Pick ML by Label Budget, Not by Algorithm

The question is not which machine learning type is best. It is what evidence you have and what you can afford to create.

That reframing matters because most method debates happen before anyone prices the evidence. A team picks a supervised model, then discovers that the labels it needs require a domain expert who is booked for six months. Another team picks reinforcement learning because it sounds advanced, then discovers it has no safe simulator. The algorithm was never the hard part. The evidence was.

Every method is a label strategy

Supervised learning needs a target. Someone has to decide what the correct answer is for a set of examples, and someone has to produce enough of them to train and test. The cost is not just the labeling tool. It is the expert time, the adjudication of disagreements, and the maintenance of the labels as the world changes.

Unsupervised learning does not need labels. It needs a structure worth finding and a person who can interpret the result. A clustering model will always produce clusters. The question is whether anyone can act on them.

Reinforcement learning needs a simulator and a reward. The simulator has to be safe enough to explore in, and the reward has to encode what you want. A reward that is easy to measure is often not the reward you care about.

Self-supervised learning needs raw data and compute. It creates its own training signal from the structure of the data, which is why it works well for language and images. It still needs a small supervised step to connect the learned representation to your decision.

Semi-supervised learning needs a little of both. A small labeled set anchors a larger unlabeled set. It is useful when labels are expensive and the unlabeled data is representative.

The method follows the evidence. The evidence is a budget line.

A decision sequence

Start with the decision, not the data. What output does the system need to produce, and who checks it?

If you have examples of the decision with known answers, start supervised. If the positive cases are rare, consider anomaly detection or a semi-supervised approach. If the labels are expensive but the unlabeled data is plentiful, use a small labeled set and a self-supervised representation.

If you do not have labels, ask whether you need structure or a decision. Structure is an unsupervised problem: clustering, segmentation, dimensionality reduction. A decision without labels needs a simulator, a proxy, or a rules-based system until you can create one.

If you cannot measure the outcome, do not use machine learning. Use a rule. A rule is auditable, cheap, and easy to change. A model trained on an unmeasurable outcome is a guess with a confidence score.

The feedback loop matters more than the method

A model is only as good as the loop that corrects it. Who sees an error? How quickly does that error become a labeled example? Does the loop include the people who own the decision, or only the data team?

Supervised models degrade when the world changes and the labels do not. Unsupervised models produce output no one uses when there is no reviewer. Reinforcement learning learns the wrong behavior when the reward is wrong or the simulator is unrealistic.

The loop is an organizational design. A team that can route an error to a reviewer and back into the training set in a week will outperform a team with a better model and a slower loop.

Two failure modes

The first is label theater. A team buys labels that are easy to produce rather than labels that reflect the decision. The model learns a proxy, the metric improves, and the decision does not.

The second is method shopping. A team chooses a method because it is fashionable, then shapes the problem to fit it. The result is a complicated system that solves a problem no one had.

Both failures are avoided by the same discipline: write down the decision, the evidence, and the loop before choosing the algorithm.

What to tell stakeholders

The method is a consequence of the evidence. If the evidence is not affordable, change the decision or the scope, not the algorithm. A narrower decision with clean labels is a better first project than a broad decision with a proxy.

The algorithm is the last decision, not the first. Start with the evidence you can afford, and the method usually chooses itself.