Toss TASTE Cuts Model Training to One Day and Automates Comparative Evaluation
·2026.11.17 16:14
Key point
By automating four comparative evaluation processes, including feature count and model structure, the system secures decision-making time for MLEs.
Details
- Reduced model training time from two weeks to one day using the TASTE framework and automated the four comparative evaluation processes required before model adoption into the pipeline
- Used the auto_k algorithm to determine the feature count at the point where performance no longer improves, reducing features by 84% while increasing top 5% recall by approximately 1.7x with no performance loss
- Compared AITM deep learning and tree models in the funnel model using the same evaluation set, confirming that deep learning has an AUROC advantage of approximately 1.4% when data is sufficient, while trees dominate when positive cases are scarce
- Automatically generated evaluation reports by specifying only the OOT ground truth table, reducing model evaluation time from 30 minutes to 1 minute
- Stopped score loading when feature missingness exceeded 10% during inference and improved raw data processing methods by detecting future information leakage in mid-month data using standalone AUROC
- Maintained consistency in the list of top-scoring customers by pinning the feature list and hash values from training time instead of using model files
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.