Scikit-rank: Neural Network Ranking from T-Bank

The T-Bank team released the Scikit-rank library, designed for neural network ranking. The tool ensures compatibility with the scikit-learn interface and uses standard fit and predict methods.

Process Automation

The library independently handles data preprocessing, feature encoding, model training, and inference building. This eliminates the need to rewrite the entire pipeline when switching from boosting to neural networks.

Scikit-rank supports working with data in numpy, pandas, and polars formats. Models can be trained on both CPUs and GPUs.

Available Architectures and Loss Functions

• DCN v2
• FinalNet
• FinalMLP
• Destine
• TabM

Several neural network architectures are available internally for ranking tasks. For training, you can choose a suitable loss function such as BPR, LambdaRank, ListMLE, or CORAL.

A presentation of the Scikit-rank project took place at the ACM RecSys 2026 international conference in the US. More details about the project can be found here: https://habr.com/ru/companies/tbank/articles/1088224/