Improving Relational Machine Learning by Modeling Temporal Dependencies.

摘要

Networks encode dependencies between entities (people, computers, proteins) and allow us to study phenomena across social, technological, and biological domains. These networks naturally evolve over time by the addition, deletion, and changing of links, nodes, and attributes. Existing work in Relational Machine Learning (RML) has ignored relational time series data consisting of dynamic graphs and attributes, even despite the importance of modeling these dynamics.;This dissertation investigates the problem of Relational Time-series Learning from dynamic attributed graph data, with the goal of improving the predictive quality of existing RML methods. In particular, we propose a framework for learning dynamic graph representations, as well as methods for the three representation discovery tasks of (i) dynamic node labeling, (ii) weighting, and (iii) prediction. In addition, techniques for modeling relational and temporal dependencies are proposed, along with efficient methods for discovering features, ensembles, as well as classification methods. The results demonstrate the importance of modeling both the relational and temporal dependencies as well as learning an appropriate graph data representation that captures these fundamental patterns. Furthermore, while previous work has focused on static graphs that are small, non-attributed, simple, or homogeneous, we instead have carefully designed generalized relational time-series models that are: (a) efficient with linear or nearly linear runtime, (b) scalable for big graph data, (c) flexible for a variety of data types and characteristics, and (d) capable of modeling attributed and heterogeneous relational time-series data. Finally, the proposed methods are shown to be scalable, effective, and flexible for a variety of real-world applications.