New LLM-based approach improves recommendation accuracy through coordinated indexing

Large language models' strong grasp of meaning and reasoning are now being harnessed to build recommendation systems that can better predict what users want to see or buy next. A research team has developed ED² (End-to-End Dual Dynamic), a system that addresses key limitations in how LLMs are currently applied to sequential recommendation tasks.

The Problem with Current Approaches#

Today's LLM-based recommendation systems typically handle index generation and sequential recommendation as separate steps. This fragmentation means that semantic insights from language models don't fully inform the recommendation logic. Additionally, most systems overlook user-specific details, missing the chance to detect complex patterns in how individual users interact with items over time.

How ED² Works#

The dual dynamic index mechanism unites index generation and recommendation into a single, streamlined pipeline. This allows the system to leverage both semantic knowledge and collaborative signals more effectively. To help the LLM grasp these dual indices, the team introduced a multigrained token regulator that creates supervisory signals by aligning the model's understanding across different levels of representation.

The system was also trained on custom user collection data and novel instruction-tuning tasks specifically designed to capture sophisticated user-item interaction patterns.

Results#

Testing on three public datasets showed substantial gains: ED² achieved an average improvement of 19.62% in Hit-Rate (measuring recommendation accuracy) and 21.11% in NDCG (a ranking quality metric) compared to existing approaches.