Using Autoencoders for Advanced Residential Area Classification

Alex Singleton from the University of Liverpool explains how to use an autoencoder, a type of neural network, to improve cluster analysis of geographic data. Learn how this "residential area encoding" (RAE) method transforms variables to create a compressed representation, revealing nonlinear relationships and leading to more descriptive and accurate classifications. Inspired by this clip? Watch the full talk and all sessions from SDSC NYC 2025 on demand: https://spatial-data-science-conference.com/2025/newyork#Watch-on-demand #autoencoder #neuralnetworks #spatialdatascience #SDSC25 Transcript: 00:02 If we have a selection of variables. In our experiment, we took all the variables that go into the national classification I showed you at the start. What we're going to do is try and transform those in a way which means we can synthesize a set of inputs that are going to create better clusters in terms of their descriptive properties at the other end. 00:27 The way we do that is we use something called an autoencoder. We're calling this method residential area encoding. An autoencoder is a neural network and it takes a set of input data. It applies the autoencoder to it. It goes through this encoding phase where it's trying to synthesize down the properties of the data to a compressed representation which is this bottleneck layer and it optimizes this method by fitting a decoder function which then 01:03 is trying to recreate the original data. It's iteratively trying to recreate the individual data from this compressed representation. The models run and at some point produce a model which is as good as it can get and that bottleneck layer is the compressed representation that we use in the cluster analysis. 01:24 One of the things about autoencoders which are neat is the different layers that you can have in your neural network could have what they call nonlinear activation functions added to them which means that the models themselves expect nonlinear linear attributes of the data and can learn that nonlinear structure. 01:43 Picking up on some of those things I was showing you in the graph with the lines about digital access digital technologies this enables you to map out some of those nonlinear relationships in your compressed representation. Once we've got the bottleneck layer, we can do something like select how many clusters we think a model might want to comprise. 02:08 In this case we use silhouette scores and we select a model that's eight clusters which is quite useful because that also mirrors the national classification as well.