An article can carry any of 20 topic tags at once - how should you frame that target?

## Naming the task correctly Four target shapes get confused with each other: - **Binary** - one label, two outcomes. Fraud or not. - **Multiclass** - one label, more than two outcomes, **mutually exclusive**. Exactly one is true per example: a photo shows a cat *or* a dog *or* a horse. - **Multilabel** - many labels, each independently true or false. An article may be tagged `politics`, `election` and `europe` simultaneously - or nothing at all. - **Multi-output** - several separate targets predicted together, each of which may itself be binary, multiclass or continuous. The test is one sentence: **can two of these labels be true of the same example at the same time?** If yes, it is multilabel. Twenty news topics obviously can co-occur, so the tagging task is multilabel. ## What the framing changes about the output A mutually exclusive framing produces one probability vector *across* the classes that sums to one. Raising the probability of one class necessarily lowers another - they compete. That competition is exactly right when only one answer can be true, and exactly wrong here: evidence for `election` should not have to be taken away from `politics`. A multilabel framing produces one probability *per tag*, each in the range 0 to 1, with no summation constraint. Twenty tags means twenty independent yes/no questions. Consequences follow immediately: - **Output cardinality is free.** Zero tags and eight tags are both representable. You must decide as a product question whether an article is allowed to end up with no tags, and what to do when that happens - typically fall back to the highest-scoring tag, or leave it untagged for a human. - **Each tag has its own base rate.** `politics` may apply to 30% of articles and `obituaries` to 0.4%. A single shared decision rule across all 20 will over-fire on the rare tags or under-fire on the common ones, so the cut point is per tag. - **Each tag has its own difficulty.** Some are lexically obvious, some are judgement calls. Per-tag reporting is the only way to see this; a single aggregate number hides which tags are broken. ## Framings that do not work, and why **Label powerset.** Treat every observed combination of tags as its own class. It is a legitimate trick for a handful of labels because it captures correlations exactly, but with 20 tags there are 2^20 - over a million - possible combinations. Almost all of them never appear in training, and the model can only ever predict a combination it has seen. Most observed combinations will have a handful of examples each. It does not scale. **Forcing one primary tag.** Pick the single most important tag per article and run a mutually exclusive model. This is sometimes imposed by a downstream system that has one slot. The costs are real: annotators disagree about which tag is *primary*, so the labels themselves become noisy and inconsistent; the model learns an arbitrary priority ordering rather than topicality; and an article genuinely about both politics and the economy is recorded as an error whichever tag you predict. If a single slot is truly required, the cleaner design is to model all tags honestly and pick the highest-scoring one at serving time - the choice of primary tag then becomes a display rule you can change without retraining. ## Correlation between tags Independent per-tag decisions assume the tags are conditionally independent given the features, which they are not - `election` almost implies `politics`. Two common repairs: chain the decisions, feeding earlier tags' predictions as inputs to later ones, or post-process with the label hierarchy (if `election` fires, force `politics`). Both add complexity, so start with independent decisions and only add structure if the errors show you inconsistent tag sets. ## The subtlety worth raising: missing versus negative In multilabel data an absent tag is ambiguous. Did the annotator judge that `europe` does not apply, or did they stop after the first two tags they thought of? Independent binary framing treats every absent tag as a hard negative, which teaches the model that under-tagged articles genuinely lack those topics. Where labelling was partial, that pushes recall down systematically. Knowing how the labels were produced is part of framing the target - a strong answer says so. ## Hierarchy If some tags nest inside others, the target has structure the flat framing ignores. You can flatten and rely on the model to learn consistency, enforce parents at serving time, or model each level separately. Whichever you pick, decide it deliberately rather than discovering inconsistent parent/child predictions in production.