Multi-Probe Zero Collision Hash (MPZCH): Mitigating Embedding Collisions and Enhancing Model Freshness in Large-Scale Recommenders

TL;DR

The paper tackles embedding collisions and staleness in large-scale recommenders by introducing MPZCH, a Multi-Probe Zero Collision Hash that employs a two-pass linear probing scheme with auxiliary identities and metadata tensors to manage slot occupancy and TTL-based eviction. By combining GPU-accelerated kernels with per-feature TTL policies and optimizer resets on eviction, MPZCH achieves collision-free user embeddings and enhanced freshness for item embeddings, without sacrificing training or inference performance. Empirical results in production show significant gains in embedding quality and content freshness, including concrete improvements in Normalized Entropy for users and a measurable uplift in impressions for new content, with additional gains in semantic coherence and learning stability. The method integrates cleanly with TorchRec, supports online streaming updates, and is released as open-source to benefit the broader community.

Abstract

Embedding tables are critical components of large-scale recommendation systems, facilitating the efficient mapping of high-cardinality categorical features into dense vector representations. However, as the volume of unique IDs expands, traditional hash-based indexing methods suffer from collisions that degrade model performance and personalization quality. We present Multi-Probe Zero Collision Hash (MPZCH), a novel indexing mechanism based on linear probing that effectively mitigates embedding collisions. With reasonable table sizing, it often eliminates these collisions entirely while maintaining production-scale efficiency. MPZCH utilizes auxiliary tensors and high-performance CUDA kernels to implement configurable probing and active eviction policies. By retiring obsolete IDs and resetting reassigned slots, MPZCH prevents the stale embedding inheritance typical of hash-based methods, ensuring new features learn effectively from scratch. Despite its collision-mitigation overhead, the system maintains training QPS and inference latency comparable to existing methods. Rigorous online experiments demonstrate that MPZCH achieves zero collisions for user embeddings and significantly improves item embedding freshness and quality. The solution has been released within the open-source TorchRec library for the broader community.

Figures (6)

• Figure 1: A simplified example of ID insertion, lookup, and collision handling.
• Figure 2: An example of kernel execution with TTL eviction policy. Probing details are omitted for clarity, with arrows pointing to the final result slots.
• Figure 3: An example of kernel execution with LRU eviction policy. Probing details are omitted for clarity, with arrows pointing to the final result slots.
• Figure 4: The MPZCH module is co-sharded with the associated embedding table, ensuring that operations are executed locally within each shard.
• Figure 5: Top: Video embeddings from selected creators in the baseline model appear dispersed and uncorrelated. Bottom: With MPZCH, embeddings for the same creators exhibit significantly higher correlation and tighter clustering.