|Friday, February 21|
|CS15 Rumors and Recommendations||
Fri, Feb 21, 3:15 PM - 4:45 PM
When Some Recommendations Are More Important Than Others: Combining Weighted Least Squares and Matrix Factorization for Recommender Systems (302765)*Calvin Price, American Express
Keywords: recommender systems
When producing recommendations it is common to first create a set of predicted ratings between users and items. In this paper we show how rating predictions can be produced so that certain observations receive more emphasis than others, in the sense of getting a lower prediction error. This is achieved by adding a weighted least squares component into the traditional SVD cost function for recommender systems. An example is given using ratings from the MovieLens dataset. Here we consider the case where predictions for extreme ratings, meaning a rating value of 1 or 5, are considered more critical to get right than other rating outcomes. Measures of recall increase sharply for these rating outcomes after using WLS (recall for other rating outcomes do suffer moderate decreases, but in some cases also improve).
This principle is very flexible and can be applied in a number of other manners. More generally, it can be applied anywhere a user has some basis for wanting some observations to be predicted better than they would normally be when all observations are treated equally. Other examples could include wanting better ratings for those users that have rated a small (or large) number