The impact of data selection strategies on distributed model performance

Abstract

Distributed Machine Learning, in which data and learning tasks are scattered across a cluster of computers, is one of the answers of the field to the challenges posed by Big Data. Still, in an era in which data abounds, decisions must still be made regarding which specific data to use on the training of the model, either because the amount of available data is simply too large, or because the training time or complexity of the model must be kept low. Typical approaches include, for example, selection based on data freshness. However, old data are not necessarily outdated and might still contain relevant patterns. Likewise, relying only on recent data may significantly decrease data diversity and representativity, and decrease model quality. The goal of this paper is to compare different heuristics for selecting data in a distributed Machine Learning scenario. Specifically, we ascertain whether selecting data based on their characteristics (meta-features), and optimizing for maximum diversity, improves model quality while, eventually, allowing to reduce model complexity. This will allow to develop more informed data selection strategies in distributed settings, in which the criteria are not only the location of the data or the state of each node in the cluster, but also include intrinsic and relevant characteristics of the data.