Exploration of Outliers in If-Then Rule-Based Knowledge Bases.

Area of Science:

• Data Science
• Artificial Intelligence
• Knowledge Representation

Background:

• Rule-based knowledge bases are complex data structures requiring robust analysis methods.
• Identifying unusual or outlier rules is crucial for domain experts to ensure knowledge base accuracy and utility.
• Existing methods often lack specialized techniques for outlier detection within rule sets.

Purpose of the Study:

• To analyze the effectiveness of four outlier detection algorithms (Local Outlier Factor, Connectivity-based Outlier Factor, K-MEANS, and SMALLCLUSTERS) applied to rule-based knowledge bases.
• To evaluate the impact of outlier removal on the quality of rule clusters using various quality indices.
• To determine which outlier detection methods yield the most significant improvements in rule cluster quality.

Main Methods:

• Application of clustering algorithms to rules within domain knowledge bases.
• Utilizing outlier detection algorithms (LOF, COF, K-MEANS, SMALLCLUSTERS) to identify rare and unusual rules.
• Quantifying outlier rules at percentages of 1%, 5%, and 10%, and within small clusters (<=5%).
• Assessing rule cluster quality using seven distinct indices before and after outlier removal.
• Conducting experiments on six diverse knowledge bases.

Main Results:

• Local Outlier Factor (LOF) and Connectivity-based Outlier Factor (COF) algorithms demonstrated superior performance in identifying unusual rules.
• The removal of detected outliers frequently led to an improvement in the quality of rule clusters across multiple indices.
• LOF and COF were most consistently associated with enhanced rule cluster quality in the experimental evaluations.
• The effectiveness of outlier detection varied depending on the specific knowledge base and the chosen quality indices.

Conclusions:

• Outlier detection algorithms, particularly LOF and COF, are valuable tools for analyzing and refining rule-based knowledge bases.
• Identifying and potentially removing unusual rules can significantly enhance the overall quality and interpretability of knowledge bases.
• This research highlights the importance of specialized outlier mining techniques for complex rule sets in AI and data science.