Stream-based selective sampling is an active learning strategy tailored for scenarios where data arrives in a stream (e.g., real-time or sequential data). In this setting, each incoming unlabeled instance is evaluated by the model in turn, and the model (or query algorithm) decides whether to request a label for that instance or to ignore it and let it pass​.The decision is typically based on an informativeness criterion: for example, the model might query labels for data points on which it is uncertain or which would most reduce future error. If the model is currently performing well (confident) on the stream, it may label few instances; if performance degrades, it queries more aggressively​.This approach is efficient for continuously generated data, since it doesn’t require labeling everything—only the most useful samples are queried. Stream-based sampling thereby reduces labeling costs while enabling the model to adapt to changes in data distribution over time​.