Graph financial fraud classification¶

Access this AI accelerator on GitHub

This accelerator demonstrates the end-to-end implementation of a Graph Isomorphism Network (GIN) as a custom model task in DataRobot. It serves multiple purposes, including model development, local testing, and integration with DataRobot’s custom model framework.

The primary objectives of this accelerator are to demonstrate the complete pipeline for adding a custom graph-based task to DataRobot, and to outline DataRobot's custom model hooks implementation. These hooks include:

• transform: Preprocess JSON graph data into DGL format. DRUM hooks automatically call transform before executing the fit and score hooks.
• fit: Train the GIN model and implicitly save it.
• score: Score the data in prediction mode.
• load_model: Load pre-trained models.

The accelerator's workflow includes three major steps:

1. Manually implement the DRUM process. Follow a step-by-step breakdown of data transformation, model training, and prediction via DRUM.

2. Test models locally with DRUM to validate hooks before integrating the model with DataRobot.

3. Implement a custom task in DataRobot with automated validation and threshold optimization.