The Cross-Industry Standard Process for Data Mining: a six-phase, iterative methodology for data and analytics projects that remains the default reference in AI certification syllabi.

When to use: For any data science or AI initiative that must deliver a business outcome — and as the exam-level answer for how an AI project is structured end to end.

Machine Learning Project Lifecycle

The production-oriented ML lifecycle — scoping, data, modeling, deployment and monitoring — covering the MLOps concerns that CRISP-DM predates.

When to use: When an ML model has to run in production rather than in a notebook — the framing used by AI-102, AIF-C01 and the Google ML Engineer exams.

Formulas library

Every formula with variables, interpretation thresholds and a worked example.

Accuracy

Accuracy = (TP + TN) / (TP + TN + FP + FN) x 100

TP: True positives (count)
TN: True negatives (count)
FP: False positives (count)
FN: False negatives (count)

> 90 — High accuracy
70 - 90 — Moderate accuracy
< 70 — Low accuracy

Worked example

A support-ticket classifier is evaluated on 1,000 tickets: 380 true positives, 520 true negatives, 60 false positives and 40 false negatives.

Accuracy = (380 + 520) / 1,000 x 100 = 90%

Nine in ten tickets are routed correctly, but the 40 missed positives may matter more than the headline number.

Try: A support-ticket classifier is evaluated on 1,000 tickets: 380 true positives, 520 true negatives, 60 false positives and 40 false negatives.

Precision

Precision = TP / (TP + FP) x 100

TP: True positives (count)
FP: False positives (count)

> 90 — Very precise
70 - 90 — Moderate precision
< 70 — Low precision

Worked example

The same classifier flagged 440 tickets as urgent; 380 truly were.

Precision = 380 / (380 + 60) x 100 = 86.36%

About one in seven urgent flags is a false alarm — acceptable for triage, too high for auto-escalation.

Try: The same classifier flagged 440 tickets as urgent; 380 truly were.

Recall (Sensitivity)

Recall = TP / (TP + FN) x 100

TP: True positives (count)
FN: False negatives (count)

> 90 — High recall
70 - 90 — Moderate recall
< 70 — Low recall

Worked example

There were 420 genuinely urgent tickets; the classifier caught 380 and missed 40.

Recall = 380 / (380 + 40) x 100 = 90.48%

Nine in ten urgent tickets are caught; the 40 missed cases are the ones to investigate for a pattern.

Try: There were 420 genuinely urgent tickets; the classifier caught 380 and missed 40.

F1 Score

F1 = 2 x (Precision x Recall) / (Precision + Recall)

Precision: Precision (%)
Recall: Recall (%)

> 90 — Excellent balance
70 - 90 — Acceptable balance
< 70 — Poor balance

Worked example

The ticket classifier scores 86.4% precision and 90.5% recall.

F1 = 2 x (86.36 x 90.48) / (86.36 + 90.48) = 88.37

A balanced model — neither false alarms nor misses dominate the error profile.

Try: The ticket classifier scores 86.

Confusion Matrix helper (accuracy, precision, recall, F1)

TP / FP / TN / FN -> Accuracy, Precision, Recall, F1

TP: True positives (count)
FP: False positives (count)
TN: True negatives (count)
FN: False negatives (count)

> 90 — Strong classifier
70 - 90 — Usable classifier
< 70 — Weak classifier

Worked example

Support-ticket classifier over 1,000 tickets: TP 380, FP 60, TN 520, FN 40.

Accuracy 90.00% - Precision 86.36% - Recall 90.48% - F1 88.37%

Errors are split fairly evenly between false alarms and misses, so threshold tuning trades one for the other.

Try: Support-ticket classifier over 1,000 tickets: TP 380, FP 60, TN 520, FN 40.