[D] Is this considered unsupervised or semi-supervised learning in anomaly detection?

A seemingly niche question from a Reddit user working on anomaly detection has gone viral, exposing a fundamental and often ambiguous classification issue in machine learning. The user describes a common real-world setup: their model is trained exclusively on "normal" or benign data to learn a representation of standard behavior, a process that is inherently unsupervised or one-class learning. No anomaly labels are used during this training phase. The core of the debate arises during evaluation, where they use a separate, labeled validation set to select a decision threshold that maximizes the F1-score, introducing a form of supervision after the fact.

This practical methodology sits in a grey area between textbook definitions. Purists might argue the initial training is unsupervised, making the entire pipeline "unsupervised anomaly detection with threshold calibration." Others contend that the use of labels to optimize the final operational boundary qualifies the overall process as semi-supervised. The discussion highlights a critical gap between academic taxonomy and applied AI engineering, where hybrid approaches are the norm for tasks like fraud detection, network intrusion, and industrial defect inspection. The consensus leaning from the thread suggests "semi-supervised" is often the safer, more accurate term for papers, as it acknowledges the use of some labels to guide the model's final decision-making capability, preventing claims of a purely label-free system.