Regularisation is typically applied to the optimisation objective of matrix factorisation methods in order to avoid over-fitting. In this paper, we explore the use of regularisation to enhance the diversity of the recommendations produced by these methods. Given a matrix of pairwise item distances, we add regularisation terms dependent on the item distances to the accuracy objective of a learning to rank matrix factorisation formulation. We examine the impact of these regularisers on the latent factors produced by the algorithm and show that such regularisation does indeed promote diversity. The regularisation comes at a cost of performance in terms of accuracy and ultimately the approach cannot greatly enhance diversity without a consequent fall-off in accuracy.
展开▼
