Can One k Always Maximize kNN Leave-One-Out Accuracy?
Company: C3 AI
Role: Data Scientist
Category: Machine Learning
Difficulty: medium
Interview Round: Online Assessment
# Can One k Always Maximize kNN Leave-One-Out Accuracy?
A question asks which value of k maximizes leave-one-out cross-validation accuracy for k-nearest-neighbor classification: 1, 3, 5, any k gives the same result, or none of the above. No dataset, distance metric, or labels are supplied. What can you conclude, and how would you choose k when data are available?
### What a Strong Answer Covers
- Recognition that no universal k maximizes leave-one-out accuracy across datasets.
- Correct exclusion of each held-out observation from its own neighbor set.
- The bias-variance trade-off and leakage-safe preprocessing.
```hint The held-out point is absent
The zero distance from a training observation to itself cannot give k=1 perfect leave-one-out accuracy.
```
### Follow-up Questions
- Why is training accuracy at k=1 different from leave-one-out accuracy?
- How should feature scaling be fitted within the validation procedure?
Overview: Explain why kNN has no universally optimal leave-one-out k and choose neighbors with correct holdout, preprocessing, and evaluation.
Read the full C3 AI Data Scientist interview experience this question came from