itembased

Itembased, in recommender systems, refers to item-based collaborative filtering, a memory-based approach that predicts a user's interest in an item from its similarity to items the user has engaged with. It complements user-based methods and is common in e-commerce and media platforms for its efficiency and interpretability.

Construction: It operates on a user-item interaction matrix. Item-item similarities are computed from this data using metrics such as cosine similarity, Pearson correlation, or adjusted cosine to offset user bias. The result is an item similarity matrix S, where S(i, j) measures relatedness between items.

Prediction: For user u and item i, a typical score is r_hat(u,i) = sum_j S(i, j) * r(u,j) divided by sum_j |S(i, j)|, over items j that user u has rated or interacted with. Implicit-feedback variants use confidence weights instead of explicit ratings.

Practical aspects: Similarities can be precomputed offline, enabling fast online recommendations. The method scales well with proper sparsity handling, but remains sensitive to data sparsity and the cold-start problem for new items.

Evaluation and limitations: It is often evaluated with RMSE/MAE for rating prediction and precision/recall or NDCG for ranking. Advantages include transparency and interpretability; limitations include bias toward popular items and reduced accuracy when few co-rated items exist.

History and usage: Item-based collaborative filtering gained prominence in the early 2000s, notably in work by Sarwar and colleagues, and remains a common baseline and component in modern recommender systems, sometimes combined with model-based approaches for improved performance.