A Few Games Marked, Most Not

Opening Scene

The team’s marquee games — the ones broadcast on television — get full, careful stat tracking: every shot marked, every possession logged. But most of the season is scrimmages and lower-profile games, recorded on a sideline camera with nobody sitting there marking anything. By the end of the year, the coaching staff has a small pile of richly labeled marquee-game data and a much larger pile of raw, unmarked footage from everything else.

Throwing away the unmarked footage feels wasteful — there’s real information in it, just not neatly labeled. Using only the marquee games feels limiting — there’s too little of it to learn much. The natural instinct is to find some way to use both piles together, and that instinct is exactly what semi-supervised learning formalizes.

In Plain English

Semi-supervised learning trains a model using a small amount of labeled data together with a much larger amount of unlabeled data. The labeled portion anchors the model to real, known answers; the unlabeled portion — despite having no answer key — still carries useful structure about how the data is generally shaped, and a model can learn from that structure too. It sits deliberately between Article 1’s fully supervised approach and Article 2’s fully unsupervised one.

The Old Way

Before this had a formal name, people intuitively blended scarce certainty with abundant rough observation:

A scout with detailed reports on a few star players and only rough impressions of everyone else, still forming a reasonably informed overall read on the league.
A doctor with confirmed diagnoses for a handful of patients and vague symptom notes for many more, using both to sharpen general intuition about a condition.
A teacher with fully graded work from a few students and quick glances at everyone else’s, still forming a broad sense of how the whole class is doing.

None of these approaches were rigorous by modern standards, but the underlying instinct — scarce certainty plus abundant rough signal beats either alone — was sound.

What’s Changing (and Why AI Is the Reason)

Modern semi-supervised techniques can meaningfully close the gap with fully labeled training, using the unlabeled majority to make the most of a small labeled anchor set — a genuine technical advance over simply discarding unlabeled data.
This matters more now because unlabeled data has become extremely cheap and abundant, while labeling remains genuinely expensive. Semi-supervised approaches let teams get real value from that imbalance instead of being stuck choosing between “not enough labeled data” and “too expensive to label it all.”
Semi-supervised methods increasingly work alongside self-supervised techniques, covered directly in Article 11, blurring the line between “small labeled anchor plus raw data” and “learning structure before any labels exist at all.”

The Metaphor, Fully Extended

Team Footage	Semi-Supervised Learning Concept
Fully stat-tracked marquee games	The small labeled portion of the dataset
Raw sideline footage from the rest of the season	The much larger unlabeled portion
The general shape of how players move across all footage	Structure a model can learn from unlabeled data alone
Using the marquee games to anchor what “good play” looks like	The labeled data grounding the model in known answers
Applying that grounded understanding to the rest of the footage	Extending learned patterns from labeled to unlabeled data
A scouting report built from both piles together	A semi-supervised model’s final output

For Beginners: What to Actually Do

Recognize when you’re sitting on this exact situation — a small labeled set and a much larger unlabeled one — it’s extremely common and often underused.
Don’t assume unlabeled data is worthless just because it lacks an answer key; it still carries real, usable structure.
Start by understanding how much the labeled anchor set alone can achieve, so you have a clear baseline to measure any semi-supervised improvement against.

For Practitioners and Leaders: The Deeper Layer

Semi-supervised approaches can meaningfully reduce labeling costs, but they don’t eliminate the need for a reasonably representative labeled anchor set — garbage anchors still produce garbage results.
Be cautious about how much the unlabeled data actually resembles the labeled data’s real-world context; a mismatch there can quietly mislead the model rather than help it.
This approach is often the most practical bridge for teams with a genuine labeling budget constraint but abundant raw data — worth considering before assuming a fully supervised approach is the only option.

Quick Recap

Semi-supervised learning combines a small labeled dataset with a much larger unlabeled one to train a model.
It mirrors a common human instinct — blending scarce certainty with abundant rough observation — done more rigorously.
It’s especially valuable now because unlabeled data is cheap and abundant while labeling remains expensive.
A representative labeled anchor set is still essential; semi-supervised learning stretches labeled data further, it doesn’t replace the need for it.

Where This Fits in the Series

Article 9 covered the problem of too little labeled data; this article covered one practical answer — blending it with abundant unlabeled data. Article 11 pushes further into the same territory, looking at what a model can learn before any labels exist at all.