OM4AnI: A Novel Overlap Measure for Anomaly Identification in Multi-Class Scatterplots.

Area of Science:

• Data Visualization
• Machine Learning
• Human-Computer Interaction

Background:

• Scatterplots are crucial for anomaly detection in multi-class datasets.
• Large datasets diminish scatterplot effectiveness due to resolution limits.
• Existing methods lack robust anomaly identification assessment.

Purpose of the Study:

• Introduce OM4AnI (Overlap Measure for Anomaly Identification), a novel Visual Quality Measure (VQM).
• Quantify scatterplot overlap for effective anomaly identification in multi-class settings.
• Provide a metric to estimate anomaly visibility and optimize scatterplots.

Main Methods:

• Compute anomaly index based on data point position relative to class clusters.
• Discretize scatterplots into pixel-level grids, calculating coverage per pixel.
• Integrate anomaly index and visual features (shape, size, order) for a quality score.

Main Results:

• OM4AnI outperforms six baseline methods in efficiency, effectiveness, and sensitivity.
• OM4AnI shows more monotonic trends against ground truth compared to baselines.
• OM4AnI demonstrates greater sensitivity to rendering order than baseline methods.

Conclusions:

• OM4AnI effectively informs users about scatterplot anomaly identification support.
• OM4AnI shows strong potential as an evaluation metric for scatterplots.
• OM4AnI can optimize scatterplots via automatic visual parameter adjustment.