Learning Interpretable Rules for Scalable Data Representation and Classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Rule-based models like decision trees offer high interpretability but are difficult to optimize on large datasets.
• Existing methods to improve performance, such as ensemble methods or fuzzy rules, often compromise model interpretability.
• Scalability and interpretability are often mutually exclusive in rule-based machine learning models.

Purpose of the Study:

• To propose a novel classifier, the Rule-based Representation Learner (RRL), that achieves both good scalability and interpretability.
• To develop an effective training method for non-differentiable rule-based models.
• To enable end-to-end discretization of continuous features within a rule-based framework.

Main Methods:

• Developed the Rule-based Representation Learner (RRL), a classifier that learns interpretable non-fuzzy rules.
• Proposed a novel training method, Gradient Grafting, to optimize discrete models using gradient descent by projecting them into a continuous space.
• Designed novel logical activation functions to enhance RRL's scalability and enable end-to-end feature discretization.

Main Results:

• RRL demonstrated superior performance compared to competitive interpretable approaches across ten small and four large datasets.
• The proposed Gradient Grafting method effectively trained the non-differentiable RRL model.
• The novel logical activation functions improved the scalability of RRL and allowed for end-to-end feature discretization.

Conclusions:

• RRL offers a scalable and interpretable solution for classification tasks, overcoming limitations of traditional rule-based models.
• The Gradient Grafting training method and logical activation functions are key innovations enabling RRL's effectiveness.
• RRL provides flexibility to balance classification accuracy and model complexity for diverse applications.