Evaluation Metrics
CS229
Anand Avati
Topics
● Why?
● Binary classifiers
○ Rank view, Thresholding
● Metrics
○ Confusion Matrix
○ Point metrics: Accuracy, Precision, Recall / Sensitivity, Specificity, F-score
○ Summary metrics: AU-ROC, AU-PRC, Log-loss.
● Choosing Metrics
● Class Imbalance
○ Failure scenarios for each metric
● Multi-class
Why are metrics important?
- Training objective (cost function) is only a proxy for real world objective.
- Metrics help capture a business goal into a quantitative target (not all errors
are equal).
- Helps organize ML team effort towards that target.
- Generally in the form of improving that metric on the dev set.
- Useful to quantify the “gap” between:
- Desired performance and baseline (estimate effort initially).
- Desired performance and current performance.
- Measure progress over time (No Free Lunch Theorem).
- Useful for lower level tasks and debugging (like diagnosing bias vs variance).
- Ideally training objective should be the metric, but not always possible. Still,
metrics are useful and important for evaluation.
Binary Classification
● X is Input
● Y is binary Output (0/1)
● Model is ŷ = h(X)
● Two types of models
○ Models that output a categorical class directly (K Nearest neighbor, Decision tree)
○ Models that output a real valued score (SVM, Logistic Regression)
■ Score could be margin (SVM), probability (LR, NN)
■ Need to pick a threshold
■ We focus on this type (the other type can be interpreted as an instance)
Score based models
Score = 1
Positive labelled example
Negative labelled example
#positives
Prevalence =
#positives + #negatives
Score = 0
Score based models : Classifier
Label positive Label negative
Th
Predict Positive
0.5
Th=0.5
Predict Negative
Point metrics: Confusion Matrix
Label Positive Label Negative
Th
9 2
Predict Positive
0.5
Th=0.5
Predict Negative
1 8
Point metrics: True Positives
Label positive Label negative
Th TP
9 2
Predict Positive
0.5 9
Th=0.5
Predict Negative
1 8
Point metrics: True Negatives
Label positive Label negative
Th TP TN
9 2
Predict Positive
0.5 9 8
Th=0.5
Predict Negative
1 8
Point metrics: False Positives
Label positive Label negative
Th TP TN FP
9 2
Predict Positive
0.5 9 8 2
Th=0.5
Predict Negative
1 8
Point metrics: False Negatives
Label positive Label negative
Th TP TN FP FN
9 2
Predict Positive
0.5 9 8 2 1
Th=0.5
Predict Negative
8
1
FP and FN also called Type-1 and Type-2 errors
Could not find true source of image to cite
Point metrics: Accuracy
Label positive Label negative
Th TP TN FP FN Acc
9 2
Predict Positive
0.5 9 8 2 1 .85
Th=0.5
Predict Negative
1 8
Point metrics: Precision
Label positive Label negative
Th TP TN FP FN Acc Pr
9 2
Predict Positive
0.5 9 8 2 1 .85 .81
Th=0.5
Predict Negative
1 8
Point metrics: Positive Recall (Sensitivity)
Label positive Label negative
Th TP TN FP FN Acc Pr Recall
9 2
Predict Positive
0.5 9 8 2 1 .85 .81 .9
Th=0.5
Predict Negative
8
1
Point metrics: Negative Recall (Specificity)
Label positive Label negative
Th TP TN FP FN Acc Pr Recall Spec
9 2
Predict Positive
0.5 9 8 2 1 .85 .81 .9 0.8
Th=0.5
Predict Negative
1 8
Point metrics: F score
Label positive Label negative
Th TP TN FP FN Acc Pr Recall Spec F1
9 2
Predict Positive
0.5 9 8 2 1 .85 .81 .9 .8 .857
Th=0.5
Predict Negative
1 8
Point metrics: Changing threshold
Label positive Label negative
Th TP TN FP FN Acc Pr Recall Spec F1
7 2
Predict Positive
0.6 7 8 2 3 .75 .77 .7 .8 .733
Th=0.6
Predict Negative
3 8
Threshold Scanning Threshold TP TN FP FN Accuracy Precision Recall Specificity F1
1.00 0 10 0 10 0.50 1 0 1 0
0.95 1 10 0 9 0.55 1 0.1 1 0.182
0.90 2 10 0 8 0.60 1 0.2 1 0.333
0.85 2 9 1 8 0.55 0.667 0.2 0.9 0.308
Score = 1
Threshold = 0.80 3 9 1 7 0.60 0.750 0.3 0.9 0.429
1.00
0.75 4 9 1 6 0.65 0.800 0.4 0.9 0.533
0.70 5 9 1 5 0.70 0.833 0.5 0.9 0.625
0.65 5 8 2 5 0.65 0.714 0.5 0.8 0.588
0.60 6 8 2 4 0.70 0.750 0.6 0.8 0.667
0.55 7 8 2 3 0.75 0.778 0.7 0.8 0.737
0.50 8 8 2 2 0.80 0.800 0.8 0.8 0.800
0.45 9 8 2 1 0.85 0.818 0.9 0.8 0.857
0.40 9 7 3 1 0.80 0.750 0.9 0.7 0.818
0.35 9 6 4 1 0.75 0.692 0.9 0.6 0.783
0.30 9 5 5 1 0.70 0.643 0.9 0.5 0.750
0.25 9 4 6 1 0.65 0.600 0.9 0.4 0.720
0.20 9 3 7 1 0.60 0.562 0.9 0.3 0.692
Threshold =
0.00 Score = 0 0.15 9 2 8 1 0.55 0.529 0.9 0.2 0.667
0.10 9 1 9 1 0.50 0.500 0.9 0.1 0.643
0.05 10 1 9 0 0.55 0.526 1 0.1 0.690
0.00 10 0 10 0 0.50 0.500 1 0 0.667
How to summarize the trade-off?
{Precision, Specificity} vs Recall/Sensitivity
Summary metrics: ROC (rotated version)
Score = 1
Score = 0
Summary metrics: PRC
Score = 1
Score = 0
Summary metrics: Log-Loss motivation
Score = 1 Score = 1
Model A Model B
Score = 0 Score = 0
Two models scoring the same data set. Is one of them better than the other?
Summary metrics: Log-Loss
Score = 1 Score = 1
● These two model outputs have same ranking, and
therefore the same AU-ROC, AU-PRC, accuracy!
● Gain =
● Log loss rewards confident correct answers and
heavily penalizes confident wrong answers.
● exp(E[log-loss]) is G.M. of gains, in [0,1].
● One perfectly confident wrong prediction is fatal.
● Gaining popularity as an evaluation metric (Kaggle)
Score = 0 Score = 0
Calibration
Logistic (th=0.5):
Precision: 0.872
Recall: 0.851
F1: 0.862
Brier: 0.099
SVC (th=0.5):
Precision: 0.872
Recall: 0.852
F1: 0.862
Brier: 0.163
Brier = MSE(p, y)
Unsupervised Learning
● logP(x) is a measure of fit in Probabilistic models (GMM, Factor Analysis)
○ High logP(x) on training set, but low logP(x) on test set is a measure of overfitting
○ Raw value of logP(x) hard to interpret in isolation
● K-means is trickier (because of fixed covariance assumption)
Class Imbalance: Problems
Symptom: Prevalence < 5% (no strict definition)
Metrics: may not be meaningful.
Learning: may not focus on minority class examples at all (majority class can overwhelm
logistic regression, to a lesser extent SVM)
Class Imbalance: Metrics (pathological cases)
Accuracy: Blindly predict majority class.
Log-Loss: Majority class can dominate the loss.
AUROC: Easy to keep AUC high by scoring most negatives very low.
AUPRC: Somewhat more robust than AUROC. But other challenges.
- What kind of interpolation? AUCNPR?
In general: Accuracy << AUROC << AUPRC
Multi-class (few remarks)
● Confusion matrix will be NxN (still want heavy diagonals, light off-diagonals)
● Most metrics (except accuracy) generally analysed as multiple 1-vs-many.
● Multiclass variants of AUROC and AUPRC (micro vs macro averaging)
● Class imbalance is common (both in absolute, and relative sense)
● Cost sensitive learning techniques (also helps in Binary Imbalance)
○ Assign $$ value for each block in the confusion matrix, and incorporate those into the loss
function.
Choosing Metrics
Some common patterns:
- High precision is hard constraint, do best recall (e.g search engine results,
grammar correction) -- intolerant to FP. Metric: Recall at Precision=XX%
- High recall is hard constraint, do best precision (e.g medical diag). Intolerant
to FN. Metric: Precision at Recall=100%
- Capacity constrained (by K). Metric: Precision in top-K.
- Etc.
- Choose operating threshold based on above criteria.
Thank You!
