Representation Learning for Subset Generation
DescriptionIn many data-driven applications like online retailing and social influence analysis, the dataset we can observe is often a collection of subsets of a ground set. The underlying distribution over all subsets is very important in many decision making algorithms. Learning such a distribution is a challenging task and existing studies mainly focus on modeling the frequency of each single item such that complex correlations among items are not captured well. Moreover, a generator that produces random subsets according to the underlying subset distribution is often missed in existing studies. In this proposal, we propose a learning framework that learns representations of items in the ground set which can help us generate random subsets via sampling an item graph. Our framework incorporates adversarial learning such that the performance of our generator can be improved through a competition with a discriminator. We also provide a solution for learning dynamic generator of subsets when the subset distribution evolves over time. To the best of our knowledge, we are the first to combine representation learning and adversarial learning to learn subset distribution and subset generator. Moreover, many valuable applications can benefit from the subset generator learned by our framework, such as data-driven viral marketing and front-end warehouse inventory management.
|Effective start/end date||1/01/21 → …|