Data Science Desktop Survival Guide
by Graham Williams
Collaborative filtering is used in recommendation systems. It generates recommendations based on ratings of items by users.
The typical algorithm represents users and items in a matrix and records the ratings in the cells of the matrix. The aim is to match a new user's preferences for items with other existing users' preferences. The similarity measures used for matching include cosine similarity, Pearson correlation, and probability-based similarity. For the most similar users any items that the new user has not utilised but have been utilised by matching users, will be recommended.
Model-based collaborative filtering trains a predictive model from user-item ratings matrix. Modellers include Bayesian, latent semantic model, and SVD. Typically the training data consists of user-item pairs and the rating.