// DESCRIPTION
Trie-Based Mixture-of-Experts Recommendation
EncRec introduces a novel trie-based routing mechanism for Mixture-of-Experts (MoE) recommendation systems, replacing the standard top-k gating with a hierarchical prefix trie that maps user interaction sequences to expert specializations. This structural approach enables more interpretable and stable expert assignment compared to learned softmax gating.
The trie structure is constructed from user behavior patterns, where shared prefixes in interaction sequences naturally cluster users with similar interests. Each trie node is associated with a subset of experts, and routing follows the longest matching prefix, providing a deterministic and cache-friendly dispatch mechanism that eliminates the load balancing issues common in standard MoE architectures.
On large-scale recommendation benchmarks (Amazon Reviews, MovieLens-25M, Yelp), EncRec achieves 3-7% improvement in NDCG@10 over dense baselines and 1-3% over standard MoE approaches, while reducing expert load imbalance by 60%. The trie structure also enables efficient online updates as user behavior evolves, without requiring full model retraining.
The architecture is designed for production serving with sub-millisecond routing latency and natural support for A/B testing at the expert level, enabling fine-grained experimentation in live recommendation systems.
// HIGHLIGHTS
- 3-7% NDCG@10 improvement over dense baselines on large-scale recommendation benchmarks
- Trie-based routing eliminates MoE load balancing issues with 60% reduction in expert load imbalance
- Sub-millisecond deterministic routing latency, cache-friendly for production serving
- Online trie updates without full model retraining as user behavior evolves
- Evaluated on Amazon Reviews, MovieLens-25M, and Yelp at scale