Visual Neural Decomposition to Explain Multivariate Data Sets.

Area of Science:

• Data Science
• Machine Learning
• Data Visualization

Background:

• Analyzing relationships between variables in high-dimensional datasets is crucial for data analysts.
• Identifying specific input variable ranges that influence target variable values can be challenging with increasing complexity.
• Current methods for exploring these correlations can be time-consuming due to numerous variable combinations.

Purpose of the Study:

• To propose a novel, scalable approach for visualizing correlations between numerous input variables and a target output variable.
• To develop a visual model based on neural networks that facilitates guided exploration of data relationships.
• To enhance the interpretability of neural network models for data analysis.

Main Methods:

• Training a neural network to predict a target variable from input variables.
• Visualizing the internal workings of the trained neural network to understand data relationships.
• Introducing a new regularization term for backpropagation to promote visually interpretable representations.

Main Results:

• The proposed method effectively visualizes correlations in multi-dimensional data, scaling to hundreds of variables.
• The visual model allows for guided exploration, helping analysts identify key relationships.
• The new regularization technique improves the interpretability of neural network models.

Conclusions:

• The developed visual model offers an efficient and scalable solution for understanding complex variable relationships in large datasets.
• This approach aids data analysts in discovering and interpreting correlations that might otherwise be obscured.
• The method demonstrates utility across both artificial and real-world data applications.