T-Technologies teams from the T-Bank ecosystem brought three works to the ACM RecSys 2026 conference in Minneapolis — this is the 20th ACM conference on recommender systems, taking place from September 27 to October 2, 2026. The main practical release is the open-source library Scikit-Rank: it wraps neural rankers in a scikit-learn-compatible fit/predict interface and allows substituting a neural network for CatBoost or LightGBM with almost no code changes. The second work is the cross-service personalization framework Perseus, which combines user signals from all company products into a single model, and the third is the ScaL³AE method, making recommendations with LLM enrichment applicable to million-item catalogs. Scikit-Rank code is already open on GitHub, ScaL³AE implementation is available anonymously, so the claimed results can be verified on your own data right now.

image
image
image

What happened

T-Technologies (T-Bank ecosystem) presented three works at ACM RecSys 2026 — the 20th ACM conference on recommender systems, taking place in Minneapolis from September 27 to October 2, 2026. The first is the open-source library Scikit-Rank, which is installed via pip install scikit-rank: it includes neural rankers DCN v2, FinalNet, FinalMLP, Destine, and TabM, built-in processing of numerical and categorical features, supports numpy, pandas, and polars, and offers more than 20 loss functions for training, including BPR, LambdaRank, ListMLE, and CORAL. The library is designed for CTR prediction and offer ranking. The second work is Perseus, a Python framework that builds a single common user event sequence from different services: views, purchases, searches, and transactions are collected together instead of logs from a single product. The third work is an article on ScaL³AE, “Scaling LLM-Enhanced Linear Autoencoders to Industrial-Size Catalogs”: the method replaces dense n×n matrices of linear autoencoders L³AE with a low-rank decomposition of the semantic similarity matrix and sparse approximation SANSA.

Context

RecSys is an annual ACM conference entirely dedicated to recommender systems; in 2026, its 20th edition takes place, and talks there undergo expert peer review. It is symptomatic that all three T-Technologies works address long-standing pain points of this market. First: production ranking — from offer selection to CTR prediction — has been held up by gradient boosting for years, as neural network models required their own training wrapper and were expensive to maintain. Second: in ecosystems with multiple applications, personalization is usually trained separately for each product, and user signals from neighboring services are lost. Third: when recommendations are enriched with LLM semantics, everything hits a memory wall — dense n×n item similarity matrices grow quadratically and stop fitting on million-item catalogs. Notably, some of the answers are not one-off experiments but tools that can be put into practice: a library, a framework, and a scalable method.

Why this matters for the industry

For the industry, the main point here is the sharp reduction in the cost of transitioning from boosting to neural ranking: for teams that have long been eyeing neural rankers but did not want to maintain custom PyTorch training, the barrier to entry drops to the level of replacing a classifier in a familiar API. The second meaning is ecosystem: Perseus shows a measurable gain from combining user signals from all company products into a single personalization model; according to the authors, the T-ECD metric increases from 16 to 39% across different domains. For banks and ecosystems with multiple applications, this is a ready-made pattern — one personalization for all services instead of separate models for each product, which can be reproduced and measured in-house. Finally, ScaL³AE removes the quadratic memory dependency of LLM-enhanced recommendations and makes linear autoencoders with LLM semantics applicable to million-item catalogs; according to the authors, the gain grows precisely on cold and sparse data, i.e., where classical approaches are weakest.

Why this matters for users

All of this can be verified on the day of publication. Scikit-Rank is installed with a single command pip install scikit-rank, works on CPU and CUDA, and training is launched via PyTorch and Hugging Face Accelerate with mixed precision and multi-GPU support; the scikit-rank/scikit-rank repository is open on GitHub. The fastest first step for a team is to run the library on its own ranking dataset and compare the result with working boosting in terms of quality, speed, and cost. ScaL³AE code is posted at anonymous.4open.science/r/scal3ae: it can be read and run, for example, to test the method on its own product catalog. A separate value for RecSys practitioners is a map of working solutions from a large Russian ecosystem: Scikit-Rank shows which architectures and which loss functions are actually used in its production, and Perseus at this stage is useful as a design template for designing its own cross-service data infrastructure.

What is still unknown / limitations

All key figures here are author-reported, and there is no external independent verification of them yet. The increase in the T-ECD metric from 16 to 39% relates to Perseus: from open materials, it does not follow what exactly stands behind it — recall@k, NDCG, or online A/B, and on what base the percentage is calculated; moreover, a significant part of the gain is probably explained by the very appearance of cross-domain signals, not a specific architecture. Perseus itself is formatted as an ACM RecSys 2026 demo talk — its goal is to show the feasibility of a prototype, not to provide a methodologically rigorous benchmark. ScaL³AE code is published only in anonymized form for review: it is suitable for reading and experiments, but not yet for industrial integration. Finally, for Scikit-Rank, the key question for the coming months is independent comparisons with gradient boosting on public CTR datasets: without them, it is unclear whether the library will become a working migration tool or remain a wrapper “for experiments.”

Sources

Author

Look at AI, editorial team