Recommender systems--where products, services, and even people can be recommended based on some behind-the-scenes computation--are quite well established as an interesting and sometimes complex area of computer science. Increasingly, Internet aficionados are embracing such technologies as naturally useful and intelligent tools to help them cope with large volumes of music, movies, and online social networks of friends. However, each recommender system has historically been based on single computing algorithms that can offer good recommendations, but usually have attendant-particular weaknesses.
This paper provides a brief overview of the most common types of recommender algorithms, and usefully lists the main weaknesses of each. For example, the “cold start” problem arises in recommender systems that use the collaborative filtering technique. This means that a user has to interact with the recommender and rate a number of action movies, for example, before the system can glean enough information to be able to recommend a movie.
The approach taken here is a common solution, where different recommender techniques are bundled together to produce a hybrid recommender system. In this case, the approach is to compare and combine the content-based user profiles in a collaborative way, where “content-based profiles are built to detect similarities between users.” The profiles are constructed into an ontology, enabling the resulting recommendations to avail themselves of the inferencing capabilities of such constructs. Unfortunately, this aspect of the work is not presented in any detail here.
Does it work? Well, results using the de facto MovieLens reference data show that the authors’ algorithms yield fewer errors in predicting relevant items than pure content-based techniques. Interestingly, the approach taken in the hybrid recommender does address problems such as “cold start” by using the ontological constructs.