Classification
Picking one of the bins that were already decided in advance
- Classification puts something incoming into one of the bins that were already decided in advance. The bin list exists first, then things go in.
- A model doesn't pick a bin outright. It scores every bin and hands back whichever one scored highest.
- A person decides the bins. Only after showing it plenty of labeled examples does it learn what each bin looks like.
- Even something outside the list still gets forced into one of the existing bins. Answering "none of these" has to be built in separately.
- If the answer is a bin, it's classification; if the answer is a number on a continuous scale, it's regression. That difference decides which method and which scorecard you use.
Contents
1The analogy
Behind a post office, regional bins stand in a row: one labeled New York, one Chicago, one Boston, and each piece of mail that comes in drops into one of them. The list of bins was fixed before any mail arrived, and a strange address today doesn't get a new bin made for it. Classification is picking one already-set bin like this.
What decides the pick is a few clues on the envelope: the first digits of the zip code, the neighborhood name, the kind of stamp. Clear clues settle the bin fast; smudged handwriting leaves you hesitating between two.
Mail with no address at all still can't sit outside every bin. It either goes into whichever bin seems closest, or there needs to be a separate "needs review" bin set up ahead of time.
2In detail
The bin list has to exist first
The very first step in classification is a person deciding the list of bins that count as valid answers. Whether you split photos into dog and cat, or dog, cat, and other, changes what model gets built entirely. Fix the bins at two, and every new photo can only come back as dog or cat.
Once the bins are set, you need plenty of examples for each one. Label hundreds of dog photos "dog" and hundreds of cat photos "cat" — these labeled examples are the training material.
How the bins get cut isn't a technical decision, it's a judgment call. Whether a post office splits bins by state or by district depends on delivery logistics, not on anything the model decides for you.
Score every bin, then pick the highest
A model doesn't reach straight for one bin. Internally, it scores every bin: New York 0.82, Chicago 0.11, Boston 0.07, something like that. Then it hands back whichever bin scored highest.
That score reveals a difference the plain answer hides. A split of 0.82 and 0.11, and a split of 0.42 and 0.39, both come back saying "New York" — but the second case is really close to a coin flip.
That's why real services usually only auto-process once a score clears some threshold, and route anything lower to a person for a look. Where you draw that line changes the balance between what gets missed and what gets misfiled.
What's on the envelope decides the grade
A classifier's grade depends far more on what clues it was given than on the model itself. A classifier fed only a zip code and one fed a zip code plus a neighborhood name won't perform the same. Feed it a pile of useless clues instead, and it gets more confused, not less.
The tricky part is when the wrong clue happens to work well. If, by chance, only New York-bound mail carried a blue sticker, a model can score great just by watching for the sticker, right up until the sticker policy changes, at which point it collapses.
A lopsided count of examples per bin causes trouble too. If nine out of ten pieces are New York-bound, just always guessing New York gets 90% right. The scorecard looks great, but the model hasn't actually learned anything.
When something outside every bin shows up
A classifier can only answer from the bins it knows. Feed a model trained only on dogs and cats a photo of a car, and it can't say "car" — it picks dog or cat instead, often with a fairly high score attached.
For a post office, that's like international mail arriving with only domestic bins available. It has to go somewhere, so it lands in whichever bin looks closest, and things keep going wrong from there.
That's why real classifiers usually get a separate "other" bin, or a mechanism that holds off answering when the score is too low. The moment you assume the bin list is complete is the most dangerous moment of all.
3More precisely
Classification maps an input onto one of a set of predetermined categories. Two possible answers makes it binary classification; three or more makes it multiclass. Training needs labeled examples, which puts it under supervised learning, a different starting point from clustering, which groups similar things without any labels at all. The scores a model produces are usually adjusted so they add up to 1 and read like probabilities, though without careful calibration those numbers can drift away from how often the model is actually right.
The analogy breaks down in places. Post office bins get decided by a solid fact, an address, but a classification model has no such solid ground. It only has statistical tendencies pulled from examples, so the same piece of mail could land in a different bin if the training data had been different. The bins themselves aren't always as cleanly separated as a post office either. A single message can be both a question and a complaint at once, which is why some setups let a model attach more than one bin instead of forcing a single pick.
4Try it yourself
5Common misconceptions
It's easy to think classification and clustering are the same task, but actually classification starts with the bins already fixed, while clustering groups similar things with no bins at all.
It's easy to think high accuracy means a good classifier, but actually a model can score high accuracy without learning anything, just by favoring whichever bin has the most examples.
It's easy to assume it'll say "I don't know" for something unfamiliar, but actually it picks one of the bins it knows, and attaches a confident-looking score to it.
7One-line summary
In shortClassification scores every predetermined bin and hands back the highest-scoring one, and how you draw up that bin list decides half the outcome.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02