Recommender Systems for Intelligence Analysts

Homeland security intelligence analysts need help finding relevant information quickly in a rapidly increasing volume of incoming raw data. Many different AI techniques are needed to handle this deluge of data. This paper describes initial investigations in the application of recommender systems to this problem. It illustrates various recommender systems technologies and suggests scenarios for how recommender systems can be applied to support an analyst. Since unclassified data on the search behavior of analysts is hard to obtain we have built a proof-ofconcept demo using analogous search behavior data in the computer science domain. The proof-of-concept collaborative recommender system that we developed is described. 1. Problem Description Homeland security and other intelligence analysts spend too much time on the mechanics of retrieving relevant information and not enough time on deep analysis. Retrieval usually needs to be initiated by the analyst (i.e., information pull). Existing information push technologies (that automatically find information for the analyst) such as RDF Site Summary (RSS) have very course grained channels that lead to information overload. Retrieval becomes even more complex for analysts who require multi-INT data from diverse heterogeneous sources (e.g., text, imagery, geospatial). Intelligence analysts also spend too much time-sharing information in multi-organizational teams. Information sharing is usually accomplished by person-to-person interactions (e.g., phone calls, email). Collaboration tools like whiteboards and chat rooms still require significant human effort. We believe that recommender systems can help solve these knowledge management problems. Assistant agents apply user profiles to proactively retrieve, share and recommend relevant information as shown in Figure 1. We are investigating hybrid aggregate knowledge representations and machine learning techniques for user profiles. Semantic Web based user profiles were described in [Kogut, 2004]. These Semantic Web based user profiles were entered by manually selecting classes and properties from a Web Ontology Language (OWL) ontology. This paper describes a recommender system approach that can automatically build a different form of user profiles, and can leverage existing profiles entered by the user. The approach takes advantage of similarities among user profiles for suggesting items of high importance to the analysts. The paper discusses ongoing efforts to develop and apply recommender systems for Homeland Security applications.