# Technical Document Extraction: Violation of Equalized Odds Analysis
## Overview
The image contains **8 line graphs** arranged in a 2x4 grid, comparing algorithmic performance metrics across different dimensionality settings. All graphs share consistent labeling conventions and color-coded data series.
---
### **Legend & Color Mapping**
- **Legend Location**: Top-left corner of all graphs
- **Color Assignments**:
- **Blue**: Oracle
- **Green**: FairICP
- **Cyan**: FairCP
- **Red**: FDL
---
### **Graph 1: Violation of Equalized Odds (KPC, dim=1)**
- **X-axis**: KPC: dim = 1 (0.000 → 0.100)
- **Y-axis**: Increased Dimensions (2.00 → 2.25)
- **Trends**:
- Red (FDL) starts highest (2.25) and declines steeply
- Blue (Oracle) and Green (FairICP) converge near 2.00
- **Key Data Points**:
- At X=0.000: FDL=2.25, Oracle=2.18, FairICP=2.18
- At X=0.100: All ≈2.00
---
### **Graph 2: Violation of Equalized Odds (Statistical Power, dim=1)**
- **X-axis**: Power: dim = 1 (0.2 → 1.0)
- **Y-axis**: Increased Dimensions (2.00 → 2.25)
- **Trends**:
- Red (FDL) dominates early (2.25 at X=0.2), then declines
- Cyan (FairCP) and Blue (Oracle) show gradual convergence
- **Key Data Points**:
- At X=0.2: FDL=2.25, FairCP=2.18
- At X=1.0: All ≈2.00
---
### **Graph 3: KPC Loss (dim=5)**
- **X-axis**: KPC: dim = 5 (0.10 → 0.30)
- **Y-axis**: Loss (15 → 50)
- **Trends**:
- Red (FDL) spikes sharply at X=0.15 (≈45)
- Blue (Oracle) and Green (FairICP) remain stable
- **Key Data Points**:
- At X=0.10: FDL=38, Oracle=22
- At X=0.30: All ≈18
---
### **Graph 4: Statistical Power Loss (dim=5)**
- **X-axis**: Power: dim = 5 (0.6 → 1.0)
- **Y-axis**: Loss (15 → 50)
- **Trends**:
- Red (FDL) declines from 40 to 20
- Cyan (FairCP) shows gradual decrease
- **Key Data Points**:
- At X=0.6: FDL=40, FairCP=45
- At X=1.0: All ≈18
---
### **Graph 5: KPC Loss (dim=10)**
- **X-axis**: KPC: dim = 10 (0.10 → 0.25)
- **Y-axis**: Loss (50 → 175)
- **Trends**:
- Red (FDL) spikes dramatically at X=0.20 (≈150)
- Blue (Oracle) and Green (FairICP) remain below 75
- **Key Data Points**:
- At X=0.10: FDL=175, Oracle=50
- At X=0.25: All ≈50
---
### **Graph 6: Statistical Power Loss (dim=10)**
- **X-axis**: Power: dim = 10 (0.85 → 1.00)
- **Y-axis**: Loss (50 → 175)
- **Trends**:
- Red (FDL) declines from 175 to 50
- Cyan (FairCP) shows moderate decrease
- **Key Data Points**:
- At X=0.85: FDL=175, FairCP=125
- At X=1.0: All ≈50
---
### **Cross-Graph Observations**
1. **Dimensionality Impact**:
- Higher dimensions (dim=10) show more extreme loss spikes (FDL)
- KPC-based metrics exhibit sharper declines than power-based metrics
2. **Algorithm Performance**:
- FDL consistently shows highest violation/loss in early ranges
- Oracle and FairICP/FairCP demonstrate better stability
3. **Convergence**:
- All algorithms converge toward baseline loss (≈2.00/50) at maximum X-values
---
### **Critical Notes**
- All y-axis scales are **linear** (no log transformations)
- X-axis ranges vary significantly by dimension (0.000–0.100 for dim=1 vs 0.85–1.00 for dim=10)
- No textual annotations present beyond axis labels and legends
