Finding the needle in the haystack-An interpretable sequential pattern mining method for classification problems.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Analyzing discrete sequential data like event logs presents challenges due to the high number of possible patterns.
• Identifying meaningful sequences and extracting actionable insights from complex data is difficult.

Purpose of the Study:

• To propose a novel feature selection algorithm integrating unsupervised sequential pattern mining with supervised machine learning.
• To develop an interpretable and efficient method for uncovering important sequential patterns in classification tasks.

Main Methods:

• The algorithm integrates unsupervised sequential pattern mining with supervised machine learning.
• It determines important sequential patterns during the mining process, avoiding post-hoc classification.
• A local, class-specific interestingness measure is introduced for inherent interpretability.

Main Results:

• The algorithm was evaluated on diverse datasets for churn prediction, malware analysis, and synthetic data.
• It achieved classification performance comparable to established feature selection algorithms.
• The method demonstrated reduced computational costs while maintaining interpretability.

Conclusions:

• The study presents a practical and efficient approach for sequential pattern discovery in classification.
• The algorithm offers an interpretable and efficient alternative to existing methods.
• This work advances sequential data analysis by combining interpretability with predictive performance.