Scalable Sequential Recommendation under Latency and Memory Constraints

TL;DR

This work tackles scalable sequential recommendation under strict latency and memory constraints, where attention-based transformers struggle with quadratic complexity. It proposes HoloMambaRec, a hybrid architecture that binds item IDs with discrete attributes via circular convolution (holographic binding) and processes the bound sequence with a shallow selective state space encoder, achieving linear-time processing and constant-time recurrence. The approach delivers consistent gains over SASRec and competitive results with GRU4Rec under a tight 10-epoch budget, while maintaining modest memory usage and forward compatibility with temporal bundling and multi-attribute extensions. The findings demonstrate the practicality of metadata-aware, long-horizon modeling in production environments and highlight the potential benefits of richer attribute signals and end-to-end bundling in future work.

Abstract

Sequential recommender systems must model long-range user behavior while operating under strict memory and latency constraints. Transformer-based approaches achieve strong accuracy but suffer from quadratic attention complexity, forcing aggressive truncation of user histories and limiting their practicality for long-horizon modeling. This paper presents HoloMambaRec, a lightweight sequential recommendation architecture that combines holographic reduced representations for attribute-aware embedding with a selective state space encoder for linear-time sequence processing. Item and attribute information are bound using circular convolution, preserving embedding dimensionality while encoding structured metadata. A shallow selective state space backbone, inspired by recent Mamba-style models, enables efficient training and constant-time recurrent inference. Experiments on Amazon Beauty and MovieLens-1M datasets demonstrate that HoloMambaRec consistently outperforms SASRec and achieves competitive performance with GRU4Rec under a constrained 10-epoch training budget, while maintaining substantially lower memory complexity. The design further incorporates forward-compatible mechanisms for temporal bundling and inference-time compression, positioning HoloMambaRec as a practical and extensible alternative for scalable, metadata-aware sequential recommendation.

Figures (2)

• Figure 1: HoloMambaRec architecture
• Figure 2: Hit Rate and NDCG learning curves on Amazon Beauty and MovieLens-1M