DGF 0.1.0

DGF 0.1.0 Release Notes

Features and improvements

• Temporal & Time-Series Support:

  • Added end-to-end time-series and temporal graph support in the high-level training API (temporal sampling for link prediction, time-series features in node prediction).
  • Modular time-series feature encoders (including 1D-CNN embedding in EmbedFeatureSet).
  • Automated temporal train/validation splitting (split_train_valid_temporal and automatic temporal splits).
  • Built-in temporal normalizers in AutoNormalizer (time delta extraction, sinusoidal encoding, and calendar feature normalizers).
  • Snapshot-to-temporal graph conversion utility (dgf.data.convert_snapshot_to_temporal).

• Sampling & Scalability:

  • Python API and workflows for GCP distributed offline graph sampling.
  • Temporal-aware sampling in the offline distributed sampler, including automated inference of edge timestamps from node timestamps.
  • Performance optimizations for the in-process sampler (reduced mutex contention, faster edge deduplication, and multi_visit=False option).
  • Support for sampling individual seed nodes multiple times, tracking #seed_id and #sample_id, and filtering seeds by feature values.

• Graph Formats & IO:

  • Added TFRecord and RecordIO container support alongside Parquet for graph storage and offline sampling pipelines.
  • Support for explicit file lists as dataset inputs in addition to sharded glob patterns.
  • Direct GCS-aware filesystem helpers (write_text, copy_dir, enhanced exists).

• Training, Evaluation & Inspection:

  • Device-agnostic model reloading (models trained on GPU can now be reloaded on CPU/TPU).
  • Enhanced model.describe() report with visual graph schema diagrams and captured training warnings/diagnostics.
  • Checkpoint persistence for PRNG random state and early stopping state.
  • Added Average Precision (AP) metric and missing label handling during evaluation.
  • Validation batch auto-splitting for large temporal subgraphs and configurable padding_margin.
  • Track and report skipped training examples, failing if excessive padding skips occur (>10%).

• Graph Algorithms & Ecosystem:

  • PyTorch Geometric (PyG) interoperability and training examples.
  • New traffic and temporal benchmark datasets (METR-LA, PEMS-BAY, Jena climate).

Bug fixes

• Fixed graph sampler crash/inconsistency when the same node was visited multiple times from different paths in a sampling plan.
• Fixed feature collision in AutoNormalizer where features sharing a name across distinct node sets (or matching the target feature name) were erroneously dropped.
• Fixed unintended 2D matrix conversion when ragged NumPy time-series happened to share identical lengths across all nodes.
• Fixed inconsistent caching behavior in prepare_from_existing_one for temporal link prediction datasets.
• Fixed handling of multi-dimensional feature shapes (e.g., (None, 2)) across IO readers.
• Fixed Graphland dataset cache collision by isolating cache paths per mask split.
• Fixed scheduler worker configuration (set_num_parlay_workers) and resolved contention timeouts in parallel algorithms.
• Added validation guards to catch and report NaN losses and features during high-level API training.