Guidelines
Suggest edit

Retune the scatterplot preset for anomaly detection

For extreme-value detection, prefer task-specific preset tuning on scatterplots of quantitative data to improve anomaly-finding fidelity and address reuse of presets built for other reading tasks in preset-reuse workflows.

  • purpose:refine
  • basis:empirical
  • task:extreme
  • chart:scatter
  • data:quantitative
  • shape:outlier-rich
  • quality:fidelity:use
  • lever:encoding

advice

Scatterplot preset for anomaly finding

Retune the scatterplot preset when the job is to find anomalies. For example, use a task-tuned scatterplot preset instead of reusing a preset carried over from a prior correlation-reading study; in the evaluated comparison, the task-tuned design beat that reused preset on both anomaly accuracy and anomaly time.

reason

Why task-specific anomaly presets work

A scatterplot preset tuned for one reading task can emphasize the wrong tradeoffs for another. When the task shifts to finding singular points, a preset reused from correlation reading can slow search and reduce correct detections.

Mechanism: Task-specific scatterplot tuning aligns the visual design with anomaly finding instead of carrying forward a preset optimized or validated for a different reading task.

Evidence: In the collated extraction, the optimizer-generated scatterplot ranked above the prior-study scatterplot for both anomaly-detection accuracy and time, with significant differences in both metrics (Micallef et al., 2017; Zeng & Battle, 2023).

Notes: This guideline is supported only for the contrast between the task-tuned design and the reused prior-study preset.

context

Use when the preset is being reused across tasks

  • User Goal: Find anomalous points in a scatterplot.
  • Task: Find anomalies rather than estimate correlation.
  • Data: Quantitative point data where anomalous cases can be checked against known answers.
  • Chart Setting: A scatterplot preset is being selected or reused across tasks.
  • Success Criterion: Correct anomaly detection and completion time both matter.

exceptions

Do not use when package-default accuracy is the main criterion

Break it when: The real decision is between a task-tuned preset and a standard package preset, and anomaly accuracy matters more than speed. Why: In this comparison, the standard package presets outperformed the task-tuned design on anomaly-detection accuracy.

costs

Costs of retuning for anomaly detection

Sacrifice: You may give up some anomaly-detection accuracy relative to a standard package preset. Risk: You can overgeneralize and assume any task-tuned preset beats every existing preset. Mitigation: Benchmark the task-tuned preset against the package preset when accuracy is the main decision criterion.

mistakes

Common preset-reuse failure

Mistake: Reuse a scatterplot preset validated for correlation reading as the anomaly-detection preset. Why it fails: The reused preset was both slower and less accurate for anomaly finding than the task-tuned preset.

check

Check the anomaly preset against the reused preset

Failure Sign: Reviewers miss known anomalies or take longer than expected with the current reused preset. Quick Check: Run an A/B comparison between the reused preset and a task-tuned preset on a dataset with known anomalies, and record both correct detections and completion time. Stronger Test: Keep the task-tuned preset only if it improves both anomaly accuracy and anomaly time over the reused preset.

fix

Fix the reused preset problem

  • Create a separate scatterplot preset for anomaly detection instead of carrying over the preset from correlation reading.
  • Compare the anomaly preset directly against the reused preset on data with known anomalous points.
  • If anomaly accuracy is the top priority, benchmark the anomaly preset against a standard package preset before adopting it.

References

Micallef, L., Palmas, G., Oulasvirta, A., & Weinkauf, T. (2017). Towards Perceptual Optimization of the Visual Design of Scatterplots. IEEE Transactions on Visualization and Computer Graphics, 23(6), 1588–1599. https://doi.org/10.1109/TVCG.2017.2674978
Zeng, Z., & Battle, L. (2023). A Review and Collation of Graphical Perception Knowledge for Visualization Recommendation. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, 1–16. https://doi.org/10.1145/3544548.3581349