Supervised learning trains a model with examples that include a known answer. The answer acts as a label. A photo may carry the label “cat.” A house record may include its sale price. The model learns a link between the input and that target.
Classification chooses a category
A classification model predicts from a set of labels. An email filter chooses spam or not spam. A quality system marks a product as acceptable or damaged. A plant tool identifies a species from a leaf image.
Some tasks allow several labels at once. A photo can contain a beach, sunset, and person. The model returns a score for each label rather than one final category.
Regression predicts a number
Regression handles a numeric target. A model may estimate delivery time, energy use, or apartment rent. It learns how input features relate to the number in past examples.
The word regression names the type of prediction. It does not mean the system moves backward. Both simple statistical models and neural networks can perform regression.
Training follows its errors
The model makes a prediction for a labeled example. A loss function measures the difference between that prediction and the known answer. The training process adjusts model values to reduce the loss across many examples.
A low training loss can mislead. The model may memorize the training set and fail on new cases. This problem is overfitting. Teams keep validation and test examples away from training. Those sets show how well the learned pattern travels to unseen data.
Labels need a clear definition
Human reviewers may disagree about whether a comment counts as harassment or a product photo shows damage. Written labeling rules reduce random differences. Difficult cases can receive review from more than one person.
Historical labels can carry old decisions and bias. A hiring model trained on past choices may copy unfair patterns. A medical dataset may represent one age group better than another. More rows do not fix a label that measures the wrong thing.
Data leakage creates false confidence
Leakage happens when training receives information that would not exist at prediction time. A hospital model cannot use a discharge code to predict a condition at admission. Duplicate records across training and test sets can also inflate scores.
The split should reflect future use. A time-based test can train on earlier months and test on later ones. A patient-level split keeps records from the same person on one side. These choices prevent easy shortcuts.
Supervised learning works well when the goal has a clear answer and enough representative examples. It struggles when labels cost too much, change often, or capture only part of the real task. The model learns the target it receives, not the goal someone meant to describe.