| dc.description.abstract | Recommender systems have cemented themselves in the daily online activities of most people, and they have been successfully applied across a range of different domains. However, they have yet to make the big breakthrough in complex areas such as tourism and gastronomy. A reason for this is that the attributes of items within these domains are seldom readily quantifiable, and people s opinions on items like hotels and restaurants are dependent on a large number of factors. A successful recommender system in such a complex domain could have a great impact on further advancements of this technology.
In this thesis, we present RestRec: a novel, personalized, context-aware, hybrid recommender system for restaurants. We perform a literature review showing that even though contextual information has huge potential, it has been largely ignored in research. In order to learn more about what factors affect people's choices in restaurants, and the nature of their social settings when attending them, we conduct a survey. The subsequent results are then incorporated into our system with the purpose of improving recommendations. Furthermore, we address the cold-start user problem which pertains to the difficulty of providing high quality recommendations to new users where little information exists. This problem is particularly dominant in the restaurant domain where the majority of users are cold-start users. To address the effects of a cold start, we employ a combination of collaborative filtering with demographic information and content-based filtering.
To evaluate RestRec we perform user-based evaluation conducted on friends, family, and fellow students. Our experiments show that we are indeed successful in identifying patterns with regards to social setting and use this to make better recommendations. The overall predictive accuracy of the system exceeds 70 %, showing the feasibility of our approach. | |
