Skip to main content

Making Predictions

Once a model has been trained using .fit(), it is ready for Inference. In Scikit-Learn, this is handled by a consistent set of methods that allow you to generate outcomes for new data.

1. Point Predictions with .predict()​

The most common way to get an answer from your model is the .predict() method. It returns a single value for each input sample.

  • In Regression: Returns the predicted continuous value (e.g., 250,000250,000).
  • In Classification: Returns the predicted class label (e.g., "Spam").
from sklearn.ensemble import RandomForestClassifier

# Assuming model is already trained
# X_new contains unseen samples
predictions = model.predict(X_new)

print(f"Predicted labels: {predictions}")

2. Predicting Probabilities with .predict_proba()​

In many classification tasks, knowing the label isn't enough; you need to know how confident the model is. Most Scikit-Learn classifiers provide the .predict_proba() method.

It returns an array of shape (n_samples, n_classes), where each value represents the probability of a sample belonging to a specific class.

# Returns probabilities for [Class 0, Class 1]
probs = model.predict_proba(X_new)

# Example output: [0.1, 0.9] means 90% confidence it is Class 1
print(f"Confidence levels: {probs}")

Why use probabilities?​

  1. Risk Management: In medical diagnosis, you might only take action if the probability is .
  2. Threshold Tuning: By default, .predict() uses a threshold. You can manually set a higher threshold to reduce False Positives.

3. Decision Functions​

Some models, like SVM (Support Vector Machines) or Linear Classifiers, provide a .decision_function().

Unlike probabilities (which range from to ), the decision function returns a "signed distance" to the decision boundary.

  • Positive value: Predicted as the positive class.
  • Negative value: Predicted as the negative class.
  • Magnitude: Indicates how far the point is from the boundary (certainty).

4. Predicting in Regression​

For regression models, .predict() returns the expected numerical value. Note that standard Scikit-Learn regression models do not provide "probabilities" because the output is a continuous range, not a discrete set of classes.

y^=w1x1+w2x2+...+b\hat{y} = w_1x_1 + w_2x_2 + ... + b

5. Deployment Checklist: The "Input Shape" Trap​

One of the most frequent errors during prediction is a Shape Mismatch.

  • Scikit-Learn estimators expect a 2D array for the input .
  • If you are predicting for a single sample, you must reshape it.
# Error: model.predict([1, 2, 3]) 
# Correct:
single_sample = [1, 2, 3]
model.predict([single_sample]) # Wrapped in a list to make it 2D

6. The Workflow Summary​

References for More Details​

Predictions are useless if they aren't accurate. Now that you know how to get answers from your model, you must learn how to verify if those answers are correct.