Researchers at T-Technologies have introduced two new approaches to improve recommendation systems for items lacking user history. The E2L3AE model integrates recommendation algorithms with text analysis neural networks, improving recommendation recall by 20.3% and ranking quality by 17.2% compared to the L3AE baseline.
The second innovation, the MoE-IGR architecture, combines graph-based recommendation models with a Mixture of Experts approach. This system analyzes item characteristics to assign specialized neural network experts, resulting in a 35% improvement in ranking quality for new items compared to existing solutions.
The team also demonstrated that the methodology for model validation significantly impacts performance. Models selected without considering user interaction chronology performed up to 81% worse when tested against future data compared to those validated with chronological awareness.
E2L3AE is designed for services with detailed item descriptions, such as e-commerce platforms for cosmetics or office supplies, and its source code is available on GitHub. MoE-IGR is particularly suited for industries with rapidly expanding catalogs, including film, publishing, and restaurants.
