Collaborative Filtering based Generative Networks

Collaborative Filtering, a popular method for recommendation engines, models its predictions using past interactions between the entities in question (aka users/movies or customers/products etc). The method does not rely on the explicit properties of the entities, the identification of which may be intractable. In this work, we leverage this advantage rendered by Collaborative Filtering where the explicit features need not be defined apriori by evaluating its application to the domain of Ligand based Virtual Screening. We further attempt to address the drawback of Collaborative Filtering , ie the lack of interpret ability of the factors discovered through collaborative filtering by creating a novel class of generative deep learning models ,called Collaborative Filtering based Generative Networks (CFGenNets). We show the utility of CFGenNets in 2 domains 1) Ligand based Virtual Screening and 2) Image generation from keyword tags/meta descriptors