A method for explaining individual predictions in neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• High-performance machine learning models, particularly artificial neural networks, often function as "black boxes," limiting the explainability of their predictions.
• While explainability methods exist for image classification, they are less effective for tabular data in classification and regression tasks.
• There is a growing need for transparent and interpretable prediction models in various scientific domains.

Purpose of the Study:

• To develop a method for explaining individual prediction results from neural network models.
• To address the challenge of interpreting black box models, especially for tabular data.
• To provide a clear rationale for the predictions made by neural networks.

Main Methods:

• The proposed method leverages the fundamental principle that neural network output is a weighted sum of inputs.
• It calculates the contribution of each input value to the final output by analyzing the network weights and input values.
• The contribution is quantified using the formula (input value * weight value) / weighted sum, tracking influence through the network layers.

Main Results:

• The developed method successfully demystifies neural networks, making them non-black box models.
• Predictions from neural networks are effectively explained, regardless of network architecture complexity (hidden layers, nodes).
• The method is applicable to both classification and regression tasks and is available as an easy-to-use Python library.

Conclusions:

• The proposed technique enhances the interpretability of neural network predictions for tabular data.
• This approach provides a transparent and reliable way to understand model behavior.
• The Python library facilitates the practical application of this explainability method in machine learning workflows.