· Xiaojing Yang · Machine Learning · 2 min read
中文Bias–Variance Trade-off in Machine Learning
A practical diagnosis map for underfitting, overfitting, model complexity, and generalization.
Core idea
Bias–variance is a diagnosis tool: it tells me whether the model is too rigid or too sensitive.
1. The core idea
High bias means the model makes strong simplifying assumptions and misses the signal. High variance means the model reacts too much to the training sample.
High bias, underfitting
Useful signal captured
High variance, overfitting
2. What you see in practice
| Pattern | Diagnosis | Possible response |
|---|---|---|
| Train bad, validation bad | High bias | richer features, larger model, better target |
| Train good, validation bad | High variance | regularization, more data, simpler model |
| Big seed variation | High variance | repeated runs, stronger constraints |
| Stable but mediocre | High bias | improve representation |
3. sklearn-style check
Learning curves are a practical way to diagnose whether more data might help.
from sklearn.model_selection import learning_curve
train_sizes, train_scores, val_scores = learning_curve(model, X, y, cv=5, scoring="accuracy")4. AI/NLP connection
In low-resource adaptation, a small model may underfit domain terminology, while a large fine-tuned model may memorize a noisy corpus. Bias–variance language helps explain why parameter-efficient tuning can be attractive.
Interview version
Bias is error from too-simple assumptions; variance is error from sensitivity to training data.
Project version
I check train/validation gaps, seed variation, domain errors, and whether more data or stronger regularization helps.
Takeaway
Bias–variance trade-off turns “the model is bad” into a useful question: what kind of generalization failure is it?
Interview pattern
When this appears in an interview, I would answer in four layers:
- give the short definition;
- explain the intuition;
- name the common failure mode;
- connect it to a real evaluation or deployment decision.