Stratification by Quantiles and Compact Genetic Algorithms for Dataset Compression in Nonlinear Regression

Abstract

K-nearest neighbors (KNN) is a supervised learning technique that predicts the outcome of a sample based on the samples surrounding it. One of the main drawbacks of this method is the memory required for prediction. All samples must be stored, so the inference time increases with the size of the dataset. For this reason, this work presents a technique based on genetic algorithms (GA) to select the most relevant samples, thereby reducing the complexity of the KNN algorithm. The main objective of GA is to use the samples that contribute most to the prediction in a regression system, reducing the number of samples and the computational load of KNN. The results of this work show a significant improvement in inference times when combining GA with the KNN method.