MLOps Pipelines for Continuous Deployment of Recommendation Systems in Retail
Author(s): Udit Agarwal, Aditya Gupta
Publication #: 2601004
Date of Publication: 07.01.2025
Country: United States
Pages: 1-7
Published In: Volume 11 Issue 1 January-2025
DOI: https://doi.org/10.62970/IJIRCT.v11.i1.2601004
Abstract
The integration of machine learning (ML) models into production environments necessitates a functional operational framework to ensure robustness, scalability, and long-term maintenance. This paper reviews the Machine Learning Operations (MLOps) paradigm as an essential system for the continuous deployment (CD) of personalized recommendation systems (RecSys) within the fast-paced retail sector. MLOps unifies ML development with system operations, facilitating automation across key stages of the model lifecycle from data preparation to deployment and continuous monitoring.2 The framework addresses the unique challenges of retail RecSys, specifically mitigating model drift caused by continuously evolving user preferences and market trends.4 Key architectural components, notably the low-latency Feature Store, are analyzed for their role in maintaining training-serving consistency and enabling real-time inference.6 The paper also examines the critical role of Dynamic Data Management (DDM) integrated into the automated retraining pipeline, which uses data reduction and feature selection to ensure resource-efficient, adaptive model updates. Contemporary approaches to continuous deployment commonly utilize dual metric systems that correlates rank-aware evaluation metrics, such as Normalized Discounted Cumulative Gain (NDCG), with tangible business outcomes like Average Order Value (AOV) lift, alongside operational efficiency metrics such as Mean Time to Resolution (MTTR).9
Keywords: MLOps, Continuous Deployment, Recommendation Systems, Retail, Model Drift, Feature Store, Observability, CI/CD, Dynamic Data Management, NDCG.
Download/View Count: 70
Share this Article