Majority baseline
The accuracy you get by always predicting the most common class. A model has to beat it to be worth anything.
What it means
The majority baseline, or majority-class baseline, is the accuracy you get by always predicting the most common class. It takes no learning at all. If 81% of customers stay, then always saying "stays" scores 81%.
It's the first number to compute for any classifier, because it tells you what an accuracy score is really worth. On this data, 80% accuracy isn't good. It's worse than doing nothing.
An example in Python
from collections import Counter truth = ["stays"] * 81 + ["leaves"] * 19 # The majority baseline: always guess the most common class. most_common, count = Counter(truth).most_common(1)[0] print(f"majority baseline: always '{most_common}', accuracy {count / len(truth):.0%}") # A model that finds 5 of the 19 who leave, but wrongly flags 8 who stay. predicted = ["leaves"] * 8 + ["stays"] * 73 + ["leaves"] * 5 + ["stays"] * 14 right = sum(p == t for p, t in zip(predicted, truth)) print(f"model accuracy: {right / len(truth):.0%}")
majority baseline: always 'stays', accuracy 81% model accuracy: 78%
The model scores 78%, below the 81% baseline. Yet it found 5 of the 19 customers who leave, and the baseline finds none. So it isn't useless: accuracy is the wrong measure for it. When classes are unbalanced, look at precision and recall as well.
In practice
In scikit-learn, DummyClassifier(strategy="most_frequent") gives you this baseline, and you can score it like any other model. For predicting a number, the equivalent is always predicting the average. Either way, it's a floor to beat, not a goal.
Try it now
One score per class, turned into probabilities that add up to 1. Drag a score and watch every probability move.
Learn it properly
Multiclass classification · Softmax and cross-entropy for more than two classes. About 30 minutes.
Other terms from the same lesson
- confusion matrix: A table counting a classifier's predictions against the true answers: one row per true class, one column per predicted class. For yes/no it shows true and false positives and negatives.
- multiclass classification: Classification with more than two possible answers, such as sorting customers by their favorite product category.
- softmax regression: Logistic regression for more than two classes: one score per class, turned into probabilities by softmax.
Never trained an AI?
Teach a small one to write, right in your browser. It takes about 20 seconds and needs no account.