Learning Query Intent for Sponsored Search

As sponsored search contributes the major income of many search engines, how to deliver ads effectively is studied intensively from both industrial and academia. Among various previous studies, many of them are keyword relevance based and few considered the underlying user intents behind queries for ads delivery. In this paper, we propose to classify search queries into different search intent categories such as “car buyer”, “car reviewer” and “car maintainer” about cars for ads delivery. However many classical machine learning models for query classification may fail in categorizing queries into the predefined intent categories since it is hard to collect data for training and in addition, the traditional bag of words features cannot accurately reflect user intents. In this work, we propose a random walk based solution on top of Web click-through graph, which optimize the features for measuring user intent during random walk as well. Through this way, we can automatically generate large scale training data, weight features according to query intent, and classify queries into intent categories simultaneously. The experimental results on real world user search click-through data show that our approach can improve ad click-through rate (CTR) as high as 62.5% in contrast to the relevance based ads delivery solutions without taking user intent into consideration.