Manipulace kapacitou doporučovacích modelů pro optimalizaci v Recall-Coverage rovině

Abstract

Traditional approaches in Recommender Systems ignore the problem of long-tail recommendations. There is no systematic approach to control the magnitude of long-tail recommendations generated by the models, and there is not even proper methodology to evaluate the quality of long-tail recommendations. This thesis addresses the long-tail recommendation problem from both the algorithmic and evaluation perspective. We proposed controlling the magnitude of long-tail recommendations generated by models through the manipulation with capacity hyperparameters of learning algorithms, and we dene such hyperparameters for multiple state-of-the-art algorithms. We also summarize multiple such algorithms under the common framework of the score function, which allows us to apply popularity-based regularization to all of them. We propose searching for Pareto-optimal states in the Recall-Coverage plane as the right way to search for long-tail, high-accuracy models. On the set of exhaustive experiments, we empirically demonstrate the corectness of our theory on a mixture of public and industrial datasets for 5 dierent algorithms and their dierent versions.

Traditional approaches in Recommender Systems ignore the problem of long-tail recommendations. There is no systematic approach to control the magnitude of long-tail recommendations generated by the models, and there is not even proper methodology to evaluate the quality of long-tail recommendations. This thesis addresses the long-tail recommendation problem from both the algorithmic and evaluation perspective. We proposed controlling the magnitude of long-tail recommendations generated by models through the manipulation with capacity hyperparameters of learning algorithms, and we dene such hyperparameters for multiple state-of-the-art algorithms. We also summarize multiple such algorithms under the common framework of the score function, which allows us to apply popularity-based regularization to all of them. We propose searching for Pareto-optimal states in the Recall-Coverage plane as the right way to search for long-tail, high-accuracy models. On the set of exhaustive experiments, we empirically demonstrate the corectness of our theory on a mixture of public and industrial datasets for 5 dierent algorithms and their dierent versions.