Applying Optimal Weight Combination in Hybrid Recommender Systems

We propose a method for learning weighting schemes in weighted hybrid recommender systems (RS) that is based on statistical forecast and portfolio theory. An RS predicts the future preference of a set of items for a user, and recommends the top items. A hybrid RS combines individual RS in making the predictions. To determine the weighting of individual RS, we learn so-called optimal weights from the covariance matrix of available error data of individual RS that minimize the error of a combined RS. We test the method on the well-known MovieLens 1M dataset, and, contrary to the “forecast combination puzzle”, stating that a simple average (SA) weighting typically outperforms learned weights, the out-of-sample results show that the learned weights consistently outperform the individually best RS as well as an SA combination.