The Solution: Machine Learning
To solve these challenges, we designed, built, and deployed a system called T-Rex, or the NASA Technology Taxonomy Recommender System. The contributions of T-Rex is notable across multiple areas from model selection to optimization and deployment. Our recommender is integrated into TechPort in a novel, effective manner and other groups at NASA are now using T-Rex to programmatically classify technology through an exposed API. The system has since been used in other projects to train and build recommenders, such as the NASA Technology Target Destination system. We found our model achieves over 96% accuracy in k-fold training, and over 97.5% accuracy in measured user updates.
T-Rex is an amalgamation of classifier models, custom models, and voting optimization and pruning methods for efficient, accurate, deployed models. The system automatically optimizes, selects, and combines multiple individual models. It uses prior classification information only when accurate to suggest a subset of classes as recommendations for other models. Our training algorithm optimizes weighting from individual models from a matrix of individual model outputs to reduce training time.
T-Rex has been deployed now for over two years and has recommended over 40,000 classifications. The methodology has since been extended to other applications within the Agency, including the identification of target technology destinations and gap analysis.
Read more about T-Rex in the 2022 whitepaper: https://ntrs.nasa.gov/citations/20220016471
TechPort can be found at https://techport.nasa.gov.