Download files
Abstract
Many recommender systems rely heavily on user ratings and historical data to produce recommendations. When lacking such data, these systems do not perform well. Knowledge-based recommender systems on the other hand are ideal for naive and casual users. However, the challenge for these systems lies in the knowledge engineering process to produce the knowledge base.
This thesis proposes a novel knowledge acquisition method to produce a knowledge base for a critiquing-based recommender system. A critiquing component allows users to critique the recommendations, and explore the qualitative search space towards their target products. A knowledge acquisition tool is used by a (human or simulated) domain expert to produce ranking functions for each qualitative attribute, based on the concrete features of the items. These functions are used to rank the product set based on these qualitative attributes.
To test this approach, two knowledge bases were built in domains of movies and cars. Then we built qualitative critiquing recommender systems using these knowledge bases and conducted user studies to evaluate the systems. The first user study examined whether users could use a car recommender system to find a set cars satisfying the typical high-level needs of a naive user searching for new cars. The study showed that the users were able to fulfil the tasks. Based on the questionnaires and systems logs, we concluded that these users were satisfied with their choices.
The second user study, a comparative user study, compared our movie recommender system with a similarity-based recommender which used concrete attributes of the movies to produce recommendations. The results of this study showed that the performance of the qualitative knowledge-based recommender was superior to the similarity-based recommender, and that users could reach better results using our system.