Contrastive Multi-View Representation Learning on Graphs

Autodesk deals with complex and mostly non-Euclidean data structures such as 3D models in Fusion, Revit, etc. This data usually does not carry any human annotations. Graphs are powerful data structures that can represent non-Euclidean data and also can carry other types of information such as semantic, visual, and geometric information in addition to the structure. Hence, one can treat graphs as a universal representation for Autodesk related data. With this assumption, this project aims to learn representations (i.e., dense low-dimensional vectors) of such graphs. Once this is done, one can for example compare the similarity between Revit or Fusion models, or can use only a few labels to categorize such models. This research can directly impact Future of AEC and Industry Initiatives and also can be utilized to enhance research on other initiatives such as robotics and HCI research.

Abstract

Contrastive Multi-View Representation Learning on Graphs

Kaveh Hassani, Amir Hosein Khasahmadi

International Conference on Machine Learning 2020

We introduce a self-supervised approach for learning node and graph level representations by contrasting structural views of graphs. We show that unlike visual representation learning, increasing the number of views to more than two or contrasting multi-scale encodings do not improve performance, and the best performance is achieved by contrasting encodings from first-order neighbors and a graph diffusion. We achieve new state-of-the-art results in self-supervised learning on 8 out of 8 node and graph classification benchmarks under the linear evaluation protocol. For example, on Cora (node) and Reddit-Binary (graph) classification benchmarks, we achieve 86.8% and 84.5% accuracy, which are 5.5% and 2.4% relative improvements over previous state-of-the-art. When compared to supervised baselines, our approach outperforms them in 4 out of 8 benchmarks.