Machine Learning for On-line Advertising Using Contextual Information

Abstract

This thesis considers di erent methods of utilising the contextual information on webpages and ads in order to improve the tting of a Bayesian Poisson model to historic data using L-BFGS. The data and optimization algorithm is provided by Admeta, an advertising optimization company that uses the model for click-rate predictions. The di erent methods tried to get added contextual information include categorization and developing di erent similarity measures between web-pages and ads using keywords. The similarity measures are based on WordNet, a large lexical database, and Word2Vec an open source tool that represents words as vectors. The categorization of web-pages gives good results as does some of the similarity measures. As WordNet is limited to the words found in its databaseWord2Vec is deemed more exible and a superior source. For certain similarity measures it is shown that the click rate increases with the similarity. In the end using the average of the cosine distance between all keyword's vector pairs seams to give the best results among the di erent similarities tried for Word2Vec.