Efficient Estimation of Events with Rare Occurrence Rates Using Taxonomies

摘要

Methods for predicting the click-through rates of Internet advertisements placed into web pages are disclosed. Specifically, a click-through rate prediction is generating using a hybrid system with two terms. The first term is constructed using a machine learning model that incorporates a limited number of important factors. The second term is constructed using a look-up table that is built using a complex statistical analysis of various web page and advertisement combinations. To construct the second term, the field of multi-level hierarchical modeling is used. Specifically, a tree-structured Markov model is used to process the training data and construct the adjustment factor look-up table. To reduce the complexity of the statistical analysis, Kalman-filters are used to estimate parameters in the traditional multi-level hierarchical models for scalability.