What is Content Filtering?
Content-based (also referred to as case-based or search-based) filtering (CBF) originated in the information retrieval community, with search engines being the most common application. These recommender systems rely on the availability of features that describe the items in the dataset. The features are then exploited, and often weighted, either globally or per user, to discriminate between significant attributes. User profiles are created which describe a user’s interests using the descriptive (optionally weighted) features of items previously purchased or rated. The user profiles are then matched to items which contain similar features. The two tables below show sample content cases: one provides the representation of a movie that could be recommended when applying CBF and the second the profile of a user that may like the movie.
Example Movie Profile
Ali MacGraw, Ryan O’Neal
Example User Profile
Lorenzo's Oil, The Best of Youth
Marco Tullio Giordana
Evan Rachel Wood, Nick Nolte
Drama, Romance, Thriller
Different methods are employed to compare user profiles and item descriptions, in order to identify possible recommendations that may be of interest to the user. For example, a movie recommender system may suggest films by the same director or genre. Unfortunately, the important features of an item are subjective, and different users prefer items for different reasons. In addition, the only factor influencing recommendations is the small set of movies rated or purchased by the individual user.
Some relevant research papers
- NewsWeeder: Learning to Filter Netnews
NewsWeeder is a netnews-filtering system that addresses the problem of user profile creation by letting the user rate his or her interest level for each article being read (1-5), and then learning a user profile based on these ratings.
- Using Content-Based Filtering for Recommendation
In this paper the recommender system PRES is described that uses content-based filtering techniques to suggest small articles about home improvements.
Advantages of Content-Based Filtering
CBF approaches have the advantage of being able to uniquely characterise each user and are less affected by the first rated, cold-start and grey sheep problems. This is because the techniques employed are applicable across the entire data space, meaning that an item can be recommended before anyone has rated it, a user can receive recommendations as soon as they show interest in a single item, and that the recommendations for a user are not reliant on other users having similar tastes.
On my book shelf
Disadvantages of Content-Based Filtering
The ability of CBF systems to perform effectively is based on the assumption that the item descriptions and suitable similarity metrics are available. This is sometimes not the case and even when descriptions are available, the expense of coding them may be too much. Furthermore, the algorithms must make decisions about subjective attributes and attempt to distinguish between high and low quality ones; this can differ between users, as one user may be more interested in the genre of movies whereas another may only be interested in movies that include a certain actor. In addition, CBF algorithms suffer from overspecialisation and a lack of diversity in the recommendation set. As recommendations are based solely on the items previously rated by the user, they all tend to be very similar. Methods presented in the literature for tackling this issue include explicitly introducing diversity to the generated recommendation set and building hybrid algorithms which incorporate some collaborative aspects of recommendation.