intent change dynamically, so let’s divide the query log steam to stream bins (b₁, b₂… b_n). Query log stream bins are interactions between user and the search engine in specific times or depend on number of queries like in our experiment for simplicity.

Each stream bin contains several actions where each action a_i (1 <= i <= n) is either a query submitted by the user or a URL clicked by the user The actions a₁, a₂, …, a_n are ordered by the time of their occurrences, with a₁ having the earliest occurrence time [Hsin et al., 2011].

Stream bins in query stream can be divided into different pairs (H₁, F₁), (H₂, F₂)… (H_n-1, F_n-1) where H_j and F_j are two action sequences [Hsin, et al., 2011]. For each stream bin, it is possible to view H_j as a history of the actions that a user has performed during time t_n-1 to t_n, and F_j as the future action that the user will perform during the time t_n to t_n+1. The goal is to dynamically predict F_j through the stream while given that H_j is known.