The use of Self-Organizing Maps in Recommender Systems 2006-08-22
Abstract
This is a thesis about recommender systems, and the different approaches recommender systems use to solve the information overload problem. Our focus lies on two common approaches, collaborative filtering and content based filtering. Both of these approaches have their weaknesses and strengths. To overcome the weaknesses of each approach, various hybrid filters have been developed. We will start by analyzing these three approaches based on previous research literature and will then proceed to implement different variants of these approaches, including our own filtering approach for the movie domain. These implementations will be done in Java and open sourced for further development by other researchers in this area. The results will be evaluated and compared against previous research in this area in order to validate our implementations. Evaluation will be done by using standard metrics that are commonly used for evaluating the accuracy of recommender systems.
Various algorithms from the machine learning community have been used in the effort to improve and solve some of the problems in the previously mentioned approaches. We will concentrate on one such algorithm, Kohonen's self-organizing map algorithm. The self-organizing map algorithm is an unsupervised learning algorithm which we believe is suitable for recommender systems in the movie domain. Our implementation of this algorithm will be used together with collaborative filtering approaches in the effort of designing a recommender system for movies, called MOVSOM. The result of this approach will be evaluated and compared against the results we got from our previous implementations and discussed in the context of previous results from the recommender systems research community.
Evaluating the effectiveness of recommender systems is often done by analyzing the accuracy of the recommendations produced by the techniques used to implement the different approaches. However, the goal for a recommender system is not only to give accurate recommendations but also to conceive to the user trust and encourage the user to explore the recommendations. This is more of an interface issue than an algorithmic issue, we have chosen to call this the recommendation interface problem. Similar conclusions have been drawn by other researchers and different attempts to solve this has been done. We will summarize and discuss proposed solutions. We will introduce and describe what we call visual recommendations, and show how this approach solves the recommendation interface problem by creating a visual recommender system called MOVSOM.
The testing and evaluation will be done on the well used MovieLens dataset, as well on a larger dataset taken from an e-commerce site selling DVDs, together with movie attributes provided by the IMDb.
Our empirical evaluation results shows that MOVSOM produces recommendations of movies that are comparable to state of the art techniques and with the combination of our solution to the recommendation interface problem we believe that this approach has a very promising future as a recommender system for movies.
@unpublished{gabrielsson06somrs,
author={Sam Gabrielsson and Stefan Gabrielsson},
title={The use of {S}elf-{O}rganizing {M}aps in {R}ecommender {S}ystems,
note={Uppsala University, Department of Information Technology},
year={2006}
}
Paper SOMs
SOMs discussed in the paper.
| Dataset | SOM | |||
| CML | Item | 20 x 20 | U-Matrix | Hitmap |
| User | 20 x 20 | U-Matrix | Hitmap | |
| DS | Item | 20 x 20 | U-Matrix | Hitmap |
| User | 20 x 20 | U-Matrix | Hitmap | |
| DS2 | Item | 20 x 20 | U-Matrix | Hitmap |
| User | 20 x 20 | U-Matrix | Hitmap | |
| IMDBK | Item | 60 x 60 | U-Matrix | Hitmap |