The "T-Tekhnologii" Group has published the Perseus framework in open access, designed for training personalization models based on data from various services. The development allows combining user events into a single temporal sequence, forming a general understanding of customer behavior even in the absence of interaction history with individual products.
Perseus supports tasks such as recommendation search, ranking, classification, and regression. During training, a temporal data processing principle is applied: the model uses only events that occurred before a specified moment, which excludes the influence of "future" information. According to the developers, during testing, the quality of recommendations increased by 3–17%, and the accuracy of predicting the first basket grew by approximately 10% compared to the baseline approach.
At "T-Bank", the solution is already used for shopping services, analyzing responses to increased cashback offers, and predicting customer inquiries. The project is published on GitHub under the Apache 2.0 license and can be used for working with a single data source as well as for creating cross-domain personalization models.