Cluster-Based Improved Isolation Forest.

Area of Science:

• Data Mining
• Machine Learning
• Anomaly Detection

Background:

• Outlier detection is crucial in data mining, but existing methods like Isolation Forest face challenges with unstable results and low efficiency due to random feature partitioning.
• The random feature selection in standard Isolation Forest can lead to suboptimal partitioning and reduced performance in identifying anomalies.

Purpose of the Study:

• To propose a novel algorithm, Cluster-based Improved Isolation Forest (CIIF), to address the limitations of the standard Isolation Forest for outlier detection.
• To enhance the stability and efficiency of outlier detection by combining clustering techniques with the Isolation Forest framework.

Main Methods:

• The CIIF algorithm employs k-means clustering to group data points.
• A selection matrix is constructed based on cluster results to guide feature selection within the Isolation Forest.
• Multiple isolation trees are built, and outlier scores are calculated using average search lengths, with top-scoring objects identified as outliers.

Main Results:

• Comparative experiments on eleven real-world datasets demonstrated that CIIF significantly outperforms six other algorithms.
• The CIIF algorithm achieved an average improvement of 7% in Area Under the Curve (AUC) compared to the standard Isolation Forest algorithm.
• The proposed method shows enhanced performance in terms of detection accuracy and stability.

Conclusions:

• The Cluster-based Improved Isolation Forest (CIIF) algorithm offers a more robust and efficient approach to outlier detection.
• Integrating k-means clustering with Isolation Forest effectively improves anomaly detection capabilities.
• CIIF represents a significant advancement in the field of data mining for outlier analysis.