feat(matplotlib): implement calibration-curve

[github-actions]

Implementation: calibration-curve - matplotlib

Implements the matplotlib version of calibration-curve.

File: plots/calibration-curve/implementations/matplotlib.py

---

🤖 impl-generate workflow

[claude]

AI Review - Attempt 1/3

Image Description

The plot consists of two vertically stacked subplots. The main upper subplot shows three calibration curves against a dashed black diagonal reference line representing perfect calibration. The "Well-Calibrated" model (blue line with circle markers) follows closely along the diagonal. The "Overconfident" model (yellow line with square markers) shows a steep S-curve pattern, jumping sharply from 0 to 1 around the 0.4-0.6 probability range. The "Underconfident" model (pink/magenta line with triangle markers) shows a flatter curve. Each model displays its Brier score in the legend (0.101, 0.020, 0.181 respectively). The lower subplot shows a histogram of predicted probability distributions for all three models, clearly showing that the overconfident model clusters predictions near 0 and 1, while the underconfident model clusters near 0.5, and the well-calibrated model has a more spread distribution. All text is clearly readable, colors are distinct and colorblind-friendly, and the layout is well-balanced.

Quality Score: 93/100

Criteria Checklist

Visual Quality (38/40 pts)

• VQ-01: Text Legibility (10/10) - Title at 24pt, axis labels at 20pt, tick labels at 16pt, legend at 16pt - all perfectly readable
• VQ-02: No Overlap (8/8) - No overlapping text or elements anywhere
• VQ-03: Element Visibility (7/8) - Markers at size 12 with linewidth 3 are clearly visible; could be slightly larger but acceptable
• VQ-04: Color Accessibility (5/5) - Blue, yellow, and pink/magenta are distinguishable for colorblind users
• VQ-05: Layout Balance (5/5) - Two-subplot layout with 3:1 height ratio uses canvas effectively
• VQ-06: Axis Labels (1/2) - Labels are descriptive ("Mean Predicted Probability", "Fraction of Positives", "Count") but lack units
• VQ-07: Grid & Legend (2/2) - Grid at alpha=0.3 with dashed style is subtle, legends well-placed

Spec Compliance (25/25 pts)

• SC-01: Plot Type (8/8) - Correct calibration/reliability diagram with diagonal reference
• SC-02: Data Mapping (5/5) - X-axis shows mean predicted probability, Y-axis shows fraction of positives
• SC-03: Required Features (5/5) - Has diagonal reference line, 10 bins, Brier scores displayed, histogram subplot for prediction distribution, multiple model comparison with distinct colors and legend
• SC-04: Data Range (3/3) - Both axes range from 0 to 1 as appropriate for probabilities
• SC-05: Legend Accuracy (2/2) - Legends correctly identify each model with Brier scores
• SC-06: Title Format (2/2) - Uses exact format "calibration-curve · matplotlib · pyplots.ai"

Data Quality (18/20 pts)

• DQ-01: Feature Coverage (7/8) - Shows well-calibrated, overconfident, and underconfident models demonstrating key calibration patterns; histogram clearly shows distribution differences
• DQ-02: Realistic Context (6/7) - Simulated classifier outputs are plausible; using 35% positive rate is realistic for imbalanced classification
• DQ-03: Appropriate Scale (5/5) - 2000 samples, probabilities correctly bounded 0-1, Brier scores in realistic range

Code Quality (10/10 pts)

• CQ-01: KISS Structure (3/3) - Follows imports → data → plot → save structure, no functions or classes
• CQ-02: Reproducibility (3/3) - Uses np.random.seed(42)
• CQ-03: Clean Imports (2/2) - Only imports matplotlib.pyplot and numpy, both used
• CQ-04: No Deprecated API (1/1) - All APIs are current
• CQ-05: Output Correct (1/1) - Saves as 'plot.png'

Library Features (2/5 pts)

• LF-01: Uses distinctive library features (2/5) - Uses matplotlib correctly with subplots and gridspec_kw for height ratios, but doesn't leverage more distinctive matplotlib features like fill_between for confidence intervals or custom tick formatting

Strengths

• Excellent multi-model comparison showing well-calibrated, overconfident, and underconfident classifiers
• Includes histogram subplot as suggested in spec for showing prediction distributions
• Brier scores integrated into legend for quick comparison
• Clean separation of calibration curve calculation logic
• Colorblind-friendly palette with distinct marker shapes for each model
• Professional layout with appropriate subplot height ratios

Weaknesses

• Axis labels lack units or additional context (e.g., could specify "probability" units explicitly)
• Could use more distinctive matplotlib features like fill_between for confidence bands

Verdict: APPROVED