That is, the prediction is based on the user’s average plus the average deviation from the user mean for the item in question across all users in the training set. We also compare to the item-based approach that is reported to work best [14], which uses the following adjusted cosine similarity measure, given two items i and j: