Skip to content

model_monitroing_metrics.md

Latest commit

 

History

History
275 lines (190 loc) · 10.6 KB

model_monitroing_metrics.md

File metadata and controls

275 lines (190 loc) · 10.6 KB

Model Monitoring Metrics

1. Population Stability Index (PSI)

Purpose: Measures shift in the distribution of a continuous variable (like model confidence) between a baseline (training or reference) and current population. Used for data drift / model input shift monitoring.

Formula:

$$ PSI = \sum_{i=1}^{n} (P_i - Q_i) \cdot \ln\left(\frac{P_i}{Q_i}\right) $$

Where:

  • $n$ = number of bins
  • $P_i$ = % of baseline (expected) data in bin i
  • $Q_i$ = % of current (actual) data in bin i

Calculation Steps:

  1. Divide baseline data into n bins (quantiles or fixed width)
  2. Count % of baseline data in each bin → $P_i$
  3. Count % of current/live data in each bin → $Q_i$
  4. Apply the formula above

Interpretation / Thresholds:

  • PSI < 0.1 → No significant change
  • 0.1 < PSI < 0.25 → Moderate shift
  • PSI > 0.25 → Significant shift (investigate data/model)

Use Cases:

  • Monitor model confidence distributions (module/date softmax)
  • Detect population drift in features or outputs

2. Characteristic Stability Index (CSI)

Purpose: Similar to PSI, but calculated per feature, i.e., measures input feature distribution drift.

Formula: Same as PSI but applied to feature distributions.

$$ CSI_{feature} = \sum_{i=1}^{n} (P_i - Q_i) \cdot \ln\left(\frac{P_i}{Q_i}\right) $$

Calculation Steps:

  1. For each feature, split baseline into n bins
  2. Compute baseline and live % per bin
  3. Apply formula

Thresholds & Interpretation:

  • CSI < 0.1 → Feature stable
  • 0.1–0.25 → Moderate drift
  • > 0.25 → Strong drift (may require retraining)

Use Cases:

  • Track input features in tabular models or embedding dimensions

3. Concept Drift / Model Drift

Purpose: Detects when the relationship between inputs and outputs changes. Unlike PSI/CSI which track input distribution, concept drift checks if model predictions become unreliable.

Common Methods:

a) Prediction Distribution Drift

  • Compare predicted probability distribution over time vs training
  • Use PSI formula on model predictions (softmax probabilities)
  • Example: Confidence PSI for module/date predictions

b) Performance-based drift

  • Track metrics like accuracy, F1-score, AUC over a sliding window of live data
  • If metric drops significantly → drift detected

Formula:

For confidence / probability drift: same as PSI applied to predicted class probabilities.

For performance: simple moving average difference:

$$ \Delta Accuracy = Accuracy_{baseline} - Accuracy_{live} $$

Thresholds:

  • Accuracy drop > 5–10% may indicate drift
  • PSI of predictions > 0.1–0.25 → moderate to severe drift

4. Embedding Drift / Representation Drift

Purpose: Measures change in feature embeddings (often from neural networks) between baseline and live data.

Method: Cosine similarity is commonly used.

Formula:

$$ \text{Cosine Similarity} = \frac{\vec{e}_{live} \cdot \vec{e}_{baseline}}{|\vec{e}_{live}| |\vec{e}_{baseline}|} $$

$$ \text{Embedding Drift} = 1 - \text{Cosine Similarity} $$

Steps:

  1. Compute mean embedding from training set → $\vec{e}_{baseline}$
  2. For each live query, compute its embedding → $\vec{e}_{live}$
  3. Calculate cosine similarity → subtract from 1 to get drift score

Interpretation:

  • Drift < 0.1 → embeddings stable
  • 0.1–0.3 → moderate change
  • > 0.3 → significant drift (investigate input or model)

Use Cases:

  • NLP embeddings for BERT/TinyBERT outputs
  • Image embeddings in CV pipelines

5. Prediction Entropy

Purpose: Measures uncertainty of model predictions.

Formula:

$$ H(p) = -\sum_{i} p_i \log(p_i) $$

Where $p_i$ = softmax probability for class i.

Interpretation:

  • Low entropy → confident predictions
  • High entropy → uncertain predictions → may trigger human review or alert

Thresholds:

  • Relative to training distribution; monitor moving average

6. Low Confidence Count

Purpose: Tracks how often the model outputs low-confidence predictions.

Method:

  • Count predictions where max(softmax_probs) < threshold
  • Threshold is usually 0.5 (or domain-specific)

Interpretation:

  • Sudden rise → model unsure on current population → potential concept drift

MLOps Monitoring Metrics - Comprehensive Guide

This guide covers all commonly used MLOps monitoring metrics for tabular, NLP, CV, and classical ML applications, including formulas, interpretations, and threshold guidance.

1. Data Drift / Input Drift Metrics

These metrics monitor changes in the input feature distributions between training and live data.

Metric Formula / Method Use Case Threshold / Meaning
Population Stability Index (PSI) $\sum_{i=1}^{n} (P_i - Q_i) \ln(P_i/Q_i)$ Continuous/categorical feature drift <0.1 stable, 0.1–0.25 moderate, >0.25 severe drift
Characteristic Stability Index (CSI) PSI per feature Categorical/continuous input features Same as PSI
Kolmogorov–Smirnov (KS) Test $D = \max |F_1(x) - F_2(x)|$ Continuous feature drift p-value < 0.05 → significant drift
Jensen-Shannon Divergence (JSD) $JSD(P || Q) = \frac{1}{2}KL(P || M) + \frac{1}{2}KL(Q || M)$ Distribution comparison 0–1, higher → more divergence
Wasserstein Distance (Earth Mover's Distance) Measures "cost" to move baseline distribution to live Continuous features Domain-specific thresholds

✅ Use in all applications: tabular, NLP embeddings, CV features (pixel distributions, extracted features).

2. Concept Drift / Model Drift Metrics

These metrics measure changes in the relationship between inputs and outputs.

Metric Formula / Method Use Case Threshold / Meaning
Prediction Distribution Drift (Confidence PSI) PSI on softmax probabilities per class Classification models <0.1 stable, >0.25 alert
Accuracy / F1 / AUC Moving Average Compare baseline metric vs live metric over window All supervised tasks Drop > 5–10% → drift
KL Divergence of Predictions $KL(P_{train} || P_{live})$ Classification output drift Higher → model behavior changed
Prediction Entropy $H(p) = -\sum p_i \log(p_i)$ Uncertainty monitoring High → model unsure
Low Confidence Count Count(max_prob < threshold) Classification confidence Rising count → alert

3. Embedding / Representation Drift Metrics

For deep learning embeddings (text, images, tabular):

Metric Formula / Method Use Case Threshold / Meaning
Cosine Similarity Drift $1 - \frac{\vec{e}{live} \cdot \vec{e}{baseline}}{|\vec{e}{live}| |\vec{e}{baseline}|}$ NLP embeddings, CV embeddings <0.1 stable, 0.1–0.3 moderate, >0.3 severe
Mahalanobis Distance $d_M = \sqrt{(x - \mu)^T \Sigma^{-1} (x - \mu)}$ Detect OOD in embeddings > threshold → anomaly
Euclidean / L2 Distance $|x_{live} - \mu_{baseline}|_2$ Embedding drift Domain-specific

✅ Use Cases:

  • NLP: BERT/TinyBERT CLS token embeddings
  • CV: CNN feature maps or bottleneck layers
  • Tabular: Autoencoder latent representations

4. Out-of-Distribution (OOD) Detection / Novelty Detection

Metric Formula / Method Use Case Threshold / Meaning
Maximum Softmax Probability (MSP) $1 - \max(p_i)$ Detect OOD samples High → anomalous
ODIN Score Temperature-scaled softmax with small perturbation NLP, CV Higher → OOD
Mahalanobis Distance in Feature Space As above Detect OOD embeddings High → OOD
Autoencoder Reconstruction Error $|x - \hat{x}|$ Anomaly detection High error → OOD

5. Regression / Continuous Output Monitoring

Metric Formula / Method Use Case Threshold / Meaning
Mean Squared Error (MSE) Drift Compare live MSE with baseline Regression models Significant increase → drift
Mean Absolute Error (MAE) Drift Compare live MAE with baseline Regression models Significant increase → drift
Residual Distribution Drift (KS Test / PSI) Apply KS/PSI to residuals Regression Residual distribution shifts → model misalignment
Prediction Interval Coverage % of true values inside predicted interval Uncertainty monitoring Drop → model underestimates uncertainty

6. Feature Importance / SHAP Monitoring

Metric Formula / Method Use Case Threshold / Meaning
SHAP Distribution Drift PSI / KS test on SHAP values per feature Detect change in model reasoning High drift → retrain or investigate
Permutation Importance Drift Compare feature importance baseline vs live Tabular / CV Drift → model relying on different features

7. Latency and System Metrics

Metric Formula / Method Use Case Threshold / Meaning
Inference Latency Wall-clock time per request All models SLA violation → alert
Throughput / Requests per Second Count / time System performance Low → bottleneck
Error Rate Failed inference count / total Reliability High → alert

8. Summary Table (All Domains)

Metric Domain Formula / Method Interpretation
PSI Tabular/NLP/CV $\sum (P_i - Q_i) \ln(P_i/Q_i)$ Data / prediction distribution shift
CSI Tabular PSI per feature Feature drift
KL / JSD Tabular/NLP/CV $\sum P \log(P/Q)$ Distribution change
Cosine Drift NLP/CV 1 - cos_sim Embedding drift
Mahalanobis / L2 NLP/CV/Tabular Distance in feature space OOD detection
Entropy Classification $-\sum p_i \log(p_i)$ Prediction uncertainty
Low Confidence Classification Count(max_prob < threshold) Confidence monitoring
Residual / Error Drift Regression MSE/MAE comparison Model performance drift
SHAP / Feature Drift Tabular/NLP/CV PSI / KS on importance Explainability drift
Latency All Time per inference SLA monitoring

Key Points

  • PSI/CSI → distribution shift
  • Embedding / Representation Drift → changes in feature space
  • Prediction-based Drift → model concept drift
  • Entropy / Confidence → uncertainty monitoring
  • Residual / Error Drift → regression model performance
  • SHAP / Feature Drift → explainability shift
  • System Metrics → operational health