Recommender Systems in AI & Artificial Intelligence

Collaborative Filtering is the 'wisdom of the crowd'. It assumes that if User A and User B both liked the same five movies in the past, User A will likely enjoy the sixth movie that User B just rated highly. The most advanced form of this is Matrix Factorization (like SVD). By decomposing a sparse matrix of ratings into latent 'User' and 'Item' vectors, the model can predict the missing ratings in the table. This allows us to uncover hidden patterns, like a user's preference for 'Action movies with female leads' even if that category was never explicitly labeled.

Content-Based Filtering focuses on the items themselves. It uses metadata—like tags, descriptions, and keywords—to build a profile of what a user likes. If you frequently watch cooking videos, the system will recommend other cooking videos based on their textual and visual similarity. To calculate this, we use Cosine Similarity, which looks at the orientation of vectors in space. Unlike Euclidean distance, which can be biased by the 'length' of a profile (e.g., how many movies you've watched), Cosine Similarity only cares about the direction of your tastes.