Human-Assisted Refinement of Personalized News Recommendation
With the advance of machine learning from smartphones to healthcare, there have been breakthroughs in the research of both the algorithms and applications in this domain. In parallel, much research has been done on crowdsourcing and human computer interaction as well. However, very little research has been done on leveraging human interaction with the potent machine learning algorithms developed. In this project, the goal is to investigate potential ways to best combine the advances in artificial intelligence and machine learning with human knowledge and opinions in the domain of news. Contextual bandits are used to create a user model and a rankings system which can then be adjusted and manipulated directly by the user or through other indirect actions. Stacked autoencoders are used to condense a wide swath of news topics into broader categories for the user to interact with at first, with further interaction allowing the bandit system to pinpoint which topics the user prefers. In addition, much deliberation and care was put into the design of the interface to ensure the system can get user feedback in as clean and efficient a manner as possible.