Comparing State-of-the-Art Collaborative Filtering Systems

The document compares different state-of-the-art collaborative filtering systems. It finds that item-based collaborative filtering performs best with a mean absolute error of 0.6382 using probabilistic similarity and 400 neighbors. User-based approaches work best with 1500 neighbors and predicting using deviation from the mean. Cluster-based approaches have the highest error rate of 0.6736 using K-means clustering into 4 clusters and predicting with Bayes. Item-based approaches require fewer neighbors and scale better to large datasets.