Awesome_OL: A General Toolkit for Online Learning Approaches

Welcome to Awesome_OL, your comprehensive toolkit for online learning strategies and classifiers! This repository provides a collection of state-of-the-art strategies and classifiers for online active learning (OAL) and online semi-supervised learning (OSSL). Whether you're a researcher, practitioner, or enthusiast in machine learning, this toolkit offers valuable resources and implementations to enhance your projects.

OAL Strategies:

Explore a variety of online active learning strategies located in the OAL_strategies folder:

Recent Progress:

Strategy	Description	Reference	Code Source	Year	Journal/Conference
CogDQS	A dual-query strategy using Ebbinghaus’s law of human memory cognition, enabling experts to annotate the most representative samples. It employs a fixed uncertainty strategy for auxiliary judgment.	Paper	NA	2023	IEEE Transactions on Neural Networks and Learning Systems
DSA-AI	A dynamic submodular-based learning strategy with activation interval for imbalanced drifting streams, which aims to address the challenges posed by concept drifts in nonstationary environments.	Paper	Link	2024	IEEE Transactions on Neural Networks and Learning Systems
MTSGQS	A memory-triggered submodularity-guided query strategy that evaluates sample value through residual analysis and limited retraining, effectively addressing imbalanced data stream issues.	Paper	NA	2023	IEEE Transactions on Intelligent Transportation Systems
DMI-DD	A query strategy that evaluates chunk-level sample values based on model explanations.	Paper	Link	2024	IEEE Transactions on Cybernetics

OAL Classifiers:

Classifier	Description	Reference	Code Source	Year	Journal/Conference
ROALE-DI	A reinforcement online active learning ensemble for drifting imbalanced data streams, which combines uncertainty and imbalance strategies to evaluate sample value.	Paper	Link	2022	IEEE Transactions on Knowledge and Data Engineering
OALE	An online active learning ensemble framework for drifting data streams based on a hybrid labeling strategy that includes an ensemble classifier and active learning strategies	Paper	NA	2019	IEEE Transactions on Neural Networks and Learning Systems

Baseline Strategies:

Strategy	Description	Reference	Code Source	Year	Journal/Conference
RS	Random Sampling (RS) serves as a simple baseline for active learning, randomly selecting data samples for labeling without considering their informativeness.	NA	NA	NA	NA
US_fix	Uncertainty Sampling with Fixed Threshold (US_fix) selects samples with uncertainty scores exceeding a fixed threshold for labeling, effectively targeting uncertain regions of the data space.	Paper	NA	2014	IEEE Transactions on Neural Networks and Learning Systems
US_var	Uncertainty Sampling with Variable Threshold (US_var) dynamically adjusts the uncertainty threshold based on model confidence and dataset characteristics, offering improved sample selection flexibility and performance in dynamic environments.	Paper	NA	2014	IEEE Transactions on Neural Networks and Learning Systems

OSSL Classifiers:

Discover online semi-supervised learning classifiers in the OSSL_strategies folder:

Recent Progress:

Classifier	Description	Reference	Code Source	Year	Journal/Conference
OSSBLS	An online semi-supervised BLS method with a loss function incorporating static anchor points.	Paper	NA	2021	IEEE Transactions on Industrial Informatics
ISSBLS	An online semi-supervised BLS method that ignores the relationship between historical data.	Paper	NA	2021	IEEE Transactions on Industrial Informatics

Baseline Strategies:

Classifier	Description	Reference	Code Source	Year	Journal/Conference
SOSELM	A classic online semi-supervised learning method based on extreme learning machines.	Paper	NA	2016	Neurocomputing

Supervised Classifiers:

Find various online learning classifiers in the classifer folder:

Baseline Strategies:

Classifier	Description	Reference	Code Source	Year	Journal/Conference
OLI2DS	An online learning algorithm for imbalanced data streams that tackles dynamically evolving feature spaces and imbalances and empirical risk minimization using dynamic cost strategies.	Paper	Link	2023	IEEE Transactions on Knowledge and Data Engineering
DES	An online ensemble learning method designed to adapt to data drift in streams with class imbalance, employing an improved Synthetic Minority Oversampling TEchnique (SMOTE) concept.	Paper	Link	2024	IEEE Transactions on Neural Networks and Learning Systems
BLS-W	An online learning method based on the standard BLS architecture, utilizing the Sherman–Morrison Woodbury formula for incremental updates.	Paper	Link	2023*	IEEE Transactions on Cybernetics
IWDA	A novel learner-agnostic algorithm for drift adaptation, which estimates the joint probability density of input and target for the incoming data. As soon as drift is detected, it retrains the learner using importance-weighted empirical risk minimization.	Paper	Link	2023*	IEEE Transactions on Neural Networks and Learning Systems
ACDWM	An adaptive chunk-based incremental learning method is proposed for handling imbalanced streaming data with concept drift, utilizing statistical hypothesis tests to dynamically select chunk sizes for assessing classifier stability.	Paper	Link	2020	IEEE Transactions on Neural Networks and Learning Systems
ARF	An advanced online ensemble learning method that addresses changing data streams by integrating effective resampling methods and adaptive operators with ADWIN.	Paper	Link	2017	Machine Learning
SRP	An ensemble method specially adapted to stream classification which combines random subspaces and online bagging.	Paper	Link	2019	ICDM

Baseline Strategies:

Classifier	Description	Reference	Code Source	Year	Journal/Conference
OSELM	An online sequential extreme learning machine model, which tries to iteratively update with the structure of extreme learning machines without the drift detection and adaption technique.	Paper	Link	2006	IEEE Transactions on Neural Networks

The following details are summarized for such implemented methods:

Method	OAL Strategy	Classifier	Binary Classification	Multi-class Classification	Concept Drift Adaptation	Ensemble
ROALE-DI	✅	✅	✅	✅	✅	✅
CogDQS	✅		✅	✅	✅
DSA-AI	✅		✅	✅	✅
DMI-DD	✅		✅	✅	✅
MTSGQS	✅		✅	✅	✅
RS	✅		✅	✅
US-fix	✅		✅	✅
US-var	✅		✅	✅
OLI2DS		✅	✅		✅
IWDA		✅	✅	✅	✅	✅
DES		✅	✅		✅	✅
ACDWM		✅	✅		✅	✅
SRP		✅	✅	✅	✅	✅
ARF		✅	✅	✅	✅	✅

Datasets:

The datasets folder contains .csv files structured with attributes, headers, and labels, catering to the needs of various strategies and classifiers.

Visualization:

The visualization folder contains implementations for visualizing metrics such as accuracy (acc), macro F1 score, and other relevant performance measures.

Utility:

• utils.py: This component file serves as the interface between classifiers and strategies, facilitating seamless interaction within the toolkit.

Implementation:

The specific implementations are encapsulated into a unified form. Further technical details and improvements can be explored within each strategy or classifier.

Environment Setup:

Before using this library, please ensure that you have the following essential packages and their corresponding versions installed.

Package	Version
numpy	1.21.6
matplotlib	3.2.2
scikit-learn	0.22.1
scikit-multiflow	0.5.3
pandas	1.2.3
scipy	1.7.3

Alternatively, for your convenience, you can set up the required environment by running the following command:

conda env create -f env.yml

References:

Explore related resources and inspiration at:

• GitHub - deep-active-learning
• GitHub - scikit-multiflow

---

Note

We hope this toolkit serves as a valuable asset in your online learning endeavors. Our team at the THUFDD Research Group, led by Prof. Xiao He and Prof. Donghua Zhou in the Department of Automation at Tsinghua University, is dedicated to fostering innovation and excellence in machine learning for industrial applications.

Your feedback, questions, and contributions are invaluable to us. Whether you have suggestions for improvements, encounter issues, or wish to collaborate on enhancements, we welcome your participation. Together, we can continue to refine and expand this toolkit to empower researchers, practitioners, and enthusiasts in the field.

Please feel free to reach out to us via email with Zeyi Liu and Songqiao Hu. Here's to a fruitful learning journey!