Genetic Programming for Evolving a Front of Interpretable Models for Data Visualization

Abstract

Data visualization is a key tool in data mining for understanding big datasets. Many visualization methods have been proposed, including the well-regarded state-of-the-art method t-distributed stochastic neighbor embedding. However, the most powerful visualization methods have a significant limitation: the manner in which they create their visualization from the original features of the dataset is completely opaque. Many domains require an understanding of the data in terms of the original features; there is hence a need for powerful visualization methods which use understandable models. In this article, we propose a genetic programming (GP) approach called GP-tSNE for evolving interpretable mappings from the dataset to high-quality visualizations. A multiobjective approach is designed that produces a variety of visualizations in a single run which gives different tradeoffs between visual quality and model complexity. Testing against baseline methods on a variety of datasets shows the clear potential of GP-tSNE to allow deeper insight into data than that provided by existing visualization methods. We further highlight the benefits of a multiobjective approach through an in-depth analysis of a candidate front, which shows how multiple models can be analyzed jointly to give increased insight into the dataset.

Learning a Formula of Interpretability to Learn Interpretable Formulas

“…In fact, we agree that is simplistic. However, we believe that minimizing represents one of the first baselines to compare against (and it was the only one we found being used to specifically promote interpretability [22]), and that designing a competitive baseline is non-trivial. We will investigate this further in future work.…”

“…The authors of [43] study whether modern model-based GP can be useful when particularly compact symbolic regression solutions are sought, to allow interpretability. A very different take to enable or improve interpretability is taken in [22,41,45], where interpretability is sought by means of feature construction and dimensionality reduction. In [22] in particular, MOGP is used, with solution size as a simple PHI.…”

“…A very different take to enable or improve interpretability is taken in [22,41,45], where interpretability is sought by means of feature construction and dimensionality reduction. In [22] in particular, MOGP is used, with solution size as a simple PHI. Importantly, none of these works takes attempts to learn a PHI from data.…”

Model Learning with Personalized Interpretability Estimation (ML-PIE)

“…For decision trees and decision rules, besides reducing the number of features, approaches exist to restrict model size, prune unnecessary parts [5,26], aggregate local models in a hierarchy [48], or promote a trade-off between accuracy and complexity by means of loss functions [30,52] or prior distributions [33,60,61]. Regarding GP (and close relatives like grammatical evolution), perhaps the most simple and popular strategy to favor interpretability is to restrain the number of model components [16,31,57], sometimes in elaborate ways or particular settings [6,32,40,49,56]. Another strategy consists of penalizing models according to a weighted sum of the components that they include, after having pre-determined a weighing scheme [23,36].…”

Interpretable Dimensionally-Consistent Feature Extraction from Electrical Network Sensors

“…Recently there has been new application perspectives for Genetic Programming (GP) regarding the increasing need for interpretable results. GP was for example used to provide interpretable policies in reinforcement learning [14], to learn manifolds [25], to create visualizations [26] or to explain complex deep learning models [10]. For dimensionality reduction tasks, GP has also been used a lot as a feature construction method [35].…”