City Research Online

Abstract

Nowadays users have access to an immense number of media content. They are able to consume thousands of Television (TV) channels and millions of video clips from online portals like YouTube. Due to the immense number of available content, users can have the problem to find content of interest. This problem can be solved by recommendation systems. For example, recommendation systems can be used to create recommendations which fit to the preferences of users.

Recommendation systems can use two different approaches for the creation of recommendations. They can take content-based and/or collaborative-filtering techniques into account. Content-based filtering techniques use information, the so-called metadata, that describe the content in more detail. Collaborative-filtering techniques calculate similarities e.g., between users. All users are included in a dataset, the so-called community. Generally the number of user profiles within the community is quite large. Examples of such huge communities are Amazon, Netflix, MovieLens, and LastFM. The community which includes the user profiles is used to create a user-item matrix. This user-item matrix contains the preferences from users on items e.g., movies, genres, book titles, and so forth.

The quality of the recommendations depends on the accuracy of the predictions. As mentioned above, collaborative-filtering techniques calculate similarities e.g., between users. These similarities can be used to calculate predictions for an entry within the user-item matrix. If the predictions are close or equal to the preferences of a user, the used collaborative-filtering technique predicts accurately.

Generally recommendation systems only use one single collaborative-filtering algorithm for the similarity calculation. The research work of this thesis proves that a dynamic selection of the most accurate filtering algorithm by considering more algorithms is able to increase the accuracy of the predictions significantly.

In order to increase the accuracy of predictions, this thesis presents a dynamic multi-algorithm collaborative-filtering system which creates recommendations for video content, such as movies or genres. This system is able to find the most accurate filtering algorithm by considering the k-nearest neighbours. These neighbours are selected by identifying the most similar users or items e.g., movies. Besides the dynamic selection, this thesis presents newly developed collaborative-filtering algorithms which are able to overcome researched weaknesses of state-of-the-art algorithms.

The evaluation of the proposed system considers a huge dataset from MovieLens and a small dataset from an undertaken survey. The consideration of a huge and a small dataset shall prove that the system can be used in both cases.

The results of this thesis show that the proposed system is able to decrease the error rate significantly compared to existing approaches.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics
Departments:	School of Science & Technology > Engineering Doctoral Theses School of Science & Technology > School of Science & Technology Doctoral Theses

Export

Downloads