Xai feautures

.claude/settings.local.json

@@ -0,0 +1,23 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(git status *)",
+      "Bash(cat .gitignore)",
+      "Bash(../../venv/Scripts/python.exe step4_analysis.py)",
+      "Bash(venv/Scripts/python.exe -c ' *)",
+      "Bash(git add *)",
+      "Bash(git commit -m ' *)",
+      "Read(//c/Users/ddagher/AppData/Roaming/Claude/local-agent-mode-sessions/skills-plugin/eb3cff85-bb14-4303-9b21-bb0d69e3fe6d/6e246a8a-a583-4ff3-9ca1-346d79601d86/skills/pptx/**)",
+      "Bash(python scripts/thumbnail.py \"C:\\\\Users\\\\ddagher\\\\OneDrive - Capgemini\\\\Bureau\\\\Nouveau Présentation Microsoft PowerPoint.pptx\")",
+      "Bash(pip install *)",
+      "Bash(npm list *)",
+      "PowerShell(node --version)",
+      "PowerShell(npm list -g pptxgenjs 2>&1)",
+      "Bash(venv/Scripts/python.exe -c \"import pptx; print\\(pptx.__version__\\)\")",
+      "Bash(venv/Scripts/pip.exe install *)",
+      "Bash(venv/Scripts/python.exe -c \"import pptx; print\\('pptx', pptx.__version__\\)\")",
+      "Bash(ls \"C:\\\\Users\\\\ddagher\\\\Downloads\\\\slicing-simulations.zip\\\\2.2.1-10kprocessed\" 2>&1 | head -50)",
+      "PowerShell(Get-ChildItem -Recurse -Directory -ErrorAction SilentlyContinue | Where-Object { $_.Name -match 'data|dataset|sample|record' } | Select-Object -ExpandProperty FullName)"
+    ]
+  }
+}

.gitignore

@@ -0,0 +1,7 @@
+__pycache__/
+*.pyc
+venv/
+*.npy
+data/
+.DS_Store
+Thumbs.db

notes.md

@@ -0,0 +1,55 @@
+
+# Step 1 — DONE: Understand Input/Output
+Reverse-engineered delay_model.py completely. RouteNet-Fermi is a Graph Neural Network with message-passing between 3 entity types Path - Link - Queue anf Graph tensors
+> Output: per-flow delay = queue_delay + transmission_delay
+
+# Step 2 — DONE: Run predict.py on all traffic models
+Saved predictions_delay_*.npy, labels_delay_*.npy, flow_counts_delay_*.npy for all 5 TMs:
+
+Traffic Model	    Mean MAPE	Median MAPE
+constant_bitrate    	4.29%	1.15%
+onoff	                2.74%	0.99%
+autocorrelated	        2.46%	0.93%
+modulated	            5.26%	3.82%
+all_multiplexed	        4.53%	2.48%
+
+> Matches paper's reported MAPE. Model validated.
+
+
+# Step 3 — SHAP Plan
+The challenge: RouteNet-Fermi takes dict inputs with ragged/graph tensors — standard SHAP can't perturb them directly. The strategy:
+
+Focus: 12 scalar path features per flow (the traffic, packets, etc.)
+Method: shap.KernelExplainer — model-agnostic, works via perturbation
+Wrapper: flatten one flow's 12 features, call full model with that flow's features replaced
+Dataset: constant_bitrate (simplest, cleanest — good for baseline)
+Samples: 40 flows (20 low-delay + 20 high-delay)
+Background: median of 50 reference flows
+
+# Used take(200) on test set because full set was too slow for SHAP iteration 
+
+## After the execution of predict.py for delay prediction intraffic_models
+
+Ran predict.py on 200 test samples × 5 traffic models = 54,400 flows each.
+Saved predictions, labels, flow_counts as .npy files.
+
+MAPE results:
+| Traffic model      | Mean  | Median |
+|--------------------|-------|--------|
+| constant_bitrate   | 4.29% | 1.15%  |
+| onoff              | 2.74% | 0.99%  |
+| autocorrelated     | 2.46% | 0.93%  |
+| modulated          | 5.26% | 3.82%  |
+| all_multiplexed    | 4.53% | 2.48%  |
+
+Model is performing as expected (matches paper's reported MAPE).
+modulated is hardest, autocorrelated is easiest.
+
+
+# Applying SHAP to delay(one choosen sub-dataset) traffic model only; extension to jitter/loss is easy
+
+# Use constant_bitrate sub-dataset when Appling Graph SHAP on the pretrained model and Analyzing the imporatance scores of the inputs
+# why ? it's the cleanest single-model baseline
+
+
+# @RoutNet-Fermi > @traffic_models > delay > constant_bitrate