# Technical Data Extraction: Performance Comparison of ASGL vs. Baselines
This document provides a comprehensive extraction of data from a series of five line charts comparing the performance of various graph neural network models under different privacy constraints.
## 1. Document Metadata & Global Components
* **Image Type:** Multi-panel line graph (5 subplots).
* **Primary Language:** English.
* **Global Legend (Top Center):**
* **SDGNN:** Grey line with open circles ($\circ$).
* **SiGAT:** Pink line with open circles ($\circ$).
* **SGCN:** Blue line with open circles ($\circ$).
* **GAP:** Green line with open squares ($\square$).
* **LSNE:** Gold/Yellow line with open diamonds ($\diamond$).
* **ASGL (Proposed):** Red line with open triangles ($\triangle$).
* **Common X-Axis:** Privacy budget ($\epsilon$), ranging from 1 to 6.
* **Common Y-Axis:** AUC (Area Under the Curve), ranging generally from 0.50 to 0.90.
---
## 2. Component Isolation & Data Extraction
### (a) Bitcoin_Alpha
* **Trend Analysis:** All models show an upward trend as the privacy budget increases. **ASGL (Proposed)** maintains the highest AUC throughout, starting at ~0.75 and plateauing near 0.86. **LSNE** shows the most significant growth, starting as the lowest performer and surpassing GAP and SGCN by $\epsilon=6$.
* **Approximate Data Points (AUC):**
| Model | $\epsilon=1$ | $\epsilon=2$ | $\epsilon=3$ | $\epsilon=4$ | $\epsilon=5$ | $\epsilon=6$ |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
| **ASGL** | 0.75 | 0.81 | 0.83 | 0.86 | 0.86 | 0.86 |
| **SiGAT** | 0.71 | 0.73 | 0.73 | 0.75 | 0.79 | 0.82 |
| **SDGNN** | 0.68 | 0.69 | 0.71 | 0.73 | 0.80 | 0.85 |
| **GAP** | 0.57 | 0.60 | 0.64 | 0.71 | 0.73 | 0.73 |
| **SGCN** | 0.52 | 0.55 | 0.64 | 0.69 | 0.73 | 0.77 |
| **LSNE** | 0.51 | 0.54 | 0.60 | 0.65 | 0.76 | 0.81 |
### (b) Bitcoin_OCT
* **Trend Analysis:** **ASGL** dominates, reaching a peak of ~0.88. **SDGNN** and **SiGAT** follow closely. **SGCN** and **GAP** show a sharp increase between $\epsilon=2$ and $\epsilon=4$.
* **Approximate Data Points (AUC):**
| Model | $\epsilon=1$ | $\epsilon=2$ | $\epsilon=3$ | $\epsilon=4$ | $\epsilon=5$ | $\epsilon=6$ |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
| **ASGL** | 0.80 | 0.85 | 0.85 | 0.85 | 0.87 | 0.88 |
| **SDGNN** | 0.77 | 0.79 | 0.79 | 0.82 | 0.83 | 0.86 |
| **SiGAT** | 0.70 | 0.73 | 0.79 | 0.84 | 0.86 | 0.87 |
| **SGCN** | 0.56 | 0.58 | 0.67 | 0.75 | 0.76 | 0.78 |
| **GAP** | 0.58 | 0.58 | 0.65 | 0.68 | 0.70 | 0.74 |
| **LSNE** | 0.51 | 0.54 | 0.65 | 0.81 | 0.87 | 0.88 |
### (c) WikiRfA
* **Trend Analysis:** **ASGL** shows a steep improvement from $\epsilon=1$ to $\epsilon=3$, then plateaus. **LSNE** shows a steady linear-like climb, eventually matching **SDGNN** at $\epsilon=6$. **GAP** remains the lowest performer with very slow growth.
* **Approximate Data Points (AUC):**
| Model | $\epsilon=1$ | $\epsilon=2$ | $\epsilon=3$ | $\epsilon=4$ | $\epsilon=5$ | $\epsilon=6$ |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
| **ASGL** | 0.67 | 0.77 | 0.80 | 0.80 | 0.80 | 0.81 |
| **SDGNN** | 0.66 | 0.71 | 0.72 | 0.72 | 0.76 | 0.79 |
| **SiGAT** | 0.63 | 0.65 | 0.71 | 0.73 | 0.74 | 0.80 |
| **SGCN** | 0.51 | 0.65 | 0.65 | 0.70 | 0.71 | 0.71 |
| **LSNE** | 0.51 | 0.53 | 0.61 | 0.66 | 0.74 | 0.79 |
| **GAP** | 0.54 | 0.55 | 0.56 | 0.57 | 0.58 | 0.60 |
### (d) Slashdot
* **Trend Analysis:** **ASGL** and **SDGNN** are the top performers, converging at ~0.89. **LSNE** exhibits a "S-curve" trend, starting very low and rising sharply between $\epsilon=2$ and $\epsilon=4$.
* **Approximate Data Points (AUC):**
| Model | $\epsilon=1$ | $\epsilon=2$ | $\epsilon=3$ | $\epsilon=4$ | $\epsilon=5$ | $\epsilon=6$ |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
| **ASGL** | 0.79 | 0.86 | 0.88 | 0.89 | 0.89 | 0.89 |
| **SDGNN** | 0.76 | 0.84 | 0.87 | 0.88 | 0.89 | 0.89 |
| **SiGAT** | 0.71 | 0.79 | 0.84 | 0.84 | 0.85 | 0.85 |
| **LSNE** | 0.57 | 0.62 | 0.76 | 0.78 | 0.78 | 0.78 |
| **SGCN** | 0.57 | 0.62 | 0.67 | 0.72 | 0.78 | 0.81 |
| **GAP** | 0.61 | 0.64 | 0.69 | 0.71 | 0.74 | 0.75 |
### (e) Epinions
* **Trend Analysis:** **ASGL** leads significantly in the low-privacy budget range ($\epsilon=2, 3$). By $\epsilon=6$, **ASGL**, **SDGNN**, **SiGAT**, and **LSNE** all converge around 0.84-0.87. **GAP** remains significantly lower than the others.
* **Approximate Data Points (AUC):**
| Model | $\epsilon=1$ | $\epsilon=2$ | $\epsilon=3$ | $\epsilon=4$ | $\epsilon=5$ | $\epsilon=6$ |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
| **ASGL** | 0.68 | 0.82 | 0.85 | 0.87 | 0.87 | 0.87 |
| **SDGNN** | 0.68 | 0.72 | 0.72 | 0.85 | 0.85 | 0.87 |
| **SiGAT** | 0.68 | 0.71 | 0.71 | 0.75 | 0.79 | 0.84 |
| **LSNE** | 0.51 | 0.61 | 0.76 | 0.85 | 0.86 | 0.86 |
| **SGCN** | 0.62 | 0.65 | 0.70 | 0.75 | 0.81 | 0.84 |
| **GAP** | 0.59 | 0.61 | 0.63 | 0.63 | 0.65 | 0.67 |
---
## 3. Summary of Findings
Across all five datasets (Bitcoin_Alpha, Bitcoin_OCT, WikiRfA, Slashdot, and Epinions), the **ASGL (Proposed)** model consistently outperforms or matches the baseline models. It is particularly robust at lower privacy budgets ($\epsilon < 3$), where other models like LSNE and SGCN show significantly lower performance. As the privacy budget increases (allowing for less noise), all models generally improve, with most converging toward a high AUC between 0.80 and 0.90.
