Probabilistic Learning of Operator Interest in Surveillance Environments for Online Track Characterization

Abstract: Machine learning models that augment data-intensive human workflows must rely on an understanding of a user’s interests and behaviors. By understanding what a user desires, these models can help intuitively prioritize workflow and essential information for decision making, and form the baseline for trusted autonomous systems. This work considers the problem of human interest classification for missile defense surveillance. In this case, users are satellite operators who must prioritize simultaneous processing and characterization of targets across the globe for extended duration while working under strict time and accuracy constraints. Learning human interest in this context is particularly challenging because user interests are generally not static, tracks have a short lifespan, and the user pool is small. Here, we formulated the solution similarly as a binary Bayesian logistic regression problem to classify operator interest in a given candidate track, but with the added complexity of partially observable feature variables and dependencies between these variables. Our approach leverages domain knowledge to instantiate priors on human interest and, using online user interactions with the system, can continuously infer human interest for candidate tracks. We validate our user interest classification algorithm using simulated truth testing across various configurations of expected operator behavior, which overall show that the algorithm can effectively learn relative track interest with minimal training data.