TADILOF: Time Aware Density-Based Incremental Local Outlier Detection in Data Streams.

Area of Science:

• Data Mining and Machine Learning
• Time Series Analysis
• Internet of Things (IoT) Data Analytics

Background:

• Outlier detection in data streams is essential for data mining but challenged by increasing data volumes from IoT.
• Existing density-based Local Outlier Factor (LOF) methods fail to adapt to evolving data patterns and new clusters over time.

Purpose of the Study:

• To develop a novel algorithm for outlier detection in streaming data that addresses temporal variations.
• To introduce an 'approximate LOF' score estimation method using historical data.

Main Methods:

• Proposed a new algorithm: time-aware density-based incremental local outlier detection (TADILOF).
• Developed an 'approximate LOF' score estimation by leveraging historical data and removing outdated information.
• Evaluated TADILOF's performance against state-of-the-art methods using standard metrics.

Main Results:

• TADILOF demonstrated superior performance in terms of Area Under the Curve (AUC) compared to existing methods.
• The algorithm achieved comparable execution times to current state-of-the-art techniques.
• An application to an air-quality monitoring system showcased the practical utility of the proposed scheme.

Conclusions:

• TADILOF effectively handles temporal dynamics in data streams, outperforming existing outlier detection algorithms.
• The 'approximate LOF' method provides an efficient way to estimate outlier scores in dynamic environments.
• The developed approach is applicable to real-world IoT applications, such as environmental monitoring.