v0.4.0

[kayzliu]

We are excited to announce the final pre-alpha release, PyGOD v0.4, which marks a major milestone in our development. Following bug fixes and minor improvements, we plan to release v1.0. Your feedback and suggestions are appreciated. ⚠️ Please note that this version is NOT forward compatible and some APIs have changed. Here are the major changes in this release:

Enhanced Base Class

• Detector: base class for all detectors.
• DeepDetector: base class for all deep learning based detectors.

Simplied APIs

• Removed predict_proba and predict_confidence.
• Use predict(return_prob=True, return_conf=True) instead.

Modularized Detectors

We now introduce multiple modules to improve the code reusability and extendibility.

• nn: all base models inherit torch.nn.Module
• nn.encoder:
• nn.decoder:
• nn.functional: loss function, etc.
  Also, we changed the name of several modules to improve the clarity.
• models→detector
• metrics→metric

More Utility Functions

• to_edge_score: edge outlier score converter
• to_graph_score: graph outlier score converter
• init_detector: detector initializer
• init_nn: neural network initializer

Updated Requirements

• PyGOD now requires Python 3.8+
• PyTorch 2.0 and PyG 2.3.0 support
• Enabled model compile via detector(compile_model=True) (beta)

And Many More

• More comprehensive test coverage (almost 100%)
• Reorganized documentation for better readability
• Merge MLPAE and GCNAE to GAE
• Most of the deep detectors support specifying various backbone from PyG
• Retrieve learned embedding from fitted deep detectors with save_emb=True by detector.emb

---

This discussion was created from the release v0.4.0.