Network Structures, Concurrency, and Interpretability: Lessons from the Development of an AI Enabled Graph Database System

This thesis describes the development of the SmartGraph, an AI enabled graph database. The need for such a system has been independently recognized in the isolated fields of graph databases, graph computing, and computational graph deep learning systems, such as TensorFlow. Though prior works have investigated some relationships between these fields, we believe that the SmartGraph is the first system designed from conception to incorporate the most significant and useful characteristics of each. Examples include the ability to store graph structured data, run analytics natively on this data, and run gradient descent algorithms. It is the synergistic aspects of combining these fields that provide the most novel results presented in this dissertation. Key among them is how the notion of “graph querying” as used in graph databases can be used to solve a problem that has plagued deep learning systems since their inception; rather than attempting to embed graph structured datasets into restrictive vector spaces, we instead allow the deep learning functionality of the system to natively perform graph querying in memory during optimization as a way of interpreting (and learning) the graph. This results in a concept of natural and interpretable processing of graph structured data. Graph computing systems have traditionally used distributed computing across multiple compute nodes (e.g. separate machines connected via Ethernet or internet) to deal with large-scale datasets whilst working sequentially on problems over entire datasets. In this dissertation, we outline a distributed graph computing methodology that facilitates all the above capabilities (even in an environment consisting of a single physical machine) while allowing for a workflow more typical of a graph database than a graph computing system; massive concurrent access allowing for arbitrarily asynchronous execution of queries and analytics across the entire system. Further, we demonstrate how this methodology is key to the artificial intelligence capabilities of the system