Graph Machine Learning Papers

Semi-Supervised Classification with Graph Convolutional Neural Networks

Kipf & Welling (2017)

Parker's Take: A great paper to dive head first into the world of graph machine learning with. Introduces the concept of a convolution on a graph, and produced state-of-the-art results at the time of publish. Personally, I have used a GCN to find fraud and predict IPOs of companies.

Graph Attention Networks

Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, Yoshua Bengio (2018)

Parker's Take: While I still don't quite get the math, the algorithm described in this paper certainly an interesting one. Using an attention mechanism allows the algorithm to be more efficient/parallelizable when compared to things like a GCN. The model is also able to classify entire graphs, and can generalize to an unseen graph.

Graph Convolutional Neural Networks for Web-Scale Recommender Systems

Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, Jure Leskovec (2018)

Parker's Take: One of the issues with applying Kipf & Welling's GCN in the real world is that it does not scale well. This paper, written by Pinterest of all places, proposes a new framework for at-scale GCNs. They apply their algorithm to produce embeddings of each item in their 3 billion vertex graph in order for better recommendations. For graph ML at scale, this is an essential read.

Variational Graph Auto-Encoders

Kipf & Welling (2016)

Parker's Take: I think that VAEs are very cool. Applying them to graphs is even cooler. Allowing you to compress a highly-dimensional/non-dimensional structure into a relatively small vector astounds me. State-of-the-art results at link prediction because of it is even cooler.

Modeling Relational Data with Graph Convolutional Neural Networks

Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, Max Welling (2017)

Parker's Take: Another paper with Kipf & Welling involved. This time, they expand upon the GCN, describing a method that allows multiple different types of edges and vertices to be in the graph. They then use this to perform edge prediction and vertex classification. If you are using a knowledge graph, this is a must-read.

node2vec: Scalable Feature Learning for Networks

Aditya Grover, Jure Leskovec (2016)

Parker's Take: Great way to create embeddings for your vertices. Think about this as word2vec, but instead of words in a sentence, you are creating "sentences" of vertices through the randomly walking through them. The parameters p and q determine how the random walks are dispersed (focusing on local neighborhood vs. far reaches of graph), don't forget to tune them. This is a really easy method to generalize to knowledge graph content and scale.