Remediating data drifts and re-establishing ML models

Abstract

Machine learning models are significantly used in organizations for critical business processes and decision-making. The inputs and predictions from these models might drift over time due to several reasons, causing unanticipated behaviour and performance issues. COVID-19 has a significant impact on business resulting in huge dispersions of the data with unstable characteristics and irregular spikes at various time steps. When the data set experiences significant drifts due to the changes in their statistical properties, continuing with existing models based on historical data will lead to poor decision outcomes. It's hard to anticipate all significant future events and create an ML model which can withstand black swan events. Therefore, it's imperative to have an automated process for continuous monitoring in the ML lifecycle to identify the potential issues using drift detection methodologies with proper actions to avoid false predictions in silence. This study is focused on time series where data stream is represented as observations at regular intervals. Many dynamic real-world processes can be modelled using time series where data can flow quickly and change over time which is detrimental to time series analysis and forecasting. The objective of this study is to compare and evaluate different approaches for data drifts in time series models to retain its optimum performance and proposing an integrated approach for handling those drifts in real-time pipelines using timely alerts and re-establishing models.