A Genetic Programming Encoder for Increasing Autoencoder Interpretability

Autoencoders are powerful models for non-linear dimensionality reduction. However, their neural network structure makes it difficult to interpret how the high dimensional features relate to the lowdimensional embedding, which is an issue in applications where explainability is important. There have been attempts to replace both the neural network components in autoencoders with interpretable genetic programming (GP) models. However, for the purposes of interpretable dimensionality reduction, we observe that replacing only the encoder with GP is sufficient. In this work, we propose the Genetic Programming Encoder for Autoencoding (GPE-AE). GPE-AE uses a multi-tree GP individual as an encoder, while retaining the neural network decoder. We demonstrate that GPE-AE is a competitive non-linear dimensionality reduction technique compared to conventional autoencoders and a GP based method that does not use an autoencoder structure. As visualisation is a common goal for dimensionality reduction, we also evaluate the quality of visualisations produced by our method, and highlight the value of functional mappings by demonstrating insights that can be gained from interpreting the GP encoders.