T-Bank introduced the Perseus framework, designed for processing heterogeneous user event sequences and supporting cross-domain scenarios. The tool is intended for machine learning tasks such as candidate generation, ranking, classification, and regression.
The framework utilizes the T-ECD dataset, which was released in September 2025 and presented at the KDD-2026 conference in South Korea. This dataset includes data from T-Bank services like Cashback, Shopping, and Supermarkets.
ML engineers manage the workflow by loading event data into an Event Hub, preparing data representations, and creating YAML configuration files. The framework automates subsequent operations, including model training and inference.
Perseus supports various backbone architectures for sequence modeling, including modern_bert, bert, ligr, danet, hstu, and mamba. The system processes user history and context to generate predictions based on defined targets and artifacts.
