Classification and Regression
Classification & Regression
What are classification and regression?
- Classification = predicting a category (type)
- Regression = predicting a number (value)
- The two core tasks of supervised learning
Requests usually arrive as a single sentence: 'Let's forecast demand.' Whether that means 'how many units to order next month' or 'which items are at risk of stockout' changes everything, from how you label the data to the report format. One line in the project brief shapes six months of work. Before choosing a model, write down what the person receiving the prediction will do with it, and you won't get lost at this fork.
Classification comes in binary (churn/retain), multiclass (5 defect types), and multilabel (several tags on one document). Most classification models output a probability between 0 and 1, so the 0.5 cutoff is a tuning variable, not a fixed value. If the sales team can only handle 30 cases a day, raise the cutoff to 0.72 and hand over just the top 30. Evaluation metrics differ too: classification uses precision, recall, F1, and AUC, while regression uses MAE, RMSE, and MAPE. A numeric target such as purchase amount can be turned into a classification problem by splitting it into 'KRW 3 million (≈ US$2,240) or more / less', but at that moment KRW 2.99 million (≈ US$2,240) and KRW 100,000 (≈ US$75) land in the same bucket.
Accuracy is where projects break most often. At a company with 3% monthly churn, a model that labels every customer 'retained' still scores 97% accuracy. The number looks great, yet it catches zero churning customers. Regression has the same trap. Even with an R² of 0.9, if a few peak-season orders miss by more than 40%, the purchasing manager will never look at that number again. Another common mistake is using a classification result directly in a monetary decision. '80% churn probability' does not mean 'KRW 8 million (≈ US$5,990) in losses.' If you need a monetary figure, either add a separate regression model or multiply the probability by each customer's contribution margin to get the value at risk.
John KimInstructor John Kim, author of "B2B Marketing Design" and CEO of Nanum Consulting
Sungil KimGlobal pitching expert from KAIST and Samsung Electronics who coached the winning teams at the UN Tourism and Select USA IR pitch competitions| Category | Classification | Regression |
|---|---|---|
| Prediction target | Category (type) | A continuous number (value) |
| Yes | Spam/not spam, dog/cat | House prices, tomorrow's sales |
| Question | 'Which one is it?' | 'How much?' |
These are the two basic prediction problems in statistics and machine learning, and they form the foundation of supervised learning.
Consider a hypothetical case. Company A, an industrial parts distributor with 260 employees and annual revenue of KRW 110 (≈ US$1) billion, manages 1,400 customer accounts. When the number of accounts that stopped buying last year grew to 190, management told the team to 'predict churn with AI,' and the data team started a three-month project.
- In the first meeting, the head of sales asked us to "pinpoint which accounts will leave," while the finance lead said "what I want to know is how much revenue we lose." Two different questions were mixed into one project.
- The data team lead separated the two questions. 'Will the account stop ordering within 3 months?' became a classification model, and 'expected order value over the next 6 months' became a separate regression model.
- The classification model's first report showed 94% accuracy, but it caught only 22 of the 190 accounts that actually churned. Lowering the decision threshold from 0.5 to 0.28 raised recall to 68%.
- The regression model's MAPE spiked to 31% on the top 50 large accounts. After separating out small accounts and shortening the training window to the most recent 18 months, MAPE came down to 14%.
- The sales division multiplied the two outputs to rank accounts by value at risk (churn probability × expected order value), then scheduled branch manager visits for the top 120.
- Six months later, quarterly account churn fell from 48 accounts to 29, and the annualized order value retained was about KRW 2.7 billion (≈ US$2,020,100). The head of sales said, "If we had only asked for a list of at-risk accounts, we'd have gotten half the answer."