How conti748/wyze-rule-recommendation works
Project type and purpose
• Type: A machine-learning pipeline (a full-stack ML application, focused on training and inference).
• Purpose: Given a user’s smart-home device graph, propose new automation rules by casting it as a graph link-prediction task using a GNN (PyTorch-Geometric).
File structure
• Top-level scripts: train.py, test.py, test.py, test.py handle end-to-end training and evaluation.
• Configuration: cfg/ holds device/rule encodings and training hyperparameters (training.yaml).
• Data: data/ (inputs), models/ (checkpoints), results/ (output artifacts).
• Core modules:
• modules/data_pipeline/GraphDataset.py and RulesDataset.py build per-user graph objects.
• modules/utils/ holds sampling, decoding, general utilities.
• modules/LinkPredictor.py defines the GNN architecture.
• Requirements: requirements.txt declares dependencies (PyTorch, PyTorch-Geometric, HuggingFace Hub).
Extract additional insights from README
• External dependency: downloads the “wyzelabs/RuleRecommendation” dataset from HuggingFace Hub.
• Problem framing: each user’s devices = nodes, rules = directed, typed edges. Node features: one‐hot device type. Edge features: trigger/action states (45×47).
• Model: embeds states and node types → two SageConv layers → concatenated embeddings → classification head (sigmoid).
• Training: positive (leave-one-out) & negative sampling → BCE loss.
• Evaluation: compute MRR on public/private splits.
Main components and interactions
a. Data Ingestion & Configuration
• External Dataset (HuggingFace Hub)
• Local config files (cfg/)
→ feeds into
b. Data Pipeline
• GraphDataset: reads rules + device lists → builds PyG Data objects (nodes, edges, features)
• RulesDataset: helper for batching/testing
• Sampling Utils: positive/negative sampling modules
→ yields training/test batches
c. Model
• LinkPredictor (GNN):
• Embedding layers for trigger/action states
• Node feature concatenation (node type + aggregated edge embeddings)
• 2×SageConv layers
• Classification head → edge probability scores
→ outputs per‐pair link scores
d. Training & Evaluation
• train.py: loads config, instantiates datasets & model, runs training loop, logs loss/metrics, stores best_model.ckpt in models/
• test.py: loads checkpoint, loads dataset split (public/private), generates top-50 rule suggestions, computes MRR → stores in results/
e. Storage & Artifacts
• models/: serialized model checkpoints
• results/: evaluation outputs
• docs/: architecture slides & images
• Clear labels for each component and sub-component