Evaluating a Model That Fits the Development Set Too Closely
Case context: A team spent six weeks building a fraud-detection model for an online payments platform. During that time, they checked the same development set after nearly every change and adjusted features, thresholds, and network size to improve the development score. By the end of the project, the model scored 95% on the development set but only 79% on the test set.
Question: What evaluation problem should be identified in this scenario, and what should the team do next?
Sample answer: The model has likely overfit to the development set because the team repeatedly optimized against it over the six-week cycle. The large gap between development performance (95%) and test performance (79%) supports that diagnosis. The next step is to create a fresh development set and use it for future model selection.
Key points:
- Diagnose overfitting to the development set.
- Use the large development-test gap as evidence.
- Replace the dev set with a new one for future tuning.
Rubric: Full credit is given for identifying development-set overfitting caused by repeated evaluation and recommending a fresh development set.
0
1
Tags
Machine Learning
Deep Learning
Machine Learning Strategy
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Yearning @ DeepLearning.AI
Related
Keep the Test Set Out of Routine Model Decisions
What makes a model gradually adapt itself to the dev set during development?
A large gap with dev performance much better than test performance can indicate overfitting to the dev set.
If repeated evaluation has made the old development set misleading, get a _____ development set.
Match each dev-set overfitting concept to its description.
Put the dev-set overfitting process in the correct order.
After several rounds of tuning, your validation score is much better than your test score. What should you do next?
It is a good practice to keep checking the test set after every major training change so you can choose the best model.
Repeatedly choosing models based on dev-set results can cause the system to gradually _____ to the dev set.
Match each development-set practice or signal with what it means in model selection.
Order the steps for investigating a validation-versus-test performance gap.
How Repeated Tuning Can Distort the Development Set
Evaluating a Model That Fits the Development Set Too Closely
Sign That the Development Set Has Been Overused