Most real-world ML use cases reduce to a small number of recurring problem types. Recognizing which type you're facing is the first step toward picking the right algorithm and evaluation metric.
The Recurring Problem Types
| Problem Type | Description | Example | Typical Approach |
|---|---|---|---|
| Regression | Predict a continuous number | House price prediction | Linear Regression, Random Forest |
| Binary classification | Predict one of two classes | Spam / not spam, churn / no churn | Logistic Regression, SVM |
| Multi-class classification | Predict one of several classes | Categorizing a support ticket into one of 10 departments | Random Forest, Softmax Regression |
| Multi-label classification | Predict several labels at once, not mutually exclusive | Tagging a news article with multiple relevant topics | One-vs-rest classifiers |
| Clustering | Group similar data points with no predefined labels | Customer segmentation | K-Means |
| Anomaly detection | Identify data points that don't fit the normal pattern | Fraud detection, equipment failure prediction | Isolation Forest, DBSCAN, statistical thresholds |
| Recommendation | Predict which items a user will prefer | Product/movie recommendations | Collaborative filtering, matrix factorization |
| Ranking | Order a set of items by relevance | Search results ranking | Learning-to-rank models |
| Time series forecasting | Predict future values from a sequence of past values | Sales forecasting, demand planning | ARIMA, gradient boosting with lag features |
How to Map a Business Question to a Problem Type
# A simple decision checklist
if target_variable_is_continuous_number:
problem_type = "regression"
elif target_variable_has_exactly_2_categories:
problem_type = "binary classification"
elif target_variable_has_3_plus_mutually_exclusive_categories:
problem_type = "multi-class classification"
elif no_target_variable_exists_at_all:
problem_type = "clustering (or anomaly detection, or dimensionality reduction)"Practical Use Cases
Recognizing the problem type early narrows down candidate algorithms and, critically, the right evaluation metric — using accuracy on a rare-fraud (anomaly detection) problem, for example, is a common and misleading mistake; see Imbalanced Data.
Common Mistakes
- Forcing every problem into a classification/regression mold — some business questions (e.g. "what are our natural customer segments?") are genuinely unsupervised.
- Picking a metric before identifying the problem type — the right metric (RMSE vs F1-score vs precision@k) depends entirely on which type of problem it is.
Interview Relevance
Q: "A company wants to predict which of 5 shipping delay reasons applies to a late order. What kind of ML problem is this?" Multi-class classification (5 mutually exclusive categories) — not regression, and not binary classification.
Practice Question
Identify the problem type for each: (a) predicting a customer's lifetime value in rupees, (b) tagging a support ticket with all applicable issue categories (can be more than one), (c) grouping retail stores by sales pattern with no predefined categories.
