Enabling Causal Hypothesis Management via Automated Feature Extraction in Graph Databases

Physical Sciences Inc. (PSI) and Pacific Northwest National Laboratory (PNNL) propose to develop a software system designed to enable military analysts to identify and verify adversary courses of actions within large-scale intelligence surveillance and reconnaissance (ISR) datasets. The proposed Graph Feature ExTraction for Causal Hypotheses (Graph-FETCH) will ingest, align and analyze multiple diverse ISR data streams to generate and maintain a knowledge graph (KG). Graph search will be streamlined for analysts by leveraging language models to translate readable text into graph constraint searches. Machine learning enabled graph pattern extraction algorithms will be leveraged to identify patterns that are predictive future activities and events. Automated pattern extraction will enable the application of causal inference and discovery algorithms. Discovered predictive hypotheses will be translated into causal-graphs and quantitatively tested using causal-graph refutation algorithms. Graph-FETCH will leverage transitioned, operational graph capabilities for entity alignment, anomaly detection, and graph pattern search, developed by PSI under the DARPA SIGMA+ program. The proposed development aims to enhance the operational picture and planning capabilities of military analysts by accurately predicting future modeled events, using verified hypotheses, with an accuracy greater than 85% at less than 15% false alarm probability.