Estimation of minimum sample size for identification of the most important features: A case study providing a qualitative B2B sales data set

Abstract

An important task in machine learning is to reduce data set dimensionality, which in turn contributes to reducing computational load and data collection costs, while improving human understanding and interpretation of models. We introduce an operational guideline for determining the minimum number of instances sufficient to identify correct ranks of features with the highest impact. We conduct tests based on qualitative B2B sales forecasting data. The results show that a relatively small instance subset is sufficient for identifying the most important features when rank is not important.