Cascading Succession of Models for an Enhanced Long-Tail Discernment AI System

A Spatio-Temporal Knowledge Graph (STKG) with a Type-Sensitive extension (a.k.a., TS-STKG or T2S2KG) and a quintuple representation for spatio-temporal facts is examined for prospective use. Certain embedding models — to contend with complex relationships (e.g., 1-to-N, N-to-1, N-to-N, etc.) — are utilized in a cascading fashion. Various negative sampling techniques are used at each major cascade, and this ensemble was explored against the Yet Another Great Ontology (YAGO)3-10 dataset. As the long-tail phenomenon is prevalent, with its concomitant unbalanced data (and bias towards favoring head classes), selected T2S2TKG Embedding (T2S2TKGE) techniques were utilized to better balance between head and tail classes. These were sorted into cascades based upon their performance against the various complex relationships, and their time as well as space complexities were considered as well. Various architectural constructs were explored, and a Graph Convolutional Network (GCN)-Bidirectional Long Short-Term Memory (BiLSTM)-“GraphSAGE”-inspired Graph-Attention-Network (GSGAT) mechanism along with a Robust Convex Relaxation (RCR)-based Deep Convolutional Neural Network (DCNN) Generative Adversarial Network (GAN) (DCGAN)-DCNN-1,2,3 amalgam shows promise for operationalizing the aforementioned.