# Active learning for data streams: a survey

**Authors:** Davide Cacciarelli, Murat Kulahci  
**arXiv:** [2302.08893](https://arxiv.org/abs/2302.08893)  
**Date:** 2023-02-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.08893v4/

## Key points

- Online active learning with data streams aims to minimize costs by selecting informative real-time data points.
- Obtaining annotated data remains a challenge for training complex prediction and decision-making models, hindering AI integration into real-world applications like healthcare or autonomous driving.
- Current strategies in this field include uncertainty sampling, diversity sampling, query by committee, and reinforcement learning for online classification, regression, and semi-supervised learning.
- Further research is needed for online active linear regression models and advanced methods applicable to nonlinear models beyond linear bandits.
- Future directions include exploring model-agnostic approaches for regression models and developing single-pass online sampling strategies for dynamic data streams.

## Summary

, , , , 


Online active learning with data streams is a rapidly evolving field in machine learning that focuses on selecting the most informative data points in real-time to minimize the cost associated with collecting labeled observations. The increasing volume of data generated by modern applications has made it crucial to develop effective methods for learning from data streams continuously. However, the challenge lies in obtaining annotated data to train complex prediction and decision-making models, hindering the integration of artificial intelligence into real-world applications such as healthcare, autonomous driving, and industrial production. This comprehensive survey provides an overview of the current state-of-the-art strategies for online active learning with data streams. Various techniques based on uncertainty sampling, diversity sampling, query by committee, and reinforcement learning have been explored in contexts like online classification, regression, and semi-supervised learning. The analysis highlights the need for further research into online active linear regression models and advanced methods applicable to nonlinear models beyond linear bandits. Future directions in this field include investigating model-agnostic approaches for regression models and developing single-pass online sampling strategies for dynamic data streams. While ensemble models and batch-based approaches have been dominant in online classification, there is a growing interest in exploring methods that can handle continuous streams of data without requiring batch processing. Research efforts are also directed towards leveraging Bayesian optimization for active learning in nonlinear regression problems to enhance model performance. with  is a rapidly evolving field that aims to minimize costs by selecting informative real-time data points. However, obtaining annotated data remains a challenge for training complex prediction and decision-making models, hindering AI integration into practical applications like healthcare or autonomous driving. This survey provides an overview of current strategies including uncertainty sampling, diversity sampling, query by committee, and reinforcement learning for online classification, regression, and semi-supervised learning. Further research is needed for online active linear regression models and advanced methods applicable to nonlinear models beyond linear bandits. Future directions include model-agnostic approaches for regression models and single-pass online sampling strategies for dynamic data streams. While ensemble models and batch-based approaches dominate online classification, there is a growing interest in continuous stream methods without batch processing. Research also focuses on leveraging Bayesian optimization for active learning in nonlinear regression to enhance model performance. capabilities have made  from  crucial in machine learning. However, obtaining annotated data remains a challenge for training complex prediction and decision-making models, hindering AI integration into real-world applications like healthcare or autonomous driving. This survey provides an overview of current strategies including uncertainty sampling, diversity sampling, query by committee, and reinforcement learning for online classification, regression, and semi-supervised learning. Further research is needed for online active linear regression models and advanced methods applicable to nonlinear models beyond linear bandits. Future directions include model-agnostic approaches for regression models and single-pass online sampling strategies for dynamic data streams. While ensemble models and batch-based approaches dominate online classification, there is a growing interest in continuous stream methods without batch processing. Research also focuses on leveraging Bayesian optimization for active learning in nonlinear regression to enhance model performance. through  has become crucial with the increasing volume of data generated by modern applications. This comprehensive survey provides an overview of current strategies including uncertainty sampling, diversity sampling, query by committee, and reinforcement learning for online classification, regression, and semi-supervised learning in the context of . Further research is needed for online active linear regression models and advanced methods applicable to nonlinear models beyond linear bandits. Future directions include model-agnostic approaches for regression models and single-pass online sampling strategies for dynamic data streams.