Top 15 Machine Learning Evaluation Metrics: The Ultimate Interview Q&A Guide

Machine Learning: Evaluation Metrics Explained

Building a machine learning model is a journey, not a destination. Once you've trained a model, how do you know if it's any good? How does it compare to other models? The answer lies in **evaluation metrics**.

Evaluation metrics are the tools we use to measure the performance, quality, and effectiveness of a model. Choosing the right metric is just as crucial as choosing the right algorithm, because the metric defines what you consider a "successful" outcome.

Broadly, we use different metrics for different tasks:

• Classification Metrics: Used when the model predicts a category (e.g., 'Spam' or 'Not Spam', 'Cat' or 'Dog').
• Regression Metrics: Used when the model predicts a continuous numerical value (e.g., house price, temperature).

This guide breaks down the 15 most important evaluation metrics in a simple Q&A format to help you master them for your projects and data science interviews.

Top 15 Interview Questions & Answers

1. What is a Confusion Matrix and where do we use it?

A Confusion Matrix is a table used to evaluate the performance of a classification model. It provides a detailed breakdown of how many predictions were correct and what kinds of mistakes were made.

Confusion Matrix Structure

Predicted Class

Positive

Negative

Actual

TP
(True Positive)

FN
(False Negative)

Class

FP
(False Positive)

TN
(True Negative)

It's especially useful for:

• Visualizing model performance beyond simple accuracy.
• Understanding performance on imbalanced datasets.
• Calculating more advanced metrics like Precision, Recall, and F1-Score.

2. How do we evaluate Regression and Classification models?

We use different sets of metrics for regression and classification because they solve different types of problems.

• Classification Models (predicting a category, e.g., 'Spam' or 'Not Spam'):
  • Accuracy, Precision, Recall, F1-Score: Based on the Confusion Matrix.
  • AUC-ROC Curve: Measures performance across all classification thresholds.
• Regression Models (predicting a continuous value, e.g., house price):
  • R-squared (R²) & Adjusted R-squared: Explains variance.
  • Mean Absolute Error (MAE): Average absolute error.
  • Root Mean Squared Error (RMSE): Punishes large errors more heavily.

3. What is R-squared and Adjusted R-squared? Which is better?

R-squared (R²) measures how much of the variation in your dependent variable is explained by your model. It ranges from 0 to 1 (0% to 100%).

The Problem with R-squared: It always increases when you add more features, even if they are useless. This can be misleading.

Adjusted R-squared solves this by penalizing the model for adding features that don't improve it. It only increases if a new feature is truly valuable.

Which is better? For comparing models with a different number of features, Adjusted R-squared is better as it provides a more honest assessment.

4. What is RMSE and where do we use it?

RMSE stands for Root Mean Squared Error. It's a standard metric for evaluating regression models by measuring the average magnitude of the errors.

RMSE: Visualizing Prediction Errors

Why use it? RMSE gives the error in the same units as the target variable (e.g., dollars), making it easy to interpret. A lower RMSE indicates a better model fit.

5. What are AIC and BIC?

AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) are metrics for model selection that balance goodness-of-fit with model complexity.

The goal is to find the model with the lowest AIC or BIC score.

• AIC: Can sometimes favor more complex models for better prediction.
• BIC: Puts a heavier penalty on complexity, preferring simpler models.

6. When to use RMSE vs. Adjusted R-squared?

You should often use both! They answer different questions:

• Use Adjusted R-squared to... understand the *percentage* of variance your model explains. It's a relative measure of goodness-of-fit. (e.g., "The model explains 85% of the variance.")
• Use RMSE to... understand the *absolute* prediction error in original units. It tells you how far off your predictions are on average. (e.g., "The model is off by $10,500 on average.")

7. What are Classification Report, Confusion Matrix, and Accuracy?

These are all tools to evaluate classification models:

• Accuracy: The simplest metric. Ratio of correct predictions to total predictions. Often misleading on imbalanced datasets.
• Confusion Matrix: The foundational table showing TP, TN, FP, and FN.
• Classification Report: A text summary showing the precision, recall, and F1-score for each class in one convenient report.

8. What are Precision and Recall?

Precision and Recall are crucial for imbalanced datasets and offer a more nuanced view than accuracy.

• Precision: Of all the positive predictions, how many were correct? It measures the cost of false positives.

Precision = TP / (TP + FP)

True Positives (TP)

False Positives (FP)

Precision focuses on the group the model *predicted* as positive.

• Recall (Sensitivity): Of all the actual positive cases, how many did the model find? It measures the cost of false negatives.

Recall = TP / (TP + FN)

True Positives (TP)

False Negatives (FN)

Recall focuses on the group that was *actually* positive.

9. What is F1-Score and Support in a Confusion Matrix?

F1-Score: The harmonic mean of Precision and Recall. It provides a single score that balances both metrics. Use it when both false positives and false negatives are important.

Formula: 2 * (Precision * Recall) / (Precision + Recall)

Support: In a classification report, "support" is simply the number of actual occurrences of each class in your dataset. It provides context for the other metrics.

10. What are TPR and FPR?

TPR and FPR are the two axes of the ROC curve.

• TPR (True Positive Rate): Same as Recall. What proportion of actual positives were correctly identified? TP / (TP + FN)
• FPR (False Positive Rate): What proportion of actual negatives were incorrectly classified as positive? FP / (FP + TN)

11. What are Sensitivity and Specificity?

These terms are common in medical fields and are equivalent to other metrics.

• Sensitivity: Exactly the same as Recall and TPR.
• Specificity: The True Negative Rate. It measures the ability to correctly identify true negatives. TN / (TN + FP). It is also equal to 1 - FPR.

12. What are TP, TN, FP, and FN?

These are the four quadrants of the Confusion Matrix:

• TP (True Positive): Actual: Positive, Predicted: Positive. (Correct)
• TN (True Negative): Actual: Negative, Predicted: Negative. (Correct)
• FP (False Positive) / Type I Error: Actual: Negative, Predicted: Positive. (Incorrect)
• FN (False Negative) / Type II Error: Actual: Positive, Predicted: Negative. (Incorrect)

13. What are Type 1 and Type 2 Errors? Which is more dangerous?

These errors correspond to False Positives and False Negatives.

Identifying Type 1 & Type 2 Errors

Predicted +

Predicted -

Actual +

Correct

Type 2 Error
(False Negative)

Actual -

Type 1 Error
(False Positive)

Correct

Which is more dangerous? It completely depends on the context.

• In a medical test for a disease, a Type 2 Error (telling a sick person they are healthy) is more dangerous.
• In a spam filter, a Type 1 Error (marking an important email as spam) is more dangerous.

14. What are Micro, Macro, and Weighted Averages?

When averaging metrics in multi-class classification:

• Macro Average: Averages the metric for each class, treating all classes equally.
• Micro Average: Aggregates the contributions of all classes to compute the metric globally. (Equivalent to overall accuracy).
• Weighted Average: Averages the metric for each class, but weights each by its support (number of true instances). This accounts for class imbalance.

15. What are AUC and ROC?

ROC (Receiver Operating Characteristic) Curve: A graph showing a classifier's performance across all decision thresholds. It plots the True Positive Rate (TPR) against the False Positive Rate (FPR).

AUC (Area Under the Curve): The area under the ROC curve. It provides a single number to summarize the model's performance.

AUC-ROC Curve

AUC is an excellent metric for imbalanced datasets because it measures how well the model can distinguish between classes, regardless of the chosen threshold.

By understanding these 15 metrics, you're well-equipped to evaluate your own models effectively and confidently explain your choices in a data science interview.