Guardado en:
Detalles Bibliográficos
Autores principales: Medvedev, Marko, Lyu, Kaifeng, Li, Zhiyuan, Srebro, Nathan
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:https://arxiv.org/abs/2510.25108
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866918191580577792
author Medvedev, Marko
Lyu, Kaifeng
Li, Zhiyuan
Srebro, Nathan
author_facet Medvedev, Marko
Lyu, Kaifeng
Li, Zhiyuan
Srebro, Nathan
contents We consider training and testing on mixture distributions with different training and test proportions. We show that in many settings, and in some sense generically, distribution shift can be beneficial, and test performance can improve due to mismatched training proportions, even if the components are unrelated and with no transfer between components. In a variety of scenarios, we identify the optimal training proportions and the extent to which such distribution shift can be beneficial. We show how the same analysis applies also to a compositional setting with differing distribution of component "skills'' at training and test.
format Preprint
id arxiv_https___arxiv_org_abs_2510_25108
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Shift is Good: Mismatched Data Mixing Improves Test Performance
Medvedev, Marko
Lyu, Kaifeng
Li, Zhiyuan
Srebro, Nathan
Machine Learning
We consider training and testing on mixture distributions with different training and test proportions. We show that in many settings, and in some sense generically, distribution shift can be beneficial, and test performance can improve due to mismatched training proportions, even if the components are unrelated and with no transfer between components. In a variety of scenarios, we identify the optimal training proportions and the extent to which such distribution shift can be beneficial. We show how the same analysis applies also to a compositional setting with differing distribution of component "skills'' at training and test.
title Shift is Good: Mismatched Data Mixing Improves Test Performance
topic Machine Learning
url https://arxiv.org/abs/2510.25108