A connectionist approach to generating oblique decision trees.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Neural networks offer high accuracy but lack interpretability.
• Decision trees provide interpretability but can be large and less accurate.
• Oblique decision trees offer a balance, using linear combinations of attributes.

Purpose of the Study:

• To develop a novel algorithm for generating oblique decision trees.
• To leverage the strengths of both neural networks and symbolic decision trees.
• To improve accuracy and reduce tree size in pattern classification.

Main Methods:

• A hybrid approach combining connectionist (neural network) and symbolic components.
• Construction and pruning of a three-layer feedforward neural network.
• Building a decision tree from the pruned network's hidden unit activations, expressed using original input attributes.

Main Results:

• Generated oblique decision trees maintain high accuracy comparable to neural networks.
• The algorithm produces smaller decision trees than univariate approaches.
• The oblique decision trees outperform other methods in both accuracy and size.

Conclusions:

• The novel algorithm successfully integrates neural network accuracy with decision tree interpretability.
• This method offers a superior alternative for pattern classification tasks.
• The resulting oblique decision trees are more efficient and effective than existing methods.