archi829/heterogat-v2

HeteroGAT v2 — Bottleneck Oracle Model

Overview

HeteroGAT v2 is a heterogeneous graph attention network designed to detect bottlenecks and anomalous patterns in system trace graphs.

It operates on graph-structured data derived from execution traces and learns node/edge relationships to identify suspicious or high-impact components.

---

Model Details

• Architecture: Heterogeneous Graph Attention Network (HeteroGAT)
• Framework: PyTorch + PyTorch Geometric
• Model File: heterogat_v2.pt

Key Components

• Multi-head attention on heterogeneous node types
• Learned embeddings for different entity types
• Graph-level aggregation for prediction

---

Dataset

Trained on: 👉 https://huggingface.co/datasets/archi829/bottleneck-oracle-data

Input

• Preprocessed graph objects (graphs_v2.pkl)
• Nodes represent system entities (services, processes, etc.)
• Edges represent interactions/dependencies

---

Training Pipeline