We are exploring how we can use reinforcement learning to improve the accuracy and diversity of personalized recommender systems. We are currently looking at an application in news recommendations. By incorporating natural language processing tasks of topic extraction and word embeddings, we hope to capture user dynamic preferences on news item consumption for personalized recommendations by a recommender agent.

We are looking to establish relationships between entities in a dataset by building a knowledge graph with those entities and the relation that exists among them. By so doing, we hope to use graph algorithms and machine learning methods to efficiently extract vector representations of each entity that represents/captures how they are related to all the other entities in the dataset. Consequently, we can use these vectors as embeddings/feature representations for a recommender system.

We are looking to discover ways in which unconscious bias is introduced in recommender systems by detecting possible data and machine learning algorithmic bias. Then we intend to develop models/frameworks to mitigate such biases.