Find the missing variance from total test error

Question

A model's expected test error is 20. Its squared bias is 5 and the irreducible noise is 3. Expected test error is $\text{squared bias} + \text{variance} + \text{irreducible noise}$ .

What is the model's variance?

Accepted Answer

The idea. Test error is the sum of three parts. Knowing the total and two of the parts pins down the third. Solve. 	ext{Variance} = 	ext{Test error} - 	ext{squared bias} - 	ext{noise} = 20 - 5 - 3 = 12. So variance, at 12, is the biggest contributor to this model's error, larger than the bias (5) and noise (3) put together. When variance dominates like this, the diagnosis is overfitting: the model is flexible and wobbles a lot from sample to sample. The fix is to simplify it or add regularization, not to gather more features. Takeaway. The three-way split lets you recover any missing comp…

Find the missing variance from total test error

Your answer

Discussion

Learn the concepts

Related questions