Online advertisements and search engines have become virtual libraries in today’s information-driven world, as almost everybody uses them to search for information in various fields. Given the fact that terabytes of information from various fields are collected over the Internet every day, it is important to develop highly efficient Web search models to provide information to Internet users.
Using a Bayesian network, Zhu et al. develop a general click model (GCM) and employ the expectation propagation method to perform approximate Bayesian inference. The model uses newly designed attributes, such as the local hour and the user agent, and traditional attributes, such as position and relevance. Unlike the existing models in the literature, the main goal of this model is to accurately model tail queries.
The paper provides necessary information from prior work in click models and formulates the GCM in detail, in order to make the paper self-contained in terms of mathematical notations and expressions.
The effectiveness of the model is illustrated through various experiments on advertising data and Web search data, and the performance results are compared with recent popular click models in the literature. Zhu et al. discuss in detail some of the drawbacks of the GCM, and then suggest ways to improve it in future work. This is an important contribution, in terms of generalizing the existing click models.