Towards Scalable Meta-Learning of near-optimal Interpretable Models via Synthetic Model Generations
The work introduces a scalable method for generating synthetic pre-training data to enable meta-learning for decision trees. It uses a Structural Causal Model (SCM) pipeline to create high-quality, causal synthetic datasets and corresponding near-optimal decision trees. This allows a MetaTree model to be pre-trained on these synthetic examples, which can then be used for inference on real-world data.