Incorporate labeled outlier feedback into SONAR to improve guarantees

Investigate how to incorporate labeled outlier feedback into the design of the SONAR algorithm and determine whether such feedback can improve its theoretical guarantees, including control of Type I and Type II errors in streaming non-stationary environments.

Background

SONAR is introduced as an efficient, streaming-friendly, strongly convex variant of OCSVM that achieves provable bounds on Type I/II errors and adapts to non-stationary data via changepoint detection. The presented guarantees are derived without relying on labeled outlier data, reflecting the common scenario where outliers are scarce or adversarial.

The authors explicitly identify the incorporation of labeled outlier feedback as an open direction, suggesting potential improvements to theoretical guarantees if such labels can be effectively integrated into SONAR’s learning procedure.

References

While our Type II error guarantees are expressed through margin bounds, future work could derive direct Type II error bounds for various outlier models. It is also left open to study how labeled outlier feedback could be incorporated in SONAR's design principles and further improve guarantees.

— An Efficient Variant of One-Class SVM with Lifelong Online Learning Guarantees (2512.11052 - Suk et al., 11 Dec 2025) in Section 7, Conclusion/Discussion