CmnRec: Sequential Recommendations With Chunk-Accelerated Memory Network

Recently, Memory-based Neural Recommenders (MNR) have demonstrated superior predictive accuracy in the task of sequential recommendations, particularly for modeling long-term item dependencies. However, typical MNR requires complex memory access operations, i.e., both writing and reading via a controller (e.g., RNN) at every time step. Those frequent operations will dramatically increase network training time, resulting difficulty being deployed on industrial-scale recommender systems. In this paper, we present novel general Chunk framework to accelerate significantly. Specifically, our divides proximal information units into chunks, performs certain steps, whereby number can be greatly reduced. We investigate two ways implement effective chunking, PEriodic Chunk (PEC) Time-Sensitive (TSC), preserve recover important recurrent signals sequence. Since chunk-accelerated models take account more than that from single timestep, it alleviate influence noise user-item interaction sequence large extent, thus improve stability MNR. way, proposed chunk mechanism lead not only faster prediction, but even slightly better results. The experimental results three real-world datasets (weishi, ml-10M ml-latest) show notably reduces running with up 7x & 10x inference MNR, meantime achieves competitive performance.