Discrete Missing Data Imputation Using Multilayer Perceptron and Momentum Gradient Descent.

Area of Science:

• Data Mining
• Machine Learning

Background:

• Missing values in real-world data hinder industrial productivity.
• Discrete missing data imputation is less explored than continuous data imputation.
• Efficient imputation methods are crucial for data mining and productivity.

Purpose of the Study:

• To propose a novel discrete missing value imputation method using a multilayer perceptron (MLP).
• To enhance the convergence speed of the MLP using momentum gradient descent and prefilling strategies.
• To evaluate the effectiveness of the proposed method against existing imputation techniques.

Main Methods:

• Development of an improved multilayer perceptron (IMLP) model for discrete data.
• Application of momentum gradient descent algorithm for faster convergence.
• Utilizing prefilling strategies to optimize MLP training.
• Comparative analysis of IMLP against eight common imputation methods (mode, random, hot-deck, KNN, autoencoder, MLP).

Main Results:

• The improved MLP model (IMLP) demonstrates effective discrete missing value imputation across various scenarios.
• Experimental results show competitive or superior classification accuracy compared to other methods.
• The method's performance is validated under different missing mechanisms and proportions.

Conclusions:

• The proposed IMLP method offers a robust solution for discrete missing value imputation.
• This approach can significantly improve data quality and productivity in data mining applications.
• The study highlights the potential of advanced MLP techniques for handling complex data challenges.