Graph Embeddings for Machine Learning Problems on Relational Data

Dynamic graph analysis is concerned with many tasks, directly relevant to understanding and working with modern large scale interconnected data structures. One of such tasks is representation learning, which attempts to create a standardized encoding for each object or connection in the network. These representations are intended to be leveraged to solve various practical tasks - for instance, link prediction or node classification. In this paper, we combine two state of the art approaches for representation learning on temporal networks to propose a novel method, which we call TGN-CAW(after the named approaches). Our method uses the general framework suggested in the TGN paper and enhances it by using a highly custom message generating function. This function is modeled after the highly effective representations obtained by extracting CAWs, or Causal Anonymous Walks, from the network - established in the CAW paper. Our experimental setup indicates improvement over existing models in terms of quality metrics, on various datasets and evaluation tasks. Experiments and evaluation were conducted on our novel training and validation framework. By providing unified approach to dynamic graph models and data, framework may bolster further research in the field.