Ad optimisation with a collaborative multi-armed bandit algorithm

Sammanfattning

This report describes the implementation of a collaborative multi-armed bandit algorithm in Python. The algorithm is tested on different online ad campaign data with the purpose of finding out if a collaborative bandit can come to an accurate decision faster on which ad in a set of ads is the best performer with regards to click-through rate. If the best performing ad can be found quickly, that ad can be displayed the most, thus maximizing profits for the advertiser. Analysis is also done on the ad data to see if similar ads perform similarly, which is important for the idea to work. The experiments performed show a decrease in performance when using collaboration between ads compared to no collaboration over a longer time period, however a slight increase in the short term. Some issues with the work and possible future extensions are also discussed.