🤖 Russian teams improve recommendation systems — three papers at ACM CIKM '26
Teams from T-Bank, MIPT, HSE, and Innopolis University presented three papers on recommendation systems at the ACM CIKM '26 conference in Rome (proceedings published on November 6, 2026). E²L³AE jointly trains a previously “frozen” LLM-embedding text encoder with a linear recommendation autoencoder, rather than separately: across five datasets (cosmetics, toys, groceries, office supplies, clothing), the model outperformed L³AE in all tests, achieving +20.3% in recall and +17.2% in ranking quality on office supplies. MoE-IGR (“Give Me MoE”) replaces the general update of the PinSAGE graph model with dynamic routing between “experts”: in cold-start scenarios on three real-world datasets, it achieved up to +16% Recall@20 and up to +34.4% NDCG@20 on new items. Mikhail Pechatov’s paper “Split Decisions” introduces the model-selection regret metric and shows that random train/test splits without accounting for chronology often select the wrong model for future prediction — in 17 out of 18 comparisons, such models performed worse (-49% on movies, -81% on short videos).
🌍 Cold-start is a bottleneck in industrial recommendation systems: new items and new users do not yet have interaction history. Both models address this with a verifiable mechanism. E²L³AE adapts LLM knowledge about items to the recommendation task — the fine-tuned encoder itself acts as a competitive semantic recommender. MoE-IGR places specialized “experts” in the graph model instead of general weights, and routing analysis showed their specialization on cold and “warm” items. Important caveat: the percentages cited are best points on individual datasets, not average results.
👤 All three papers are published openly under the CC BY license — links below. If you build recommendations yourself, the practical takeaway from “Split Decisions” is this: compare the model selected on a random split versus a global temporal split — the discrepancy in metrics will reveal your regret. For readers, this highlights two live industry trends: “LLM-text + classic collaborative model” combinations and “mixture of experts in graphs.” The legal status of user interaction history is not defined in the sources — in some jurisdictions, such data may be considered personal data.
Source 1: https://dl.acm.org/doi/10.1145/3799682.3839945 Source 2: https://dl.acm.org/doi/10.1145/3799682.3839967 Source 3: https://dl.acm.org/doi/10.1145/3799682.3839942