[Paper Review] Intelligent Model Update Strategy for Sequential Recommendation

Zheqi Lv, Wenqiao Zhang+3 more|arXiv (Cornell University)|Feb 14, 2023

Recommender Systems and TechniquesComputer Science47 references7 citations

TL;DR

The paper proposes an Intelligent Device-Cloud Parameter Request Model (IDEAL) built atop the DC-CDR framework, introducing Mis-Recommendation Detector (MRD) and Distribution Mapper (DM) to balance update frequency and communication budget for device-cloud sequential recommendation.

ABSTRACT

Modern online platforms are increasingly employing recommendation systems to address information overload and improve user engagement. There is an evolving paradigm in this research field that recommendation network learning occurs both on the cloud and on edges with knowledge transfer in between (i.e., edge-cloud collaboration). Recent works push this field further by enabling edge-specific context-aware adaptivity, where model parameters are updated in real-time based on incoming on-edge data. However, we argue that frequent data exchanges between the cloud and edges often lead to inefficiency and waste of communication/computation resources, as considerable parameter updates might be redundant. To investigate this problem, we introduce Intelligent Edge-Cloud Parameter Request Model, abbreviated as IntellectReq. IntellectReq is designed to operate on edge, evaluating the cost-benefit landscape of parameter requests with minimal computation and communication overhead. We formulate this as a novel learning task, aimed at the detection of out-of-distribution data, thereby fine-tuning adaptive communication strategies. Further, we employ statistical mapping techniques to convert real-time user behavior into a normal distribution, thereby employing multi-sample outputs to quantify the model's uncertainty and thus its generalization capabilities. Rigorous empirical validation on four widely-adopted benchmarks evaluates our approach, evidencing a marked improvement in the efficiency and generalizability of edge-cloud collaborative and dynamic recommendation systems.

Motivation & Objective

• Motivate improved generalization of on-device recommendation models via dynamic, cloud-assisted parameter updates.
• Develop a device-cloud collaboration framework (DC-CDR) that decouples backbone and classifier to tailor dynamic layers for real-time data.
• Introduce MRD to decide when cloud-assisted parameter updates are needed, reducing unnecessary communication.
• Propose Distribution Mapper to quantify uncertainty in semantic understanding of data and guide updates.

Proposed method

• Define a DC-CDR optimization where the backbone is fixed after training and the classifier is generated by a cloud model from real-time data.
• Split the model into static backbone and dynamic layers to improve personalization and generalization on-device.
• Use a cloud generator model to produce dynamic-layer parameters for on-device inference based on real-time sequences.
• Train the cloud generator jointly with the static layers using global history data instead of sequentially optimizing static and cloud components.
• Introduce MRD to determine update necessity by creating a MRD dataset from historical data and using a two-stage training procedure with classification and KL-divergence losses.
• Use Distribution Mapper to assess uncertainty in the model’s understanding of data semantics and guide parameter updates.

Experimental results

Research questions

• RQ1 How can device-cloud collaboration be structured to maximize recommendation performance under budget constraints?
• RQ2 Can a Mis-Recommendation Detector accurately decide when cloud-supplied parameters should be requested for on-device models?
• RQ3 How does a Distribution Mapper quantify uncertainty to improve the parameter-update strategy in sequential recommendation?
• RQ4 What are the benefits of decoupling backbone and classifier into static and dynamic layers for on-device generalization?
• RQ5 What training procedure enables joint optimization of cloud-generated dynamic parameters with on-device static layers?

Key findings

• DC-CDR improves potential generalization by enabling dynamic on-device layers guided by cloud-generated parameters.
• MRD and DM are proposed to reduce unnecessary communication while maintaining recommendation quality under budget constraints.
• The framework decouples static backbone and dynamic classifier components to support personalized yet globally informed recommendations.
• MRD uses historical data to construct a MRD dataset and enables a two-stage training procedure with KL-divergence loss.
• DM evaluates uncertainty in semantic understanding to inform when parameter updates should occur.
• The proposed approach aims to achieve high communication revenue under varying device-cloud budgets.