Title:	DRIFT-DRIVEN REGRESSION FOR PREDICTING THE EVOLUTION OF PANDEMICS
Author(s):	Khaled Jouini and Ouajdi Korbaa
ISBN:	978-989-8704-53-5
Editors:	Paula Miranda and Pedro Isaías
Year:	2023
Edition:	Single
Keywords:	Concept Drift, Incremental Learning, Collaborative Learning, Pandemic Forecasting
Type:	Full
First Page:	77
Last Page:	84
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	Pandemics will never cease to emerge and threaten public health and the global economy. Predicting the evolution of a pandemic is of paramount interest as it enables policymakers understand the potential spread of a virus and make informed decisions to mitigate its impact. Concept drifts refer to situations where the relationship between the input features (model input) and the learning targets (model output) changes or evolves over time. Concept drifts are common in pandemic curves, not only because new variants appear over time, but also due to factors such as seasonality, policy responses to the pandemic, and changes in the way the disease is treated. Without proper intervention, the accuracy of conventional (batch) machine learning models will deteriorate after drift occurs, since they were trained on outdated data. Incremental learning is an alternative to batch learning, where the training phase never ends, and the model is incrementally updated as new data becomes available. While incremental models are in general less accurate than batch models, they adapt better to concept drifts as they are continuously refined using the most recent data. Batch and incremental learning are often considered as two distinct and mutually exclusive approaches (Montiel et al., 2018a). In this work we propose CDR (Collaborative Drift- Driven Regression), a novel collaborative regression strategy where incremental and batch regressors work together to complement each other's strengths, ergo the overall predictive performance. Experiments conducted on COVID-19 pandemic data, show that CDR is an efficient collaborative learning strategy that yields better results than the underlying batch and incremental models used separately.
	Go Back