Reccomendation engine tuning for cross-media catalogs

What kind of music do you listen too? What movie genres do you tend to prefer? Comedies, blockbusters, 90′s SciFi? Are you a fiction or non-fiction reader? Do you enjoy business literature? What is your favorite sports to watch on TV?

People usually have a multitude of interests. When it comes to media and digital content, our interests and preferences can span any number of different heterogeneous genres, moods and styles, not only within any individual media domain, but also certainly across domains. But can we successfully establish a bridge between a user’s preferences in one domain (e.g., movies) and her preferences in another domain (e.g., video games)? To put it simply, does the fact that I like a certain horror movie mean that I am likely to prefer a certain strategy video game?

These questions are certainly attracting the attention of the recommender system scientific community. In the last two years, for instance, the major conference in the field – ACM Conference on Recommender Systems – has hosted a special workshop on “Information Heterogeneity and Fusion“. Similarly, the IEEE International Conference on Data Mining (ICDM) has hosted a workshop on “Mining Multiple Information Sources“.

As summarized in a short survey at the “23rd IEEE International Conference on Tools with Artificial Intelligence“, several works have concluded that transferring knowledge across different domains can effectively increase the performance of recommendation engines based on collaborative filtering. The problem is not trivial at all and user-based approaches founded on assumptions such as “if a group of users have similar tastes about movies they will have similar tastes about video games” are likely to fail. Some advanced solutions – still limited to academic applications – have been proposed. They try to take advantage of hidden relationships among:

• user ratings. Li et al. proposed, for instance, the so called CodeBook in order to share rating patterns across separated domains. (See “Can movies and books collaborate? cross-domain collaborative filtering for sparsity reduction“.
• items. For instance, Cremonesi et al. recently presented a solution for discovering relationships among items belonging to different domains. (See “Cross-domain recommender systems“).

Here’s a simple example.

A recommendation engine that applies a cross-domain collaborative technique might discover that those who like action movies in a certain demographic group are likely to give higher than average ratings to strategy video games. Now, if in the recommendation engine estimation, I have a preference for action movies, I will be recommended more strategy games, and more often, than the average user.  Advanced cross-domain algorithms are also able to discover hidden patterns among multiple domains. For instance, if players who like strategy video games also like role-playing video games, with such techniques I might be recommended both strategy and role-playing video games, just because the recommender system is aware I like action movies.