Abstract

In this thesis, various machine learning domains have been combined in order to build a video recommender system based on object detection. This combines two extensively studied research fields, computer vision and recommendation, that also are rapidly growing and popular techniques on commercial markets. To investigate the performance of the approach, three different content-based recommenders has been implemented at Spotify that are based on respective video data sources: object detections, titles and descriptions and user preference data. These are be evaluated individually and as a combined hybrid system.Evaluation of the system shows that the mean average performance scores are in range $sim$ 40-70% for the prediction algorithm, and in range $sim$ 15-70% for the top-$N$ algorithm. The reason why the system does not perform better can be explained by data sparsity. The approach based on object detection performs worse in comparison to the the other methods. Hence, there is a low correlation between the user tastes and the video contents in terms of object detection data, which therefore are bad descriptors of the video contents and not very suitable for the recommender system. However, the results can not easily be generalized over other systems before the approach has been tested on other environments, systems and for various data sets. Moreover, there are plenty of room for refinements and improvements to the system, as well as many interesting research areas for future work.