Collaborative Recommender System for Spatio-Temporal Intelligence Documents

NLP pipelines available today are getting robust for general language modeling purposes. But domain-specific data, abbreviations and lingos, and text about time or space still need a lot of tuning and training that are well beyond application of standard tool sets. Deep learning for recommendation engines is quite new, and all recommender systems, in particular for specially trained users, tend to have a high cost for collecting validation data from users. Hence the design of the user interface for the recommender system is critical for immediate and widespread adoption. Toward this end, in this proposal we propose the use of analytics tools from Topological Data Analytics (TDA). TDA-based tools have recently been used to "explain the structure" of the layers in trained CNNs for image analysis tasks. Our goal in this project will be to develop new TDA-based tools to fuse spatio-temporal information with text embedding. We will subsequently also develop novel user interface with explanation or justification of model-generated results to close the feedback loop on the recommendation system.