diff --git a/joss.07226/10.21105.joss.07226.crossref.xml b/joss.07226/10.21105.joss.07226.crossref.xml
new file mode 100644
index 0000000000..07f0cd84eb
--- /dev/null
+++ b/joss.07226/10.21105.joss.07226.crossref.xml
@@ -0,0 +1,207 @@
+
+
+
+ 20250106152355-a75a657e37a0f78a34d58134370213ee59900a4f
+ 20250106152355
+
+ JOSS Admin
+ admin@theoj.org
+
+ The Open Journal
+
+
+
+
+ Journal of Open Source Software
+ JOSS
+ 2475-9066
+
+ 10.21105/joss
+ https://joss.theoj.org
+
+
+
+
+ 01
+ 2025
+
+
+ 10
+
+ 105
+
+
+
+ DeepRiver: A Deep Learning Library for Data Streams
+
+
+
+ Cedric
+ Kulbach
+
+ FZI Research Center for Information Technology, Karlsruhe, Germany
+
+ https://orcid.org/0000-0002-9363-4728
+
+
+ Lucas
+ Cazzonelli
+
+ FZI Research Center for Information Technology, Karlsruhe, Germany
+
+ https://orcid.org/0000-0003-2886-1219
+
+
+ Hoang-Anh
+ Ngo
+
+ AI Institute, University of Waikato, Hamilton, New Zealand
+
+ https://orcid.org/0000-0002-7583-753X
+
+
+ Max
+ Halford
+
+ Carbonfact, Paris, France
+
+ https://orcid.org/0000-0003-1464-4520
+
+
+ Saulo Martiello
+ Mastelini
+
+ Institute of Mathematics and Computer Science, University of São Paulo, São Carlos, Brazil
+
+ https://orcid.org/0000-0002-0092-3572
+
+
+
+ 01
+ 06
+ 2025
+
+
+ 7226
+
+
+ 10.21105/joss.07226
+
+
+ http://creativecommons.org/licenses/by/4.0/
+ http://creativecommons.org/licenses/by/4.0/
+ http://creativecommons.org/licenses/by/4.0/
+
+
+
+ Software archive
+ 10.5281/zenodo.14601979
+
+
+ GitHub review issue
+ https://github.com/openjournals/joss-reviews/issues/7226
+
+
+
+ 10.21105/joss.07226
+ https://joss.theoj.org/papers/10.21105/joss.07226
+
+
+ https://joss.theoj.org/papers/10.21105/joss.07226.pdf
+
+
+
+
+
+ River: Machine learning for streaming data in Python
+ Montiel
+ Journal of Machine Learning Research
+ 110
+ 22
+ 2021
+ Montiel, J., Halford, M., Mastelini, S. M., Bolmier, G., Sourty, R., Vaysse, R., Zouitine, A., Gomes, H. M., Read, J., Abdessalem, T., & Bifet, A. (2021). River: Machine learning for streaming data in Python. Journal of Machine Learning Research, 22(110), 1–8. http://jmlr.org/papers/v22/20-1380.html
+
+
+ A retrospective of the tutorial on opportunities and challenges of online deep learning
+ Kulbach
+ 10.48550/arXiv.2405.17222
+ 2024
+ Kulbach, C., Cazzonelli, L., Ngo, H.-A., Le-Nguyen, M.-H., & Bifet, A. (2024). A retrospective of the tutorial on opportunities and challenges of online deep learning. https://doi.org/10.48550/arXiv.2405.17222
+
+
+ Detecting anomalies with autoencoders on data streams
+ Cazzonelli
+ Joint european conference on machine learning and knowledge discovery in databases
+ 10.1007/978-3-031-26387-3_16
+ 2022
+ Cazzonelli, L., & Kulbach, C. (2022). Detecting anomalies with autoencoders on data streams. Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 258–274. https://doi.org/10.1007/978-3-031-26387-3_16
+
+
+ Automatic differentiation in PyTorch
+ Paszke
+ 2017
+ Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., & Lerer, A. (2017). Automatic differentiation in PyTorch. https://api.semanticscholar.org/CorpusID:40027675
+
+
+ From concept drift to model degradation: An overview on performance-aware drift detectors
+ Bayram
+ Knowledge-Based Systems
+ 245
+ 10.1016/j.knosys.2022.108632
+ 2022
+ Bayram, F., Ahmed, B. S., & Kassler, A. (2022). From concept drift to model degradation: An overview on performance-aware drift detectors. Knowledge-Based Systems, 245, 108632. https://doi.org/10.1016/j.knosys.2022.108632
+
+
+ Learning under concept drift: A review
+ Lu
+ IEEE transactions on knowledge and data engineering
+ 12
+ 31
+ 10.1109/TKDE.2018.2876857
+ 2018
+ Lu, J., Liu, A., Dong, F., Gu, F., Gama, J., & Zhang, G. (2018). Learning under concept drift: A review. IEEE Transactions on Knowledge and Data Engineering, 31(12), 2346–2363. https://doi.org/10.1109/TKDE.2018.2876857
+
+
+ MOA: Massive online analysis
+ Bifet
+ Journal of Machine Learning Research
+ 52
+ 11
+ 2010
+ Bifet, A., Holmes, G., Kirkby, R., & Pfahringer, B. (2010). MOA: Massive online analysis. Journal of Machine Learning Research, 11(52), 1601–1604. http://jmlr.org/papers/v11/bifet10a.html
+
+
+ Scikit-multiflow: A multi-output streaming framework
+ Montiel
+ Journal of Machine Learning Research
+ 72
+ 19
+ 2018
+ Montiel, J., Read, J., Bifet, A., & Abdessalem, T. (2018). Scikit-multiflow: A multi-output streaming framework. Journal of Machine Learning Research, 19(72), 1–5. http://jmlr.org/papers/v19/18-251.html
+
+
+ creme, a Python library for online machine learning
+ Halford
+ 2020
+ Halford, M., Bolmier, G., Sourty, R., Vaysse, R., & Zouitine, A. (2020). creme, a Python library for online machine learning (Version 0.6.1). https://github.com/MaxHalford/creme
+
+
+ CapyMOA — capymoa.org
+ CapyMOA Developers
+ 2024
+ CapyMOA Developers. (2024). CapyMOA — capymoa.org. https://capymoa.org.
+
+
+
+
+
+
diff --git a/joss.07226/10.21105.joss.07226.pdf b/joss.07226/10.21105.joss.07226.pdf
new file mode 100644
index 0000000000..601f99fd0b
Binary files /dev/null and b/joss.07226/10.21105.joss.07226.pdf differ
diff --git a/joss.07226/paper.jats/10.21105.joss.07226.jats b/joss.07226/paper.jats/10.21105.joss.07226.jats
new file mode 100644
index 0000000000..3795ab5e8b
--- /dev/null
+++ b/joss.07226/paper.jats/10.21105.joss.07226.jats
@@ -0,0 +1,611 @@
+
+
+
+
+
+
+
+Journal of Open Source Software
+JOSS
+
+2475-9066
+
+Open Journals
+
+
+
+7226
+10.21105/joss.07226
+
+DeepRiver: A Deep Learning Library for Data
+Streams
+
+
+
+https://orcid.org/0000-0002-9363-4728
+
+Kulbach
+Cedric
+
+
+*
+
+
+https://orcid.org/0000-0003-2886-1219
+
+Cazzonelli
+Lucas
+
+
+
+
+https://orcid.org/0000-0002-7583-753X
+
+Ngo
+Hoang-Anh
+
+
+*
+
+
+https://orcid.org/0000-0003-1464-4520
+
+Halford
+Max
+
+
+
+
+https://orcid.org/0000-0002-0092-3572
+
+Mastelini
+Saulo Martiello
+
+
+
+
+
+FZI Research Center for Information Technology, Karlsruhe,
+Germany
+
+
+
+
+AI Institute, University of Waikato, Hamilton, New
+Zealand
+
+
+
+
+Carbonfact, Paris, France
+
+
+
+
+Institute of Mathematics and Computer Science, University
+of São Paulo, São Carlos, Brazil
+
+
+
+
+* E-mail:
+* E-mail:
+
+10
+105
+7226
+
+Authors of papers retain copyright and release the
+work under a Creative Commons Attribution 4.0 International License (CC
+BY 4.0)
+2025
+The article authors
+
+Authors of papers retain copyright and release the work under
+a Creative Commons Attribution 4.0 International License (CC BY
+4.0)
+
+
+
+Python
+
+
+
+
+
+ Summary
+
Machine learning algorithms enhance decision-making efficiency by
+ leveraging available data. However, as data evolves over time, it
+ becomes crucial to adapt machine learning (ML) systems incrementally
+ to accommodate new data patterns. This adaptation is achieved through
+ online learning or continuous ML technologies. Although deep learning
+ technologies have demonstrated outstanding performance on predefined
+ datasets, their application to online, streaming, and continuous
+ learning scenarios has been limited.
+
DeepRiver is a Python package for deep
+ learning on data streams. Built on top of River
+ (Montiel
+ et al., 2021) and PyTorch
+ (Paszke
+ et al., 2017), it offers a unified API for both supervised and
+ unsupervised learning. Additionally, it provides a suite of tools for
+ preprocessing data streams and evaluating deep learning models.
+
+
+ Statement of need
+
In today’s rapidly evolving landscape, machine learning (ML)
+ algorithms play a pivotal role in shaping decision-making processes
+ based on available data. These algorithms, while accelerating
+ analysis, require continuous adaptation to dynamic data structures, as
+ patterns may evolve rapidly. To address this imperative, adopting
+ online learning and continuous ML technologies becomes paramount.
+ While deep learning technologies have demonstrated exceptional
+ performance on static, predefined datasets, their application to
+ dynamic and continuously evolving data streams remains underexplored.
+ The absence of widespread integration of deep learning into online,
+ streaming, and continuous learning scenarios hampers the full
+ potential of these advanced algorithms in real-time decision-making
+ (Kulbach
+ et al., 2024). The emergence of the
+ DeepRiver Python package fills a critical void
+ in the field of deep learning on data streams. Leveraging the
+ capabilities of River
+ (Montiel
+ et al., 2021) and PyTorch
+ (Paszke
+ et al., 2017), DeepRiver offers a
+ unified API for both supervised and unsupervised learning, providing a
+ seamless bridge between cutting-edge deep learning techniques and the
+ challenges posed by dynamic data streams. Moreover, the package equips
+ practitioners with essential tools for data stream preprocessing and
+ the evaluation of deep learning models in dynamic, real-time
+ environments. Such functionality has been applied to Streaming Anomaly
+ Detection
+ (Cazzonelli
+ & Kulbach, 2022). As the demand for effective and efficient
+ adaptation of machine learning systems to evolving data structures
+ continues to grow, the integration of DeepRiver
+ into the landscape becomes crucial. This package stands as a valuable
+ asset, unlocking the potential for deep learning technologies to excel
+ in online, streaming, and continuous learning scenarios. The need for
+ such advancements is evident in the quest to harness the full power of
+ machine learning in dynamically changing environments, ensuring our
+ decision-making processes remain accurate, relevant, and agile in the
+ face of evolving data landscapes.
+
+
+ Related Work
+
Online machine learning involves updating models incrementally as
+ new data arrives, rather than retraining models from scratch. Several
+ frameworks and libraries have been developed to support this
+ paradigm:
+
+
+
scikit-multiflow
+ (Montiel
+ et al., 2018)
+
+
+
Python-based Library: Inspired by the Java-based MOA
+ framework, designed for streaming data and online learning in
+ Python.
+
+
+
Key Features:
+
+
+
Supports algorithms like Hoeffding Trees, online
+ bagging, and boosting.
+
+
+
Includes concept drift detection (e.g., ADWIN,
+ Page-Hinkley) to adapt to changing data distributions.
+
+
+
Stream generators and evaluators for real-time data
+ simulation and model assessment.
+
+
+
+
+
Limitations: Focuses mainly on traditional machine learning
+ methods, with limited support for deep learning
+ architectures.
+
+
+
+
+
creme
+ (Halford
+ et al., 2020)
+
+
+
Lightweight Online Learning: Specialized in incremental
+ learning where models are updated per instance, leading to
+ efficient, low-latency model training.
+
+
+
Provides a unified API with a broad range of online
+ learning algorithms, making it the go-to library for streaming
+ data analysis in Python.
+
+
+
Limitations: Primarily supports feature-based models with
+ limited capabilities for deep neural networks.
+
+
+
+
+
In 2020, creme merged with
+ scikit-multiflow to create
+ River, combining the strengths of both
+ frameworks.
+
+
+
Massive Online Analysis (MOA)
+ (Bifet
+ et al., 2010)
+
+
+
Java-based Pioneer: One of the earliest frameworks
+ dedicated to stream mining and online learning, widely used in
+ academic research.
+
+
+
Key Features:
+
+
+
Introduces foundational algorithms like Hoeffding
+ Trees, Adaptive Random Forest (ARF), and several drift
+ detection techniques (e.g., DDM, EDDM).
+
+
+
Excellent scalability for handling high-throughput data
+ streams in real-time.
+
+
+
Strong focus on concept drift adaptation, making it
+ robust in non-stationary environments.
+
+
+
+
+
+
+
capyMOA
+ (CapyMOA
+ Developers, 2024)
+
+
+
Python Interface for MOA: capyMOA serves as a bridge
+ between the Java-based MOA framework and Python, allowing
+ users to leverage MOA’s powerful streaming algorithms within
+ Python workflows.
+
+
+
Key Features:
+
+
+
Enables access to MOA’s core functionalities (e.g.,
+ Hoeffding Trees, Adaptive Random Forest) from Python.
+
+
+
Facilitates hybrid workflows by integrating MOA’s Java
+ algorithms with Python’s machine learning libraries.
+
+
+
Useful for Python developers looking to use MOA’s
+ advanced stream mining capabilities without switching
+ ecosystems.
+
+
+
+
+
+
+
scikit-multiflow and
+ creme (River) focus on
+ efficient online learning in Python, mainly for traditional machine
+ learning algorithms. MOA offers extensive tools for stream mining but
+ lacks deep learning support and Python compatibility. While capyMOA
+ provides Python accessibility to MOA, capyMOA is limited by the
+ underlying Java infrastructure and lacks a natural integration with
+ PyTorch’s deep learning ecosystem.
+
DeepRiver differentiates itself by
+ integrating deep learning capabilities directly into streaming data
+ workflows, enabling continuous learning for neural network models.
+ This addresses a critical gap left by existing frameworks, which are
+ predominantly focused on non-deep learning models.
+
+
+ Features
+
DeepRiver enables the usage of deep learning
+ models for data streams. This means that deep learning models need to
+ adapt to changes within the evolving data stream
+ (Bayram
+ et al., 2022;
+ Lu
+ et al., 2018) e.g. the number of classes might change over
+ time. In addition to the integration of PyTorch
+ (Paszke
+ et al., 2017) into River
+ (Montiel
+ et al., 2021), this package offers additional data stream
+ specific functionalities such as class incremental learning or
+ specific optimizers for data streams.
+
+ Compatibility
+
DeepRiver is built on the unified
+ application programming interface (API) of
+ River(Montiel
+ et al., 2021) that seamlessly integrates both supervised and
+ unsupervised learning techniques. Additionally, it incorporates
+ PyTorch’s
+ (Paszke
+ et al., 2017) extensive functionality for deep learning such
+ as using GPU acceleration and a broad range of architectures. This
+ unified approach simplifies the development process and facilitates
+ a cohesive workflow for practitioners working with dynamic data
+ streams. Leveraging the capabilities of the well-established
+ River(Montiel
+ et al., 2021) library and the powerful
+ PyTorch(Paszke
+ et al., 2017) framework, DeepRiver
+ combines the strengths of these technologies to deliver a robust and
+ flexible platform for deep learning on data streams. This foundation
+ ensures reliability, scalability, and compatibility with
+ state-of-the-art machine learning methodologies, with comprehensive
+ documentation
+ guiding users through the installation, implementation, and
+ customization processes. Additionally, a supportive community
+ ensures that all DeepRiver’s users have
+ access to resources, discussions, and assistance, fostering a
+ collaborative environment for continuous improvement and knowledge
+ sharing.
+
+
+ Adaptivity
+
DeepRiver is specifically designed to
+ cater to the requirements of online learning scenarios. It enables
+ continuous adaptation to evolving data by supporting incremental
+ updates and learning from new observations in real time, a critical
+ feature for applications where data arrives sequentially. Moreover,
+ it allows the model to dynamically adjust to changes in the number
+ of classes over time for classification tasks. It equips
+ practitioners with tools for evaluating the performance of deep
+ learning models on data streams. This feature is crucial for
+ ensuring the reliability and effectiveness of models in real-time
+ applications, enabling users to monitor and fine-tune their models
+ as the data evolves.
+
+
+
+ Architecture
+
The DeepRiver library is structured around
+ various types of estimators for anomaly detection, classification, and
+ regression. In anomaly detection, the base class
+ AnomalyScaler has derived classes
+ AnomalyMeanScaler,
+ AnomalyMinMaxScaler, and
+ AnomalyStandardScaler. Additionally, the
+ Autoencoder class, which inherits from
+ DeepEstimator, has a specialized subclass
+ called ProbabilityWeightedAutoencoder. The
+ RollingAutoencoder class inherits from
+ RollingDeepEstimator.
+
For classification, the base class
+ Classifier inherits from
+ DeepEstimator. Derived from
+ Classifier are specific classes like
+ LogisticRegression and
+ MultiLayerPerceptron. The
+ RollingClassifier class inherits from both
+ RollingDeepEstimator and
+ Classifier.
+
In regression, the base class Regressor
+ inherits from DeepEstimator. Specific
+ regression classes like LinearRegression and
+ MultiLayerPerceptron inherit from
+ Regressor. The
+ MultiTargetRegressor also inherits from
+ DeepEstimator. The
+ RollingRegressor class inherits from both
+ RollingDeepEstimator and
+ Regressor.
+
+
Architecture of
+ DeepRiver
+
+
+
Overall, the library is organized to provide a flexible and
+ hierarchical framework for different types of machine learning tasks,
+ with a clear inheritance structure connecting more specific
+ implementations to their base classes.
+
+
+ Acknowledgements
+
Hoang-Anh Ngo received an External Study Awards from the AI
+ Institute, University of Waikato, Hamilton, New Zealand for research
+ on online machine learning under the supervision of Prof. Albert
+ Bifet.
+
+
+
+
+
+
+
+
+ MontielJacob
+ HalfordMax
+ MasteliniSaulo Martiello
+ BolmierGeoffrey
+ SourtyRaphael
+ VaysseRobin
+ ZouitineAdil
+ GomesHeitor Murilo
+ ReadJesse
+ AbdessalemTalel
+ BifetAlbert
+
+ River: Machine learning for streaming data in Python
+ Journal of Machine Learning Research
+ 2021
+ 22
+ 110
+ http://jmlr.org/papers/v22/20-1380.html
+ 1
+ 8
+
+
+
+
+
+ KulbachCedric
+ CazzonelliLucas
+ NgoHoang-Anh
+ Le-NguyenMinh-Huong
+ BifetAlbert
+
+ A retrospective of the tutorial on opportunities and challenges of online deep learning
+ 2024
+ https://arxiv.org/abs/2405.17222
+ 10.48550/arXiv.2405.17222
+
+
+
+
+
+ CazzonelliLucas
+ KulbachCedric
+
+ Detecting anomalies with autoencoders on data streams
+ Joint european conference on machine learning and knowledge discovery in databases
+ Springer
+ 2022
+ 10.1007/978-3-031-26387-3_16
+ 258
+ 274
+
+
+
+
+
+ PaszkeAdam
+ GrossSam
+ ChintalaSoumith
+ ChananGregory
+ YangEdward
+ DeVitoZach
+ LinZeming
+ DesmaisonAlban
+ AntigaLuca
+ LererAdam
+
+ Automatic differentiation in PyTorch
+ 2017
+ https://api.semanticscholar.org/CorpusID:40027675
+
+
+
+
+
+ BayramFiras
+ AhmedBestoun S.
+ KasslerAndreas
+
+ From concept drift to model degradation: An overview on performance-aware drift detectors
+ Knowledge-Based Systems
+ Elsevier
+ 2022
+ 245
+ 10.1016/j.knosys.2022.108632
+ 108632
+
+
+
+
+
+
+ LuJie
+ LiuAnjin
+ DongFan
+ GuFeng
+ GamaJoao
+ ZhangGuangquan
+
+ Learning under concept drift: A review
+ IEEE transactions on knowledge and data engineering
+ IEEE
+ 2018
+ 31
+ 12
+ 10.1109/TKDE.2018.2876857
+ 2346
+ 2363
+
+
+
+
+
+ BifetAlbert
+ HolmesGeoff
+ KirkbyRichard
+ PfahringerBernhard
+
+ MOA: Massive online analysis
+ Journal of Machine Learning Research
+ 2010
+ 11
+ 52
+ http://jmlr.org/papers/v11/bifet10a.html
+ 1601
+ 1604
+
+
+
+
+
+ MontielJacob
+ ReadJesse
+ BifetAlbert
+ AbdessalemTalel
+
+ Scikit-multiflow: A multi-output streaming framework
+ Journal of Machine Learning Research
+ 2018
+ 19
+ 72
+ http://jmlr.org/papers/v19/18-251.html
+ 1
+ 5
+
+
+
+
+
+ HalfordMax
+ BolmierGeoffrey
+ SourtyRaphael
+ VaysseRobin
+ ZouitineAdil
+
+ creme, a Python library for online machine learning
+ 20200610
+ https://github.com/MaxHalford/creme
+
+
+
+
+
+ CapyMOA Developers
+
+ CapyMOA — capymoa.org
+ https://capymoa.org
+ 2024
+
+
+
+
+
diff --git a/joss.07226/paper.jats/classes.png b/joss.07226/paper.jats/classes.png
new file mode 100644
index 0000000000..4fab89fed4
Binary files /dev/null and b/joss.07226/paper.jats/classes.png differ