Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Automated Machine Learning for Unsupervised Tabular Tasks

Prabhant Singh, Pieter Gijsbers, Elif Ceren Gok Yildirim, Murat Onur Yildirim, Joaquin Vanschoren

TL;DR

LOTUS addresses unsupervised tabular AutoML by learning to select pipelines through meta-learning and dataset similarity. It relies on $GW$ (Gromov-Wasserstein) distance, computed on FastICA-transformed data, to identify the most similar prior dataset and transfer its optimal pipeline. Across outlier detection and clustering, LOTUS demonstrates strong performance against baselines and provides open-source AutoML tools, highlighting the practicality of a unified similarity-based model selection framework for unsupervised learning. By avoiding extensive meta-feature engineering and leveraging a two-phase learning strategy, LOTUS offers a scalable approach to automate unsupervised pipeline selection in diverse tabular domains.

Abstract

In this work, we present LOTUS (Learning to Learn with Optimal Transport for Unsupervised Scenarios), a simple yet effective method to perform model selection for multiple unsupervised machine learning(ML) tasks such as outlier detection and clustering. Our intuition behind this work is that a machine learning pipeline will perform well in a new dataset if it previously worked well on datasets with a similar underlying data distribution. We use Optimal Transport distances to find this similarity between unlabeled tabular datasets and recommend machine learning pipelines with one unified single method on two downstream unsupervised tasks: outlier detection and clustering. We present the effectiveness of our approach with experiments against strong baselines and show that LOTUS is a very promising first step toward model selection for multiple unsupervised ML tasks.

Automated Machine Learning for Unsupervised Tabular Tasks

TL;DR

LOTUS addresses unsupervised tabular AutoML by learning to select pipelines through meta-learning and dataset similarity. It relies on (Gromov-Wasserstein) distance, computed on FastICA-transformed data, to identify the most similar prior dataset and transfer its optimal pipeline. Across outlier detection and clustering, LOTUS demonstrates strong performance against baselines and provides open-source AutoML tools, highlighting the practicality of a unified similarity-based model selection framework for unsupervised learning. By avoiding extensive meta-feature engineering and leveraging a two-phase learning strategy, LOTUS offers a scalable approach to automate unsupervised pipeline selection in diverse tabular domains.

Abstract

In this work, we present LOTUS (Learning to Learn with Optimal Transport for Unsupervised Scenarios), a simple yet effective method to perform model selection for multiple unsupervised machine learning(ML) tasks such as outlier detection and clustering. Our intuition behind this work is that a machine learning pipeline will perform well in a new dataset if it previously worked well on datasets with a similar underlying data distribution. We use Optimal Transport distances to find this similarity between unlabeled tabular datasets and recommend machine learning pipelines with one unified single method on two downstream unsupervised tasks: outlier detection and clustering. We present the effectiveness of our approach with experiments against strong baselines and show that LOTUS is a very promising first step toward model selection for multiple unsupervised ML tasks.

Paper Structure

This paper contains 25 sections, 15 equations, 7 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. AutoML for Unsupervised tasks
  3. Our Method
  4. Preliminary: Optimal Transport
  5. Method: Learning to learn with Optimal Transport for Unsupervised Scenarios
  6. Meta-training phase
  7. Model Selection Phase
  8. Finding a distance for Tabular datasets
  9. Computational Consideration
  10. Preprocessing
  11. Task-specific Implementations
  12. Clustering
  13. Outlier Detection
  14. Experimental setup
  15. Outlier Detection Experimental Setup
  16. ...and 10 more sections

Figures (7)

  • Figure 1: Figure demonstrating our approach to measure dataset similarity between two datasets from different domains.
  • Figure 2: Figure demonstrating our extensions that populates $\mathcal{A} = \{A^*_{\lambda^*_1}, ..., A^*_{\lambda^*_n}\}$
  • Figure 3: Comparison of average rank (lower is better) of methods w.r.t. outlier detection performance across datasets in ADBench. The differences in rank of methods connected by horizontal black bars are not a statistically significant.
  • Figure 4: ROPE test result, LOTUS vs MetaOD
  • Figure 5: ROPE test result of LOTUS vs (a) ABOD (b) HBOS (c) COF (d) IForest (e) LODA (f) KNN (g) OCSVM (h) LOF
  • ...and 2 more figures