Now let's evaluate our model's performance using multiple metrics to get a comprehensive view of its accuracy. First, we'll examine the overall accuracy—the percentage of correct predictions across our entire test set. We can obtain this using the knn_model.score method, though as you'll notice, we need to provide the appropriate data parameters.
The method requires two positional arguments: X and y. This makes sense—to evaluate performance, we need both our test features (X_test) for generating predictions and our test labels (y_test) as the ground truth for comparison. When we pass in our X_test data, the model generates predictions and compares them against the actual answers to calculate our accuracy score.
Our results show 97% accuracy—an impressive performance that indicates we misclassified only 3% of our test samples. Given our test set of 30 samples, this translates to exactly one incorrect prediction. While this single error might seem insignificant, understanding where and why our model fails provides valuable insights for improvement.
Rather than manually scanning through predictions to identify the misclassified sample, we can leverage sklearn's classification report for a more systematic analysis. This tool provides granular performance metrics that reveal not just what went wrong, but which classes are most challenging for our model.
The classification report delivers three key metrics that every data scientist should understand: precision, recall, and F1-score. Precision answers "When we predicted a specific class, how often were we correct?"—essentially measuring the reliability of our positive predictions. Recall addresses "Of all actual instances of a class, how many did we successfully identify?"—capturing our model's ability to find all relevant cases. The F1-score provides the harmonic mean of precision and recall, offering a balanced metric that's particularly useful when dealing with imbalanced datasets.
To generate this comprehensive analysis, we'll import the classification_report function from sklearn.metrics. The function requires our true labels and predicted values, and we can enhance readability by including the iris dataset's target names—'setosa', 'versicolor', and 'virginica'—rather than working with numerical labels.
The resulting report reveals interesting patterns in our model's performance. We achieved perfect precision and recall (1.00) for setosa classification, indicating this species is easily distinguishable from the others based on our selected features. However, our model shows slight confusion between versicolor and virginica species, which is common given their overlapping characteristics in feature space. This granular breakdown helps us understand that while our overall accuracy is excellent, there's room for improvement in distinguishing between these two similar species.
In our next analysis, we'll dive deeper into this confusion pattern and explore techniques for improving classification performance on these challenging boundary cases.