New Methods for Enhancing Recommendation Quality for New Items
Russian researchers have developed two new approaches to solving the cold start problem in recommendation systems, improving the quality of recommendations for new products and content by up to 35%. These methods enable algorithms to effectively handle objects that lack a history of user interactions.
E2L3AE Model
The first approach utilizes the E2L3AE model, which overcomes the limitations of the basic L3AE by jointly training the text encoder with the recommendation system. Unlike separate training, this method allows the encoder to better understand key item features based on user interaction data.
Testing was conducted across five datasets: cosmetics and personal care, toys and games, groceries, office supplies, and clothing. Results demonstrated E2L3AE's superiority over L3AE in all tests: recommendation recall increased by 20.3%, and ranking quality improved by 17.2% on the office supplies dataset.
This approach is particularly effective for items with detailed descriptions, such as cosmetics, clothing, or toys. The model addresses the cold start problem by leveraging item descriptions even when there are minimal user interactions. The text encoder can be used independently from the recommendation model, where it outperforms standard text models without further fine-tuning.
MoE-IGR Architecture
The second innovation is the MoE-IGR architecture, which combines a graph model with a Mixture of Experts (MoE). The system autonomously determines which neural network expert or combination thereof is best suited for a specific item. This enables recommendations to be tailored to the specifics of each object.
Using MoE-IGR allowed for a 34.5% improvement in the ranking quality of new items on the MovieLens dataset compared to the PinSAGE method. The architecture demonstrates high efficiency when handling new data absent from the training history.
These results were presented by the research team in Rome at the ACM CIKM conference. As one of the top-tier A-level scientific conferences in information retrieval and processing, this presentation underscores the significance of these achievements in machine learning and recommendation systems.









