A Novel Hyper-Tuned Multilayer Perceptron With Effective Stochastic Learning Strategies for Missing Values Imputation

A vast amount of data in many different formats is produced and stored daily, offering machine learning a valuable resource to enhance its predictive capabilities. However, the pervasiveness of inaccuracies in real-world data presents a significant barrier that can seriously limit the effectiveness of learning algorithms. The ensemble models and hyper-tuned multi-layer perceptron (MLP) with need-based hidden neuron layers are effective frameworks for data imputation. Addressing the issue of missing data is a complex and demanding task, and much remains to be explored in developing effective and precise methods for predicting and imputing missing values across different datasets. The study offers important perspectives on using algorithms in machine learning to predict and impute gaps in data in recently updated datasets. The findings indicate that finely tuned MLP classifiers notably improve prediction accuracy and dependability compared to models with a static or reduced number of neurons. Furthermore, the study highlights the promising potential of ensemble models within the error-correcting output code (ECOC) framework as an effective approach for this task. It also suggests future research directions to refine further and strengthen machine learning-based imputation methods regarding precision and stability. ECOC framework includes all kinds of MLP classifiers and regressors such as binary classifiers, multi-class classifiers, or regression models. MLP models predict complex relationships in modern datasets. Hugging Face, COSMIC, SKlearn, and Kaggle have relevant and updated datasets. The weighted average recognition (96%) shows that the proposed MLP-based stochastic learning strategies achieved better results.