Bayesian Optimization for Parameter Tuning of Recommendation Systems

This research explores parameter optimization algorithms for recommender systems at retailers. Since evaluating the performance of a recommender is often expensive both in time and financial aspect, it is required that the optimization algorithm uses as little tests as possible. We tested several methods all in the field of Bayesian optimization: a Gaussian Process approach with a Matern and squared exponential covariance prior and a Tree Parzen Estimator approach with both a normal and uniform prior. The methods use an acquisition function, which determines the next parameter setting that will be evaluated. This research tested a new acquisition function (dynamic expected improvement) that is an adaptation to the popular expected improvement function. It was found that the TPE algorithm consistently finds better optimal values than the Gaussian Process approach. Moreover, the high impact of the prior selection is underlined by the results. It was also found that the dynamic expected improvement acquisition function performs worse than the regular expected improvement acquisition function, but could still have potential for applications with smooth objective functions.