The long-term vision of this project is to build a set of personalized intelligent assistant services completely locally on the user's personal device, without forking over private data to a cloud-service.
We start with a news recommendation service. We built a PrIA news recommendation service that collects user's personal data, builds a profile, and recommends news articles based on the profile, all locally on the user's laptop. Our small scale, IRB-approved, user study shows that while the effectiveness of PrIA's news recommendation is lower than that of Google's recommendations, the difference in performance is not significant. The user study is currently offline, but the code and instructions are available here.
In effect, in PrIA, our goal is to combine NLP, Systems, and Privacy to support intelligent assistant services without losing privacy. We are currently working on three aspects of PrIA: (1) Creating a privacy dashboard of users to learn what sensitive information about themselves can be inferred from their reading habits. We propose to use the privacy dashboard to significantly improve the PrIA news recommendation, (2) Designing a new cloud interaction architecture (for news personalization and other intelligent assistant services) without leaking private information, and (3) Extending PrIA to provide other personalization services.