Modeling diversity of sets of items is impor-tant in many applications such as productrecommendation and data summarization.Probabilistic submodular models, a familyof models including the determinantal pointprocess, form a natural class of distributions,encouraging effects such as diversity, repul-sion and coverage. Current models, however,are limited to small and medium numberof items due to the high time complexityfor learning and inference. In this paper,we proposeFLID, a novel log-submodulardiversity model that scales to large numbersof items and can be efficiently learned usingnoise contrastive estimation.We showthat our model achieves state of the artperformance in terms of model fit, but canbe also learned orders of magnitude faster.We demonstrate the wide applicability ofour model using several experiments.

Top