One-class vs multi-class CF

Collaborative filtering (CF)

Collaborative filtering is a recommendation technique that predicts the interest of a user in items from the preferences of many other users with similar tastes. It uses a user-item matrix of ratings or actions. Two main styles are user-based (find similar users) and item-based (find similar items), using similarity measures such as cosine or Pearson correlation.

Example: if Ram and Sita both liked films A and B, and Sita also liked C, then C is recommended to Ram.

Multi-class (general) collaborative filtering

The feedback is explicit and has several levels or classes, for example 1 to 5 stars. Both likes and dislikes are known, and the level of preference is known.

• Example: movie ratings on Netflix or IMDb. Ram rates film A = 5, B = 2. The system predicts his rating for film C as 4.3 and recommends high-scoring films.
• Missing entries mean "not rated" and can be ignored in training.
• Methods: neighbourhood models, matrix factorization, regression on ratings.

One-class collaborative filtering (OCCF)

Only positive feedback is available: a purchase, click, view, or like. There are no negative ratings; a missing entry may mean dislike or just unawareness.

• Example: an online shop where we only know what a customer bought; a news site where we only know clicked articles.
• The challenge is that positive examples are few, and missing values are ambiguous.
• Methods: treat all missing as weak negatives (with low weights), sample negatives, or give higher confidence to repeated actions (weighted matrix factorization, Bayesian personalised ranking).