Efficient Sequential Recommendation

• Efficient sequential recommendation is a class of techniques that delivers high-quality, low-latency predictions under constraints such as limited compute and massive item catalogs.
• It leverages methods like efficient attention architectures, compressed output heads, and hybrid sequence modeling to optimize both training and inference efficiency.
• These approaches balance recommendation accuracy with reduced memory and computational demands, enabling scalability even with billions of items and long user histories.

Efficient sequential recommendation denotes the class of modeling strategies that enable sequential recommenders to deliver high-quality, low-latency, and scalable predictions under realistic constraints, such as limited compute, large catalog sizes, and dynamic user histories. The challenge is to simultaneously maximize recommendation accuracy, minimize computation and memory requirements at both training and inference, and ensure the model is deployable at scale, even with billions of items or long temporal sequences. Recent research leverages innovations in efficient attention/memory architectures, compressed output heads, hybrid linear/nonlinear sequence modeling, and adapter- or token-based parameter efficiency to address these goals.

1. Design Principles and Efficiency Criteria

A sequential recommender is considered efficient if it satisfies several criteria:

• **Linear or near-linear