Matrix factorization is a method for data fusion that can be used for recommender systems. In the thesis, we implemented a recommender system for recommending experiences to tourists. Tourism providers promote their experiences with images and descriptions. Our goal was to choose images and descriptions that best reflect a provider's rating. We used synthetic data for users and ratings and implemented a data generator for images and descriptions. We removed noise from the data using k-nearest neighbors algorithm and for each provider selected one image or one description that represents the provider's rating. The selected images and descriptions were used as a source for matrix factorization for predicting ratings. We compared our process to matrix factorization with no additional sources and to matrix factorization where we chose random images and descriptions and used them as a source. Our process with selected images and descriptions did not improve the AUC score compared to selecting random images and descriptions and using no additional sources. We also tested on online images and descriptions of real accommodations. We used k-nearest neighbors and random forests algorithms to predict average ratings for accommodations from images and descriptions. We concluded that with the methods we used, we were not able to extract patterns from data, which would enable the prediction of ratings based on images and descriptions.
|
