Model Evaluation - lesson 5 | HackIndia Skills

Key point

One idea

A model result means little until you compare it with a real baseline. Your model should beat something simple, fixed, and fair on the same validation set.

What counts as baseline

A baseline is the simplest sensible prediction method for your task. It is not a weak model you choose just to look good.

For classification, common baselines are majority class, random guess with class balance, or a simple rule. For tabular data, a basic logistic regression or decision tree can also be a fair baseline.

Common mistake

Bad baseline choices

Do not compare your model to a number from a blog, a different dataset, or your training score. That is not fair.

Also avoid a fake baseline like 50 percent accuracy when your dataset has 80 percent of one class. A majority-class predictor would already get 80 percent accuracy there.

Your to-do

Pick one fairly

Do each one yourself, then tap it to tick it off. The ticks are only a checklist for you: they are not marked or scored.

Check your target column and count each class.
Choose the simplest method that could be used without training a complex model.
Run the baseline on the same validation split as your model.
Measure the same metric for both, such as recall, precision, or F1.
Keep the baseline result in your notes even if your model barely wins.

0 of 5 done

Example

Majority baseline code

This example uses scikit-learn. It predicts the most common class from the training labels, then evaluates on validation labels.

Use this for a first baseline in a binary or multi-class classification project.

from sklearn.dummy import DummyClassifier
from sklearn.metrics import confusion_matrix, classification_report

# X_train, X_val, y_train, y_val should already come from your clean split

baseline = DummyClassifier(strategy="most_frequent")
baseline.fit(X_train, y_train)

base_pred = baseline.predict(X_val)

print("Baseline confusion matrix")
print(confusion_matrix(y_val, base_pred))

print("Baseline metrics")
print(classification_report(y_val, base_pred, zero_division=0))

Tip

Use AI carefully

You can ask ChatGPT, Gemini, or Claude: “Suggest a fair baseline for this ML task and explain why.” Paste your target column, class counts, and metric.

Check the answer. The baseline must use only training data to decide its rule, must predict on validation data, and must be reported with the same metric as your model.