Precision vs recall vs F1
Precision asks how many predicted positives were right, recall asks how many real positives were found, and F1 balances both.
- Precision avoids false positives
- Recall avoids false negatives
- F1 is harmonic mean
Quick recall
374 cards
Supervised vs unsupervised learning
Supervised learning uses labeled targets, while unsupervised learning looks for structure without target labels.
- Supervised predicts known target
- Unsupervised finds patterns
- Examples: classification vs clustering
What is a confusion matrix?
A confusion matrix counts true and false predictions by actual and predicted class.
- TP FP FN TN for binary case
- Helps derive precision and recall
- Great for error analysis
What is a learning rate?
The learning rate controls how large each parameter update step is during training.
- Too high can overshoot
- Too low learns slowly
- Often tuned experimentally
What is data leakage in machine learning?
Data leakage happens when training uses information that would not be available at real prediction time.
- Inflates evaluation scores
- Common in preprocessing mistakes
- Think carefully about time and target leakage
What is dimensionality reduction?
Dimensionality reduction compresses features into fewer dimensions while keeping as much useful structure as possible.
- Can help visualization
- May reduce noise
- PCA is a common example
What is feature engineering?
Feature engineering transforms raw data into signals that a model can learn from more effectively.
- Encode useful domain knowledge
- Scale, bucket, aggregate, or combine
- Bad leakage ruins evaluation
What is gradient descent?
Gradient descent updates model parameters in the direction that reduces the loss function.
- Uses gradients of loss
- Step size is learning rate
- Variants include batch and mini-batch
What is regularization?
Regularization discourages overly complex models so they generalize better.
- L1 can drive weights to zero
- L2 shrinks weights smoothly
- Early stopping is also a regularization idea
What is ROC-AUC?
ROC-AUC measures how well a classifier ranks positives ahead of negatives across decision thresholds.
- Threshold-independent ranking view
- Useful for comparing classifiers
- Can mislead on highly imbalanced data
What is transfer learning?
Transfer learning starts from a pretrained model and adapts it to a new but related task.
- Saves training time
- Useful with limited labeled data
- Common in vision and NLP
Why use cross validation?
Cross validation estimates performance more reliably by training and evaluating on multiple folds of the data.
- Reduces single split luck
- Useful for small data
- Common in model comparison