The classification threshold is a crucial parameter in ML models, particularly binary classifiers, as it dictates the boundary for assigning predicted outputs to specific categories. Typically, for models like logistic regression or neural networks used in classification tasks, the output is a probability value between 0 and 1. The classification threshold, often set at a default of 0.5, determines the cutoff where predictions are labelled as one class if above the threshold and another class if below. Adjusting this threshold affects the model's sensitivity and specificity, impacting precision and recall, which is especially significant in applications requiring a balance between false positives and false negatives, such as medical diagnosis or fraud detection. The choice of threshold can thus be tailored based on the cost or consequences of different types of errors in classification.
The term "classification threshold" has been used informally in statistics for decades with the advent of classification problems, but gained specific terminology and widespread recognition in the sphere of ML around the early 2000s, as model interpretability and tuning became critical in practical applications of AI.
Key contributors to the understanding and use of classification thresholds are not easily pinpointed to specific individuals, as it evolved concurrently within the broader development of statistical and ML methodologies. However, foundational work in statistical decision theory and the rise of binary classifiers have been influenced by academic pioneers in statistics and applied math, contributing to this concept's maturation.