A model task with exactly two possible output labels — yes/no, spam/not-spam, AI/human — the simplest and most common form of supervised classification.
Binary classification is the task of assigning one of exactly two labels to each input. The model reads a piece of text (or any other input) and outputs either class 0 or class 1 — human-written or AI-generated, organic link or money link, well-formed query or keyword fragment. Everything reduces to a single decision boundary.
A binary classifier typically ends with a single output neuron that produces a probability between 0 and 1. A threshold — usually 0.5 — decides the label: above the threshold is class 1, below is class 0. The model is trained by minimising a loss function, typically binary cross-entropy, which penalises confident wrong predictions more heavily than uncertain ones.
At prediction time, the raw probability score is also useful in its own right: a score of 0.97 indicates high confidence, while 0.54 indicates the model is barely past the decision boundary. In production, many systems expose the score rather than the hard label so downstream logic can apply its own threshold.
Several DEJAN models are binary classifiers. The AI Content Detection model outputs organic vs. AI-generated. The Link Spam Algorithm outputs money-link vs. organic link. LinkBERT and LinkjeBERT assign O (not a link) or LINK to each token in a document. The Query Form Quality Classifier outputs well-formed vs. not well-formed. In each case the problem naturally has two states, and binary classification is the cleanest framing.
When a problem has more than two states — multiple intent categories, multiple sentiment levels — the task becomes multi-label classification or multi-class classification instead.
Binary classifiers are evaluated with precision, recall, F1 score, and overall accuracy. Which metric to optimise depends on the cost of each error type: a spam detector that misses spam (low recall) may be worse than one that over-flags legitimate content (low precision), or vice versa depending on the use case.
