Streaming Feature Learning (SIFTER)

Despite several commercial deployments and successes, current recommender systems require designers to hand tune features describing the data, which then remain static for months or years. Since modern applications increasingly rely on streaming data, the next generation of defense, intelligence and commercial recommender systems would need to automate the continuous updating and tuning of features in order to keep up with the trends and meet performance expectations. We propose to design, develop and validate a novel streaming feature learning method called SIFTER for producing high-performance features at the speed of data. Our effort will address several technical hurdles: (1) SIFTER must process streaming data to continuously discover new features at the speed of data; (2) SIFTER must automatically handle completely-new terms in data; (3) SIFTER must offset oldness bias against newer vocabulary terms; and (4) SIFTER must incorporate multiple iterations of training over the input data. We will identify real-world datasets and recommendation targets to validate our main hypothesis that SIFTER improves recommendation performance by both increasing quality and reducing latency. We believe that improved system recommendations would lead to better decision making by defense and intelligence analysts, eventually leading to better outcomes for our Warfighters.