Weighted Ensemble of Neural and Probabilistic Graphical Models for Click Prediction

Abstract

Predicting user behavior in web mining is an important concept with commercial implications. The user response to search engine results is crucial for understanding the relative popularity of websites and market trends. The most popular way of understanding user interests is via click models that can predict whether a user will click on a search engine result or not, based on past observations. There are two main categories of click models, namely, the neural network based models and the probabilistic graphical models. In this paper, we combine the goodness of both approaches by presenting a weighted ensemble of both types of models. The weighted sum of softmax scores integrates the predictions of the individual models. Assigning higher weights to the neural models is found to improve the performance of the ensemble. The AUC and perplexity scores of our weighted ensemble model are higher than the state of the art, as proved by experiments on the benchmark Tiangong-ST dataset.