Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.745787
Title: Pairwise preferences learning for recommender systems
Author: Qomariyah, Nunung Nurul
ISNI:       0000 0004 7227 5791
Awarding Body: University of York
Current Institution: University of York
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
Preference learning (PL) plays an important role in machine learning research and practice. PL works with an ordinal dataset, used frequently in areas such as behavioural science, medical science, education, psychology and social science. The aim of PL is to predict the preference for a new set of items based on the training data. In the application area of Recommender Systems (RSs), PL is used as an important element to produce good recommendations. Many ideas have been developed to build better recommendation techniques. One of the challenges in RSs is how to develop systems that are proactive and unobtrusive. To address this problem, we have studied the use of pairwise comparisons in preference elicitation as a very simple way of expressing preferences. Research in PL has also discovered this kind of representation and considers it to be learning from binary relations. There are three contributions in this thesis: The first and the most significant contribution is a new approach based on Inductive Logic Programming (ILP) in Description Logics (DL) representation to learn the relation of order. The second contribution is a strategy based on Active Learning (AL) to support the inference process and make choices more informative for learning purposes. A third contribution is a recommender system algorithm based on the ILP in DL approach, implemented in a real-world recommender system with a large used-car dataset. The proposed approach has been evaluated by using both offline and online experiments. The offline experiments were performed using two publicly available preference datasets, while the online experiment was conducted using 24 participants to evaluate the system. In the offline experiments, the overall accuracy of our proposed approach outperformed the other 3 baseline algorithms, SVM, Decision Tree and Aleph. In the online experiment, the user study also showed some satisfactory results in which our proposed pairwise comparisons interface in a recommender system beat a common standard list interface.
Supervisor: Kazakov, Dimitar Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.745787  DOI: Not available
Share: