Comparison of Linear Learning methods in Click-Through Rate Prediction !
http://ieeexplore.ieee.org/abstract/document/7489611/?reload=true
#Abstract: A major challenge in the current era of search engine advertising is choosing which advertisements to show in response to a user query. This significantly impacts the overall user experience, and more importantly the advertising revenue stream for the search engine provider. Predicting click-through rates (CTR) for an advertisement is a massive-scale learning problem that is central to the multi-billion dollar online advertising industry. This study examines the performance of some well-known statistical learning methods (linear and logistic) with respect to their efficiency in predicting the click through rate of an impression, where an impression can simply be defined as an instance of a particular advertisement, with each instance defined in terms of the learning parameters in our data set. Our data set consisted of three types of independent attributes to act as a regressor in predicting our dependent variable - the app through which it was clicked, the site type and the domain to which it led - with the help of other anonymised variables. Fine tuning of the algorithm parameters was done to get promising results. Besides that a dimensionality check on the data set was conducted to observe the possibilities of dimensionality reduction. Logistic loss (log-loss) was used as the validation index in all cases. Our observations led us to the conclusion that with minimal data preprocessing, linear models give competitive on-par results suited for most practical applications, where the learning method chosen should not be computationally expensive. We go on to further verify this claim by comparing the performance of linear models on various subsets of the data set attributes, showing that the performance of the linear techniques was consistent all across.
#Published in: Soft Computing Techniques and Implementations (ICSCTI), 2015 International Conference