A company wants an ML solution to detect anomalies in sensor output but has no labeled examples. Which modeling approach is appropriate for unsupervised anomaly detection in this case?

Autoencoders are a suitable unsupervised learning technique for anomaly detection when labeled data is unavailable. They learn a compressed representation (encoding) of normal data and then reconstruct it. Anomalies, being different from the normal patterns the autoencoder was trained on, will have higher reconstruction errors, making them detectable. Linear regression, classification, and decision trees are typically supervised learning methods that require labeled data for training, making them inappropriate for this scenario. While some variations exist, their primary use cases involve predicting a target variable from known examples.