Scalable graph machine studying: a mountain we will climb?

By Kevin Jung, Software program Engineer at CSIRO Information61.

Graph machine studying continues to be a comparatively new and growing space of analysis and brings with it a bucket load of complexities and challenges. One such problem that each fascinates and infuriates these of us working with graph algorithms is — scalability.

Two strategies had been positioned early on as the usual approaches to leveraging community info: Graph Convolutional Networks [1] (highly effective neural community structure for machine studying on graphs) and Node2Vec [2] (an algorithmic framework for representational studying on graphs). Each these strategies might be very helpful for extracting insights from highly-connected datasets.

However I realized first-hand that when making an attempt to use graph machine studying strategies to establish fraudulent behaviour within the bitcoin blockchain knowledge, scalability was the most important roadblock. The bitcoin blockchain graph we’re utilizing has tens of millions of wallets (nodes) and billions of transactions (edges), which makes most graph machine studying strategies infeasible.

On this article, we’ll take a more in-depth take a look at scalability for graph machine studying strategies: what it’s, what makes it tough, and an instance of a way that tries to sort out it head-on.


What’s graph machine studying?

To begin with, let’s be sure we’re on the identical web page by way of what we imply by graph machine studying.

After we say ‘graph,’ we’re speaking a couple of manner of representing knowledge as entities with connections between them. In mathematical phrases, we name an entity a node or vertex, and a connection an edge. A set of vertices V, along with a set of edges E, type a graph G = (V, E).

Graph machine studying is a machine studying approach that may naturally be taught and make predictions from graph structured knowledge. We will consider machine studying as a manner of studying some transformation perform; y = f(x), the place x is a chunk of information and y is one thing we wish to predict.

Say we take the duty of detecting fraudulent bitcoin addresses for example, and we all know the account stability for all addresses on the blockchain. A quite simple mannequin would possibly be taught to foretell that if an tackle has a zero account stability, then it’s unlikely to be fraudulent. In different phrases, our perform f(x) represents a worth near zero (i.e. non-fraudulent) when x is zero:

We will agree that merely trying on the account stability of an tackle isn’t a lot to go by when making an attempt to unravel such an issue. So at this level, we might think about potential methods to engineer further options that give our mannequin extra details about every tackle’ behaviour.

What we have already got is the wealthy community construction from the transactions occurring between payers and payees of bitcoin. By designing a mannequin that leverages this info, we’d have extra confidence within the outcomes:

We wish to make predictions about addresses not solely based mostly on their account stability but additionally based mostly on transactions made with different addresses. We will attempt to formulate f such that it’s within the type f(x, x’₀, x’₁, …) the place x’ᵢ are different knowledge factors within the native neighbourhood of x as outlined by our graph construction.

One solution to obtain that is by making use of graph construction within the type of an adjacency matrix. Multiplying the enter knowledge by the adjacency matrix (or some normalisation of) has the impact of linearly combining an information level with its adjoining factors.

Under is a illustration of a graph with three nodes as an adjacency matrix and a set of options:

The neighbourhood of node zero might be aggregated as:

That is the fundamental high-level precept adopted by algorithms like graph convolutional networks. Usually, the aggregation of native neighbourhood info is utilized recursively to extend the dimensions of the native community that’s pulled collectively to make predictions a couple of node. (Learn StellarGraph’s article Knowing Your Neighbours: Machine Learning on Graphs for a extra thorough introduction to those ideas).


What’s scalable?

A scalable mountain is a mountain that folks can climb. A scalable system is a system that may deal with rising calls for. A scalable graph machine studying methodology needs to be a way that handles rising knowledge sizes… and it additionally occurs to be an enormous mountain to climb.

Right here, I’m going to argue that the fundamental precept of naively aggregating throughout a node’s neighbourhood is not scalable, and describe the issues an algorithm should remedy as a way to be thought of in any other case.

The primary downside stems from the truth that one node might be related arbitrarily to many different nodes from a graph – even all the graph.

In a extra conventional deep studying pipeline, if we wish to predict one thing about x, we solely want details about x itself. However with the graph construction in thoughts, as a way to predict one thing about x we probably have to combination info from all the dataset. As a dataset will get bigger and bigger, all of a sudden we find yourself aggregating terabytes of information simply to make a prediction a couple of single knowledge level. This doesn’t sound so scalable.

The reason for the second downside includes understanding the distinction between a transductive and an inductive algorithm.

Inductive algorithms attempt to uncover a basic rule for the world. The mannequin takes the info as a foundation for making predictions for unseen knowledge.

Transductive algorithms try and make higher predictions for the unlabelled knowledge in a dataset by not generalising a common mannequin.

After we’re making an attempt to sort out issues in the actual world, we’re met with the problem that the info isn’t static. Gigabytes of latest knowledge might current each day, and that is what makes scalability such an essential consideration. However many graph machine studying strategies are inherently transductive because of the manner info is aggregated from all the dataset, versus simply taking a look at a single occasion of information.

Let’s check out a extra concrete instance that demonstrates this downside. Take into account a node A in a graph that’s related to 3 different nodes B, C, and D:

If we weren’t making use of any fancy graph strategies, we might merely be studying a perform that maps from the options of A to a extra helpful metric; e.g., a prediction we wish to make in regards to the node:

Nevertheless, as we wish to make use of the graph construction, we find yourself taking the options of B, C, and D as enter for the perform we’re studying:

Take into account that after we’ve skilled the mannequin, a brand new knowledge level arrives a while sooner or later that occurs to be related to our authentic node A. We ended up studying a perform that doesn’t take this connection under consideration, so we’re caught in a state of affairs the place we’re not sure whether or not the mannequin we skilled is legitimate for our new set of information.

Node E and Edge AE are launched, inflicting the Mannequin to additionally carry within the options of E when aggregating neighbourhood info to make a brand new prediction for A.

To date, our understanding of graph algorithms suggests they’re usually not very scalable, notably if the algorithm is transductive in nature. Subsequent, we’ll discover an algorithm that makes an attempt to sort out a few of these challenges.


Introducing GraphSAGE

A typical manner many algorithms attempt to sort out the scalability downside in graph machine studying is to include some type of sampling. One explicit method we’ll focus on on this part is the strategy of neighbour-sampling, which was launched by the GraphSAGE [3] algorithm.

The SAGE in GraphSAGE stands for Pattern-and-Mixture, which in easy phrases means: “for each node, take a sample of nodes from its local neighbourhood, and aggregate their features.”

The ideas of “taking a sample of its neighbours” and “aggregating features” sound fairly obscure, so let’s discover what they really imply.

GraphSAGE prescribes that we take a set measurement pattern of any given node’s native neighbourhood. This instantly solves our first downside of needing to combination info from throughout all the dataset. However what are we sacrificing by doing so?

  1. First and most clearly, taking a pattern means we’re taking an approximation of what the neighbourhood really seems like. Relying on the dimensions of the pattern we select to take, it could be a adequate approximation for our functions, however an approximation nonetheless.
  2. We hand over the prospect for our mannequin to be taught one thing from how related a node is. For GraphSAGE, a node with 5 neighbours seems precisely the identical as a node with 50 neighbours since we all the time pattern the identical variety of neighbours for every node.
  3. Lastly, we find yourself in a world the place we might make totally different predictions a couple of node based mostly on which neighbours we occurred to pattern on the time.

Relying on the issue we’d like to unravel and what we find out about our knowledge, we will attempt to take a guess at how these points might have an effect on our outcomes and decide about whether or not GraphSAGE is an acceptable algorithm for a selected use-case.

Aggregating options might be carried out in plenty of other ways, however every might be described as a perform that takes an inventory of options from the sampled neighbourhood and outputs an ‘aggregated’ characteristic vector.

For instance, the imply aggregator merely takes the element-wise imply of the options:

GraphSAGE imply aggregator.

We will then apply a second aggregation step to mix the options of the node itself and its aggregated neighbours. A easy manner this may be carried out, demonstrated above, is to concatenate the 2 characteristic vectors and multiply this with a set of trainable weights.

The native sampling nature of GraphSAGE offers us each the inductive algorithm and a mechanism to scale. We’re additionally ready to decide on the aggregation methodology to provide us some flexibility within the mannequin. Although these advantages come at a value, the place we have to sacrifice mannequin efficiency for scalability. Nevertheless, for our functions, the GraphSAGE algorithm offered an excellent method for scaling graph machine studying on the bitcoin dataset.


Success, however not with out challenges

GraphSAGE presents the neighbourhood sampling method to beat among the challenges for scalability. Particularly, it:

  • offers us an excellent approximation whereas bounding the enter measurement for making predictions; and
  • permits for an inductive algorithm.

It is a stable breakthrough however doesn’t go away us with out issues to be solved.

1. Environment friendly sampling continues to be tough

As a way to pattern the neighbours of a node with out introducing bias, you continue to have to iterate via all of them. This implies though GraphSAGE does limit the dimensions of the enter to the neural community, the step required to populate the enter includes trying via all the graph, which might be very pricey.

2. Even with sampling, neighbourhood aggregation nonetheless aggregates A LOT of information

Even with a set neighbourhood measurement, making use of this scheme recursively implies that you get an exponential explosion of the neighbourhood measurement. For instance, if we take 10 random neighbours every time however apply the aggregation over three recursive steps, this finally leads to a neighbourhood measurement of 10³.

3. Distributed knowledge introduces much more challenges for graph-based strategies

A lot of the massive knowledge ecosystem revolves round distributing knowledge to allow parallelised workloads and supply the power to scale out horizontally based mostly on demand. Nevertheless, naively distributing graph knowledge introduces a major downside as there isn’t a assure that neighbourhood aggregation might be carried out with out communication throughout the community. This leaves graph-based strategies in a spot the place you pay the price of shuffling knowledge throughout the community or miss out on the worth of utilizing huge knowledge applied sciences to allow your pipeline.

There are nonetheless mountains to climb and extra exploration to be carried out to make scalable graph machine studying extra sensible. I, for one, will probably be paying shut consideration to new developments on this house.

When you’d wish to be taught extra about graph machine studying, be happy to download the open supply StellarGraph Python Library or contact us at

This work is supported by CSIRO’s Information61, Australia’s main digital analysis community.



  1. Graph Convolutional Networks (GCN): Semi-Supervised Classification with Graph Convolutional Networks. Thomas N. Kipf, Max Welling. Worldwide Convention on Studying Representations (ICLR), 2017
  2. Node2Vec: Scalable Characteristic Studying for Networks. A. Grover, J. Leskovec. ACM SIGKDD Worldwide Convention on Information Discovery and Information Mining (KDD), 2016.
  3. Inductive Illustration Studying on Massive Graphs. W.L. Hamilton, R. Ying, and J. Leskovec. Neural Info Processing Programs (NIPS), 2017.

Original. Reposted with permission.

Bio: Kevin Jung is a Software program Engineer working within the StellarGraph group at CSIRO’s Information61, Australia’s main digital analysis community.



About the Author

Leave a Reply

Your email address will not be published. Required fields are marked *