Hyperparameter optimization based on a priori and a posteriori knowledge about classification problem

Subject of Research. The paper deals with Bayesian method for hyperparameter optimization of algorithms, used in machine learning for classification problems. A comprehensive survey is carried out about using a priori and a posteriori knowledge in classification task for hyperparameter optimization...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Nauchno-tekhnicheskiĭ vestnik informat͡s︡ionnykh tekhnologiĭ, mekhaniki i optiki Ročník 20; číslo 6; s. 828 - 834
Hlavní autoři: Smirnova, V.S., Shalamov, V.V., Efimova, V.A., Filchenkov, A.A.
Médium: Journal Article
Jazyk:angličtina
Vydáno: ITMO University 01.12.2020
Témata:
bayesian optimization
classification
gaussian processes
hyperparameter optimization
machine learning
ISSN:2226-1494, 2500-0373
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Subject of Research. The paper deals with Bayesian method for hyperparameter optimization of algorithms, used in machine learning for classification problems. A comprehensive survey is carried out about using a priori and a posteriori knowledge in classification task for hyperparameter optimization quality improvement. Method. The existing Bayesian optimization algorithm for hyperparameter setting in classification problems was expanded. We proposed a target function modification calculated on the basis of hyperparameters optimized for the similar problems and a metric for determination of similarity classification problems based on generated meta-features. Main Results. Experiments carried out on the real-world datasets from OpenML database have confirmed that the proposed algorithm achieves usually significantly better performance results than the existing Bayesian optimization algorithm within a fixed time limit. Practical Relevance. The proposed algorithm can be used for hyperparameter optimization in any classification problem, for example, in medicine, image processing or chemistry.
ISSN:2226-1494
2500-0373
DOI:10.17586/2226-1494-2020-20-6-828-834