AbstractRecommender systems are a viable solution against information overload found on daily used online services, rapid message exchanges and extended variety and complexity in data volumes. Recommender systems are widely employed to help individuals to swiftly iterate through multiple available options to find products or services that are likely to be of interest to them, e.g. news items, scholar articles or entertainment options and can be used to provide suggestions based on past user interactions, similar content selections, demographic data using artificial
intelligence or machine learning methods.
There are various methods for providing recommendations and the two most widely used are collaborative and content-based filtering that rely on the similarity of items. Collaborative filtering draws upon the ratings that a user has previously given
whereas content-based recommendations are formed according to knowledge from relevant context.
Hybrid recommendation methods or frameworks can use two or more recommendation algorithms to improve the quality of the provided recommendation. However, hybrid recommender systems in domains with high information overload, fuzziness and uncertainty are rare. This thesis proposes a novel framework to improve recommendations including the long-tail recommendation problem by applying a case-based reasoning approach based on user history. This method is extended with a multi-level algorithm that works as an add-on to existing collaborative filtering algorithms. A novel similarity measure for
item-based collaborative filtering has been developed by integrating the triangle similarity measure with a multi-level method, considering the length and the angle of rating vectors among users. Finally, a feature combination method is applied
which combines user-based collaborative filtering with attributes of demographic filtering.
The proposed framework has been thoroughly evaluated using well known datasets like MovieLens and Yahoo! Movies and metrics such as mean absolute error (MAE) and root mean squared error (RMSE). The results validate the framework by
showing that prediction accuracy is improved, while outperforming all the algorithms used as a baseline.
|Date of Award||Dec 2018|
|Supervisor||Miltiadis Petridis (Supervisor), Stelios Kapetanakis (Supervisor) & Nikolaos Polatidis (Supervisor)|