Discovery of Temporal Graph Functional Dependencies

Abstract

Temporal graphs play a crucial role in modelling dynamic applications where (node) attribute values change over time. However, in some domains, specific attribute value relationships are expected to remain consistent over a time-interval. For example, in therapeutic drug monitoring, the measurement of specific drug levels is expected to remain constant at defined time intervals. A patient must be given a drug dose within a time interval to ensure effective and therapeutic drug concentrations remain in the body. Existing work has largely focused on static graphs, and are unable to capture attribute value relationships over a time interval. Recent work has proposed a new class of dependencies, Temporal Graph Functional Dependencies (TGFDs) that model such semantics [Alipour et. al]. However, declarative specification of these constraints is challenging given the large-scale of real-life graphs.

In this thesis, we propose, ParaTGFDMiner, a parallel discovery algorithm to identify TGFDs from large-scale graphs. Given an input graph G, ParaTGFDMiner discovers TGFDs by first generating candidate constraints locally at each worker node from a subset of G. A coordinator then validates and merges these candidate constraints, across all workers, to compute high-frequency TGFDs that are satisfied in G. We propose two optimizations that implement asynchronous processing to minimize idle time, and fast graph pattern matching to identify high-likelihood candidates. Our evaluation on two real large-scale graphs demonstrate that ParaTGFDMiner is up to 60% faster than baseline approaches.