|
Abstract
|
The advent of automated machine learning (AutoML) has revolutionized the process of model development, particularly in addressing the challenge of missing data. Missing values are a common issue that can substantially compromise the performance and reliability of machine learning models. Among the techniques used for imputation, K-Nearest Neighbors (KNN) stands out due to its simplicity and proven effectiveness. This research focuses on evaluating the effectiveness of KNN as a strategy for managing missing data, highlighting its impact on improving model accuracy and reliability across diverse applications, including both classification and regression tasks.
KNN is based on the principle of proximity, classifying or predicting outcomes by considering the nearest data points in the feature space. One of its key advantages is its independence from strong assumptions about the underlying data distribution, making it highly adaptable across various datasets. However, the selection of the parameter k, which specifies the number of neighbors to consider, is critical to its performance. A small k can lead to overfitting by making the model overly sensitive to noise, while a large k may result in underfitting, overlooking essential patterns in the data. This balance is vital to achieving optimal predictive accuracy.
|