Supervised · Unsupervised Learning
Supervised / Unsupervised Learning
What is supervised learning?
- Supervised = learning from data with answers attached
- Unsupervised = finding patterns without answers
- Two approaches to machine learning
Which one you choose is not a matter of algorithm preference. It decides your project schedule and labor costs first. The moment you commit to supervised learning, the question of 'who labels how many records, and by what standard' takes up half the total effort. With unsupervised learning the labeling cost disappears, but the job of judging whether the output is usable falls entirely on the business team. So the question to ask in the first meeting is not which model to use but whether your database has a column that can serve as the answer.
Supervised learning splits further into classification (churn/retain, defective/good) and regression (next month’s demand, expected unit price). A classification model worth using in practice usually needs hundreds to thousands of labels per class, and on imbalanced data where churning customers are only 3% of the total, you can end up with a model that scores 97% accuracy while answering 'all normal' every time. Unsupervised learning is widely used for dimensionality reduction and anomaly detection, not just clustering. When you have only a few hundred labels, semi-supervised learning (mixing the two) and self-supervised learning (generating questions and answers from the data itself) fill the gap. The boundary with reinforcement learning is that, instead of answers, the model receives a reward for doing well.
Projects most often go wrong at the decision that 'we have no labels, so let's go unsupervised.' Clustering output is just numbered groups. The basis for calling Group 3 the 'price-sensitive segment' has to come from people, not the model, and changing the number of clusters from 4 to 6 makes last month’s report segments vanish entirely. Supervised learning has the opposite problem: the bias in past labels is copied as is. If you train on records where sales reps flagged accounts by gut feeling as 'this one looks like it's about to leave,' the model is not predicting churn. It is reproducing one salesperson’s judgment habits.
John KimInstructor John Kim, author of "B2B Marketing Design" and CEO of Nanum Consulting
Jaryong KooMarketing PhD and leading authority on AI, data analytics, and personal branding| Category | Supervised learning | Unsupervised learning |
|---|---|---|
| Data | With answers (labels) | No labels |
| Purpose | Prediction (classification/regression) | Discovering patterns and structure |
| Yes | Churn prediction | Customer clustering |
A foundational distinction from the early days of machine learning, dividing learning approaches by whether the data has answers (labels).
Consider a hypothetical case. Company A, an industrial consumables distributor with 240 employees and KRW 110 billion (≈ US$82,298,400) in annual revenue, manages 8,400 regular-delivery accounts. For two years its monthly churn rate has stuck at 2.8%, and the sales division kept repeating only that it did not know which accounts were at risk. A three-person data team took on the problem as a four-month project.
- When Kim, a senior manager on the data team, pulled the last 24 months of transactions from the ERP, only 1,900 of 8,400 customer accounts had a recorded churn outcome. Fewer than a quarter of the answer column was filled in.
- The team brought together the five sales team leads, agreed on one rule ('no orders for 3 consecutive months = churn'), and spent two weeks retroactively applying 6,200 labels to historical data.
- After the labels were filled in, we ran a supervised churn-prediction model, and 214 of the top 300 at-risk accounts were actually on their way out, a 71% hit rate.
- When the same data was clustered without labels, using order frequency, number of items, and average unit price, 6 groups emerged, and the head of sales said, 'We have never managed group 4 separately.'
- We assigned one dedicated CS rep to Group 4 (1,120 small-volume, multi-item accounts) and scheduled visits for the 300 at-risk accounts.
- Six months later, monthly churn fell from 2.8% to 1.6%, and annualized churned revenue dropped from about KRW 3.4 billion (≈ US$2,543,800) to KRW 1.9 billion (≈ US$1,421,500).