BCCP: An MLOps Framework for Self-Cleansing Real-Time Data Noise via Bayesian Cut-Off-Based Closest Pair Sampling

This paper aims to enable machine learning models in machine learning operations (MLOps) to automatically perform self-cleansing in real-time against noise without external intervention. Despite the significance and demand for MLOps, there is a lack of comprehensive research and experimentation involving various scenarios in operating the MLOps pipeline. In particular, exploring data corruption scenarios remains relatively under-explored. In response to this gap, we have implemented a comprehensive MLOps pipeline that covers representative computer vision tasks. We introduce novel techniques called Reuse Buffer and Bayesian cut-off-based closest pair (BCCP) sampling for image classification to mitigate performance degradation in scenarios with data corruption. Particularly, it enables effective control of data noise issues even in an unmanned setting through user feedback.