A multi-layer perceptron neural network for varied conditional attributes in tabular dispersed data

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Dispersed data sources with varying objects and attributes pose challenges for global model construction.
• Existing methods for multi-model classification often struggle with heterogeneous and incomplete datasets.

Purpose of the Study:

• To introduce a novel approach for constructing a global model from dispersed data using multilayer perceptron (MLP) neural networks.
• To evaluate the performance of the proposed method against existing homogeneous and heterogeneous multi-model classifiers.

Main Methods:

• Development of local models from local data tables, including imputed artificial objects.
• Aggregation of local models using weighted techniques.
• Retraining of the global model with global objects.

Main Results:

• The proposed approach consistently outperforms existing methods in classification accuracy, balanced accuracy, F1-score, and precision.
• Achieved an average classification accuracy improvement of 15% and balanced accuracy enhancement of 12% over baseline methods.
• Demonstrated superior performance compared to traditional ensemble classifiers and homogeneous MLP ensembles.

Conclusions:

• The novel method offers a robust and adaptable solution for global model construction from dispersed data.
• The resulting single global model is easier to use and interpret, showing promise for practical applications in healthcare and smart agriculture.