Fluctuation-based outlier detection.

Area of Science:

• Machine Learning
• Data Mining
• Anomaly Detection

Background:

• Outlier detection is crucial in machine learning, identifying rare data points deviating from the norm.
• Existing methods often rely on distance, density, or isolation, which can be computationally intensive.
• Outliers exhibit unique susceptibility to data fluctuation mechanisms.

Purpose of the Study:

• To propose a novel outlier detection method, Fluctuation-based Outlier Detection (FBOD).
• To achieve low linear time complexity for outlier detection.
• To introduce a method fundamentally different from existing distance, density, or isolation-based approaches.

Main Methods:

• Converts Euclidean datasets into graphs using random links.
• Propagates feature values through the graph structure.
• Identifies outliers by comparing an object's fluctuation to its neighbors' fluctuations.

Main Results:

• FBOD demonstrates superior performance compared to eight state-of-the-art algorithms on tabular and video datasets.
• FBOD achieves significantly lower execution times, using only 5% of the fastest competitor's time.
• Experimental validation confirms the effectiveness and efficiency of the fluctuation-based approach.

Conclusions:

• FBOD offers a computationally efficient and effective alternative for outlier detection.
• The fluctuation-based approach provides a fundamentally new perspective on identifying anomalies.
• FBOD shows strong potential for real-world applications requiring rapid and accurate outlier identification.