How aravindsankar28/dysat works
DySAT learns node embeddings from time-ordered graph snapshots, with IncSAT as an incremental alternative. The main workflow runs per-time-step training, prepares graph structure and random-walk context pairs, trains the selected attention model, then evaluates embeddings for link prediction and records experiment results. The graph preserves both model variants and their shared utilities. Some modules are indexed but unsampled; their connections are included only where the README or an excerpt supports them.
Main entry points: train.py DySAT training · train_incremental.py IncSAT training · preprocess.py Graph preprocessing · run_script.py Experiment runner
Components 17
Experiment orchestration
- Experiment runner
run_script.py - Experiment settings
flags.py - DySAT training
train.py - IncSAT training
train_incremental.py
Graph preparation
- Graph snapshots
data/Enron_new/graphs.npz - Graph preprocessing
utils/preprocess.py - Walk contexts
utils/utilities.py - Random walk engine
utils/random_walk.py - Node minibatches
utils/minibatch.py - Incremental batches
utils/incremental_minibatch.py
Embedding models
- DySAT network
models/DySAT/models.py - DySAT attention layers
models/DySAT/layers.py - Weight initializers
models/DySAT/inits.py - IncSAT network
models/IncSAT/models.py - IncSAT attention layers
models/IncSAT/layers.py
Evaluation
- Link prediction
eval/link_prediction.py
Other components
- Researcher
Connections 20
- Researcher → Experiment runner: starts run
README.md - Experiment runner → DySAT training: launches
README.md - Experiment runner → IncSAT training: launches
README.md - DySAT training → Experiment settings: reads settings
train.py - IncSAT training → Experiment settings: reads settings
train_incremental.py - DySAT training → Graph preprocessing: loads graphs
train.py - IncSAT training → Graph preprocessing: loads graphs
train_incremental.py - Graph preprocessing → Graph snapshots: reads snapshots
utils/preprocess.py - Graph preprocessing → Walk contexts: generates contexts
utils/preprocess.py - Walk contexts → Random walk engine: runs walks
utils/utilities.py - DySAT training → Node minibatches: builds batches
train.py - IncSAT training → Incremental batches: builds batches
train_incremental.py - DySAT training → DySAT network: trains model
train.py - IncSAT training → IncSAT network: trains model
train_incremental.py - DySAT network → DySAT attention layers: builds attention
models/DySAT/models.py - DySAT attention layers → Weight initializers: uses initializers
models/DySAT/layers.py - IncSAT network → IncSAT attention layers: builds attention
models/IncSAT/models.py - IncSAT attention layers → Weight initializers: uses initializers
models/IncSAT/layers.py - DySAT training → Link prediction: evaluates embeddings
train.py - IncSAT training → Link prediction: evaluates embeddings
train_incremental.py