We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Efficient AutoML Pipeline Search with Matrix and Tensor Factorization

Abstract: Data scientists seeking a good supervised learning model on a new dataset have many choices to make: they must preprocess the data, select features, possibly reduce the dimension, select an estimation algorithm, and choose hyperparameters for each of these pipeline components. With new pipeline components comes a combinatorial explosion in the number of choices! In this work, we design a new AutoML system to address this challenge: an automated system to design a supervised learning pipeline. Our system uses matrix and tensor factorization as surrogate models to model the combinatorial pipeline search space. Under these models, we develop greedy experiment design protocols to efficiently gather information about a new dataset. Experiments on large corpora of real-world classification problems demonstrate the effectiveness of our approach.
Comments: This is an extended version of AutoML Pipeline Selection: Efficiently Navigating the Combinatorial Space (DOI: 10.1145/3394486.3403197) at KDD 2020
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
Cite as: arXiv:2006.04216 [cs.LG]
  (or arXiv:2006.04216v1 [cs.LG] for this version)

Submission history

From: Chengrun Yang [view email]
[v1] Sun, 7 Jun 2020 18:08:48 GMT (6201kb,D)

Link back to: arXiv, form interface, contact.