TLDR: This research introduces a hybrid AI method that uses association rules, mined from data, to improve tree-based classifiers. It rectifies models to boost predictive accuracy and extends domain theories to generate significantly shorter and more general abductive explanations, making AI decisions more understandable.
In the evolving landscape of Artificial Intelligence, the quest for systems that are not only powerful but also understandable and trustworthy is paramount. A recent research paper by Gilles Audemard, Sylvie Coste-Marquis, Pierre Marquis, Mehdi Sabiri, and Nicolas Szczepanski introduces a novel approach that combines the strengths of data-driven and knowledge-driven AI to achieve this balance. Their work, titled “Leveraging Association Rules for Better Predictions and Better Explanations,” explores how association rules can significantly enhance the performance and interpretability of tree-based machine learning models.
The core idea revolves around Hybrid AI, which seeks to integrate symbolic information (knowledge) with data-driven techniques. The researchers specifically focus on classification tasks using tree-based models like decision trees and random forests. They propose a method where association rules, derived from the data itself, are used in two key ways: first, to improve the accuracy of predictions, and second, to generate more insightful and concise explanations for those predictions.
The process begins by learning a random forest (or a single decision tree) from a given dataset. This data is then transformed into a binarized format, representing Boolean conditions encountered in the forest. A data mining algorithm is subsequently employed to extract association rules, including those with negations. These rules are carefully selected, prioritizing those with 100% confidence and non-null support, ensuring they are highly reliable and not contradicted by existing data.
A crucial aspect of their approach is the “rectification” operation. Classification rules (rules that predict a class) are used to modify the tree-based model. This rectification ensures that if an association rule classifies an instance, the updated model will adhere to that rule’s classification. For instances not covered by these rules, the model retains its original classification. This iterative rectification process, applied rule by rule, is designed to be efficient and maintain the computational intelligibility of the model, meaning it can still support various Explainable AI (XAI) queries effectively.
Beyond improving predictive performance, the paper highlights the significant impact on explanation generation. The researchers extend the concept of “majoritary reasons” – a type of abductive explanation – by incorporating a strengthened domain theory. This extended theory includes not only the logical connections between Boolean conditions but also the non-classification association rules mined from the data. By leveraging these additional rules, the system can generate more general and often much shorter abductive explanations. These shorter explanations are typically easier for humans to understand, addressing a critical need for compactness in XAI.
Experimental results presented in the paper demonstrate tangible benefits. For a majority of the datasets tested, rectifying random forests and decision trees with classification rules led to an increase in predictive performance, sometimes exceeding 10%. More strikingly, the reduction in the size of abductive explanations was often substantial, in some cases over 96%, and effective for a large proportion of instances. This reduction in explanation size is a direct outcome of incorporating the mined association rules into the domain theory, allowing for more concise and general explanations.
Also Read:
- Unlocking Complex Simulations: How AI Makes Expensive Models Faster and More Transparent
- Enhancing AI’s Adaptability: A Human-Inspired Approach to Learning Across Diverse Environments
The authors emphasize that their approach maintains user control, allowing users to filter and select the association rules they trust. This ensures that the integration of knowledge is guided by human expertise and confidence. The computational overhead for both rectification and explanation generation remained reasonable across various datasets, confirming the practical applicability of their method. This research offers a promising direction for developing AI systems that are both high-performing and transparent, fostering greater trust and understanding in their decisions. You can read the full paper here: Leveraging Association Rules for Better Predictions and Better Explanations.
