Improved Estimation of Ranks for Learning ItemRecommenders with Negative Sampling

In recommendation systems, there has been a growth in the num-ber of recommendable items (# of movies, music, products). Whenthe set of recommendable items is large, training and evaluationof item recommendation models becomes computationally expen-sive. To lower this cost, it has become common to sample negativeitems. However, the recommendation quality can suffer from biasesintroduced by traditional negative sampling mechanisms.In this work, we demonstrate the benefits from correcting thebias introduced by sampling of negatives. We first provide sampledbatch version of the well-studied WARP and LambdaRank methods.Then, we present how these methods can benefit from improvedranking estimates. Finally, we evaluate the recommendation qualityas a result of correcting rank estimates and demonstrate that WARPand LambdaRank can be learned efficiently with negative samplingand our proposed correction technique.