Recommender systems were designed as a software solution to the problem of information overload. Recommendations can be generated based on the content descriptions of past purchases (Content-based), the personal ratings an individual has assigned to a set of items (Collaborative) or from a combination of both (Hybrid). There are issues that affect the performance of recommender systems, in terms of accuracy and coverage, such as data sparsity and dealing with new users and items. This thesis presents a comprehensive set of offline experiments and empirical results with the goal of improving the recommendation accuracy and coverage for the poorest performers in the dataset. This research suggests approaches for dealing with four specific research challenges: the standardisation of evaluation methods and metrics, the definition and identification of sparse users and items, improving the accuracy of hybrid systems targeted specifically at the poor performers and addressing the cold-start problem for new users. A selection of recommendation algorithms were implemented and/or extended, namely, user-based collaborative filtering, item-based collaborative filtering, collaboration-via-content and two hybrid prediction algorithms. The first two methods were developed with the express intention of providing a baseline for improvement, facilitating the identification of poor performers and analysing the factors which influenced the performance of recommendation algorithms. The later algorithms were targeted at the poor performers and were also examined with respect to user and item sparsity. The collaboration-via-content algorithm, when extended with a new content attribute, resulted in an improvement for new users. The hybrid prediction algorithms, which combined user-based and item-based approaches in such a way as to include information about transitive relationships, were able to improve upon the baseline accuracy and coverage results. In particular, the final hybrid algorithm saw a 3.5% improvement in accuracy for the poor performers compared to item-based collaborative filtering.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:535143 |
| Date | January 2010 |
| Creators | Redpath, Jennifer Louise |
| Publisher | Ulster University |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
Page generated in 0.0013 seconds