Mining User Profiles From Query Log

This paper introduces a novel method for mining user profiles (e.g., age, gender) using the query log in a search engine. The proposed method combines the advantage of the neural network for representation learning and that of the topic model for interpretability. This is achieved by plugging a parametric Gaussian mixture distribution layer into the neural network. Specifically, it first uses the popular convolution neural network to model the query content, generating a dense vector presentation for each query. Based on this representation, it infers the searching topic of the query, by fitting a Gaussian mixture distribution, and obtains the query topic distribution. Then, it deduces the distribution of topics that the user cares about by aggregating the query topic distribution of all the queries of the user. Profile prediction is performed based on the resulting user topic distribution. We evaluated this framework using a real search engine data set, which contains 40,000 labeled users with age, gender, and education level profiles. The experiment results demonstrated the effectiveness of our proposed model.