Final: debug.md
Browse files
debug.md
CHANGED
|
@@ -29,5 +29,17 @@
|
|
| 29 |
- Local R=2: ~0.05-0.3s each
|
| 30 |
- Total per config (20 deletions): ~20-40s
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
## Timestamp
|
| 33 |
-
Generated: 2026-04-24
|
|
|
|
| 29 |
- Local R=2: ~0.05-0.3s each
|
| 30 |
- Total per config (20 deletions): ~20-40s
|
| 31 |
|
| 32 |
+
## Gaussian-Gamma MAP Locality Issues
|
| 33 |
+
- MAP (non-convex optimization) shows weaker locality than VI models (57% positive decay vs 97-100%)
|
| 34 |
+
- Error does NOT decrease with radius for GGM, unlike PG and GG models
|
| 35 |
+
- This is consistent with MAP being non-convex: the "exact deletion" solution depends on optimization path
|
| 36 |
+
- Labeled as exploratory ablation, not as theory validation
|
| 37 |
+
|
| 38 |
+
## Real-Data Notes
|
| 39 |
+
- Last.fm: loaded via random sampling of 500k rows from 13.4M total; N=300, M=300, 2546 edges
|
| 40 |
+
- MovieLens: loaded from ashraq/movielens_ratings; N=500, M=500, ~3190 edges
|
| 41 |
+
- Both datasets show positive locality behavior qualitatively consistent with synthetic results
|
| 42 |
+
- Real-data graphs are sparser and more heterogeneous than synthetic bounded-degree graphs
|
| 43 |
+
|
| 44 |
## Timestamp
|
| 45 |
+
Generated: 2026-04-24, Updated: 2026-04-24T21:00
|