Learning to Rank for AutoML: enhancing pipeline selection with ranking information

Hernan Ceferino Vazquez, Pucci Romero Tobias, Verónica Bogado, Jorge Sánchez

Keywords: AutoML, Learning-to-Rank, Pipeline selection, Meta-learning, Ranking, Bayesian optimization, Monte Carlo tree search, OpenML datasets, Metric-agnostic, Machine Learning

TL;DR: This paper proposes a learning-to-rank framework for pipeline selection in AutoML, which improves on traditional methods by focusing on ranking rather than predicting performance, leading to more robust and metric-agnostic solutions.

Abstract: This paper introduces a learning-to-rank (LTR) framework to address the problem of pipeline selection in automated machine learning systems. The traditional approach to AutoML involves learning to predict the performance of various pipelines on a given task based on data acquired from previous tasks (i.e., meta-learning), which can be complex due to the need for different models for each task-specific metric. The proposed framework aims to select the best pipeline based on ranking rather than estimating a target metric, aligning more closely with the ultimate goal of the task (i.e., selecting pipeline candidates in order, from more to least promising). This approach enables more robust, metric-agnostic solutions that are easier to compare using ranking metrics like NDCG and MRR. The paper evaluates LTR strategies on public OpenML datasets, demonstrating a clear advantage for ranking-based methods. Additionally, the integration of LTR with Bayesian optimization and Monte Carlo tree search is explored, leading to improvements in the ranking metrics. Finally, the study found a strong correlation between ranking metrics (e.g., NDCG and MRR) and AutoML metrics, such as the task objective metric and the time to find the best solution, providing insights into how ranking-based methods could enhance AutoML systems.

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10512