Es2019 3

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/353983287
Recent trends in streaming data analysis, concept drift and analysis of

dynamic data sets
Conference Paper · August 2021
CITATIONS READS
8 318
3 authors, including:
Albert Bifet Barbara Hammer

Télécom ParisTech Bielefeld University
255 PUBLICATIONS 13,355 CITATIONS 515 PUBLICATIONS 7,330 CITATIONS
SEE PROFILE SEE PROFILE
Some of the authors of this publication are also working on these related projects:
scikit-multiflow: machine learning for data streams in Python View project
Statistical Physics of Machine Learning View project
All content following this page was uploaded by Albert Bifet on 18 August 2021.
The user has requested enhancement of the downloaded file.

ESANN 2019 proceedings, European Symposium on Artificial Neural Networks, Computational Intelligence
and Machine Learning. Bruges (Belgium), 24-26 April 2019, i6doc.com publ., ISBN 978-287-587-065-0.
Available from http://www.i6doc.com/en/.
Recent trends in
streaming data analysis, concept drift and
analysis of dynamic data sets
Albert Bifet1 , Barbara Hammer2 and Frank-M. Schleif3
1- LTCI, Telecom ParisTech,

Université Paris-Saclay
Paris, France
2- Bielefeld University,
CITEC centre of excellence
Bielefeld, Germany
3- University of Birmingham
School of Computer Science,
Edgbaston B15 2TT Birmingham,
United Kingdom.
Abstract. Today, many data are not any longer static but occur as
dynamic data streams with high velocity, variability and volume. This
leads to new challenges to be addressed by novel or adapted algorithms.
In this tutorial we provide an introduction into the field of streaming data
analysis summarizing its major characteristics and highlighting important
research directions in the analysis of dynamic data.
1 Introduction
In many application domains data are given at large scale and with high velocity,
requesting an in time analysis without the possibility to store large parts of
the data or to process them multiple times. Examples include astronomical
observations, earth sensing satellites and climate observation, genomics and post-
genomics data, data gathered by smart sensors such as smart phones or wearable
devices, IoT data, data gathered from assistive technologies such as Amazon’s
Alexa, data gathered in smart cities or smart factories, etc. Such data are widely
referred to as streaming data since measurements arrive continuously as a data
stream. In addition to the sheer data size, which typically prevents its processing
in batches, streaming data can pose additional challenges to the models which
renders standard techniques of machine learning unsuitable.
In recent years, quite a few approaches have been proposed in this context,
most of which are different from current popular learning methods for batch
learning, see e.g. [1, 2, 3, 4, 5, 6, 7, 8] for overviews. Besides the mere computa-
tional issues, online learning faces quite a few challenges which are fundamentally
different from classical batch processing. In the following, we shortly define what
we refer to as online learning first and we give an overview about challenges in
this domain. We address two major tasks: (i) How to derive supervised models
421
for streaming data? (ii) How to best represent data in the streaming setting?
We conclude with a glimpse on recent directions in this domain.
2 Problem statement and challenges

In online learning, a stream s1 , s2 , s3 , . . . st , . . . of data si ∈ S is given, and the
task is to incrementally infer a model ht after having seen instances s1 , . . . , st .
Unlike batch processing, the characteristics of the subsequent data st0 for t0 > t is
unknown, and it is not guaranteed that future data obey the same characteristics
as the already seen samples. Such settings pose a variety of challenges towards
learning, which do not occur in this form in offline learning.
Challenge 1: How to devise efficient online learning rules, which
instantaneously adapt model parameters based on streaming data?
Typically, efficient incremental learning is required since models should be avail-
able in real time while observing the data stream. Hence, model ht is typically
inferred from the data point st (or a small window of previous samples only)
and the previous model ht−1 , without explicitly storing all previous data points,
i.e. memory efficiency is given. In addition, computational efficiency is aimed
for, i.e. the learning rules should be efficient, yet effective in the sense that the
update resemble the accuracy of batch learning as far as possible (provided data
are i.i.d.) but without having access to the full data [9].
Challenge 2a: How to find suitable model meta-parameters?
For offline learning of parametric models, typically, model meta-parameters,
which characterize the complexity and degree of non-linearity (e.g. kernel width
or neural architecture) are determined prior to training by Bayesian optimiza-
tion of cross-validation. For online learning, this is not possible, since the re-
quired model complexity cannot be estimated prior to training without access to
sufficient training data. Hence model meta-parameters, which characterize the
model complexity, become model parameters in online learning — they need to be
adapted while processing the data stream, unless non-parametric modeling (e.g.
k-NN) or a priorly limited complexity (e.g. linear maps) are chosen. The prob-
lem of a priorly unclear model complexity requires suitable learning strategies
for the meta-parameters, which characterize the model – usually a challenging
task, since these meta-parameters are typically discrete-valued [4, 10].
Challenge 2b: How to preprocess data?
Besides model complexity, many models typically come with a set of further
meta-parameters such as step size, strength or regularization, etc., which need to
be robustly chosen or adapted for online learning, since its suitability can hardly
be determined prior to training. A related challenge concerns data preprocessing,
such as data normalization, or dimensionality reduction for numerical stability
and sampling in the presence of imbalanced classes. Similar to model meta-
parameters, all data preprocessing steps become part of the actual processing
pipeline and need to be adjusted while training [11].
422
(a) Rapid (b) Incremental (c) Gradual (d) Recurring
Fig. 1: Different drift types as they occur in streaming data analysis [5].
Challenge 3: How to deal with concept drift?
For online learning, it is usually not guaranteed that data come from a stationary
data source; hence the typical assumption of batch processing, the fact that data
are i.i.d., is violated. Whenever at least two points t, t0 in time exist where the
underlying probability distribution changes, i.e. Pt 6= Pt0 one speaks of concept
drift. Thereby, drift characteristics can vary, as depicted in Fig. 1, yielding
smooth or rapid drift, incremental, gradual or reoccurring drift. In addition
to such structural changes, outliers can occur, i.e. data st deviate randomly
from the underlying distribution Pt . Such settings probably constitute the most
fundamental difference of learning with streaming data to the batch setting, since
models need to adequately react to changed underlying probability distributions
during their whole lifetime. In particular if the type of drift is unknown, the
classical stability-plasticity dilemma arises and learning faces an essentially ill-
posed problem [12]: when is observed change caused by an underlying structure
and should be taken into account, and when is it given by noise and should be
neglected? Interestingly, if drift occurs, online learning can yield results which
are superior to batch training, since online learning can react to changes and
provide an optimal model at every given time step, while batch learning needs
to restrict to an average model which fits an average time point only [6].
In practice, quite a number of different models have been proposed, which
can roughly be decomposed into the following categories:
Supervised learning: Data have the form st = (xt , yt ) and the task is to learn
a model ht which predicts the subsequent output yt+1 ≈ f (xt+1 ) before actu-
ally seeing the true outcome. Such scenarios are relevant in practice whenever
an early prediction is required such as predicting the behavior of participants
in a traffic scene for early motion planning. For supervised learning, one dis-
tinguishes the notion of real drift, which refers to a change of the posterior
distribution P(y|x) and virtual drift or covariate shift, which refers to a change
of the input distribution P(x) only without affecting the posterior.
Unsupervised data representation: Data st are unlabeled, and the task is to solve
a problem related to data representation such as online generative modeling,
online compression, dimensionality reduction, clustering, or outlier detection.
Interestingly, quite a few early methods of machine learning such as principal
423
component analysis via the Oja-learning rule or self-organizing maps have been
inspired by biological counterparts and phrased as online learning algorithms for
streaming data in its original form [13, 14].
Time characteristics of data: Some methods explicitly focus on the time charac-
teristics of the data and tackle challenges, which can only be asked in the context
of data streams. Partially, these tasks occur as sub-problems of supervised or un-
supervised learning problems for streaming data. One central questions, which is
often embedded in so-called active methods for streaming data, is drift detection,
i.e. the task to detect points in time where a significant change of the underlying
probability distribution can be observed [15]. Such detectors are often coupled
by a strategy to adapt the model whenever a drift is detected [3]. Further chal-
lenges address the temporal characteristics, and focus on time-invariant motives
or possible (Granger) causal relations [16, 17, 18].
3 Supervised learning in the streaming context

Interestingly, many classical learning rules such as perceptron learning, learn-
ing vector quantization, self-organizing maps, or neural back-propagation have
been proposed as online learning rules, and these technologies, indeed, provide
an astonishing robustness when learning with drift [19]. The SVM and, more
generally, the focus on convex optimization methods and the possibility of their
strong mathematical substantiation led to a dominance of batch learners in ma-
chine learning. With the advent of large data sets and the involved necessity to
process data in patches, learning from streaming data became an issue again.
One of the first methods which have been proposed in this context is the very
fast decision tree [16], an incremental variant of a decision tree which resem-
bles its batch counterpart. Another popular early method was Learn++, which
constructs an ensemble very much in the spirit of adaptive boosting [20].
There exists a quite large variety of different approaches for supervised online
learning, which rely on different principles and try to address different challenges.
Quite a few methods focus on the computational issue how to efficiently transfer
popular batch learners into online-learning counterparts, such that their result is
equivalent to or (approximates) the result of the batch learner, a popular example
being e.g. incremental variants of SVM [9], incremental decision trees or random
forests [21], or incremental Bayesian learners [22]. Other methods investigate
how to efficiently adjust the model complexity according to the observed data,
whereby methods have to face a balance between [23, 24].
One of the most crucial challenges is given by possible concept drift and the
question, how to deal with such drift. A variety of approaches centers on the
detection and quantification of the observed drift in the first place, enabling a
reaction to the drift in so-called active approaches. Examples are window-based
approaches such as presented in the approaches [25, 15, 26], or the approaches
[27, 28], which relies on a modeling via the β-distribution. Such detectors can
be combined with random forests or other classifiers [29]. On the other side,
passive approaches continuously adapt models according to the given data or
424
an active window, respectively, this way reacting also to smooth drift. Popular
approaches often rely on non-parametric methods, such as extremely robust
k-NN based methods [24] or prototype-based approaches [30]. Many modern
technologies fall into the category of hybrid techniques, which combine active
drift detection and continuous passive adaptation, this way combining the best
of both worlds.
Many recent very popular different algorithmic approaches follow the idea
of ensemble techniques, in particular tree ensembles [31] or ensembles of local
models [32], this way displaying a high robustness and independence of model
parametrization. From an application point of view, it is interesting to inves-
tigate which types of drift the models can deal with. By design, active drift
detection methods are restricted to rapid drift, hence they are less suited for
subtle incremental changes, but, on the other hand, typically react rapidly to
the detected change. Passive or hybrid methods typically smoothly deal with
continuously changing environment. Interestingly, there exist currently only very
few approaches capable of dealing with reoccurring drift, a few of those having
been proposed in the work [33, 34, 35]. With the advent of streaming data e.g.
in personalized assistive systems such as investigated in the work [36], the rele-
vance of such methodologies, which are capable of a flexible reaction to priorly
unknown types of drift, will become even more prominent.
Quite a few further challenges have recently been addressed in the context of
supervised learning for streaming data, a few keywords being the mathematical
investigation of their convergence properties [37], learning in the context of more
complex outputs such as structured predictions and multi-label learning [38],
semi-supervised online learning [39], or learning for imbalanced data [40].
4 Unsupervised data representation in the streaming con-

text
Similar like in classical data analysis the pre-processing can have a large impact
on the effectiveness of a subsequent analysis. Still the majority of streaming data
are (multi-dimensional) vectorial but the generating sources may have very dif-
ferent qualities. For example the data may be generated at different output rates,
the noise level may be very different and the transmitted amount of information
per time window can vary to a large degree. Also other type of non-numeric
data by means of textual data, graphs or symbolic sequences are getting more
and more common [41], asking for appropriate preprocessing to enable effective
model building. Common to used pre-processing concepts is a (near) memory
less approach. The incoming data are processed on the fly and the respective
processing algorithms can only keep very limited information about the data
seen so far. Hence in the most basic pre-processings like smoothing or noise
reduction simple filter techniques like moving average are used. A survey can be
found in [42].
Also the general problem of imbalanced data is a particular challenge in the
streaming context [43], the distribution in general or with respect to class labels
425
may not be uniform and can evolve other time. One application of imbalanced
learning is anomaly detection, where the problem consists in predicting when
an anomaly appears. As anomalies appear with a very low frequency, it is a
classical example of imbalanced learning [44].
Another important aspect is the existence of outliers. As discussed before
streaming data are dynamic and hence show varying distributions and data char-
acteristics. Outliers are particular challenging in the streaming context because
the data distributions are naturally changing and it becomes very complicate
to decide whether the observation is an outlier or due to a change in the data
distributions. Early work addressing this point can be found in [45].
A classical data preprocessing approach is the principal component analysis
(PCA) to characterize the variance in the data. Also in the streaming context
initial work was provided to allow PCA like data processing [46]. The majority
of those techniques go back to traditional power iteration methods to get an
estimate of the underlying eigen functions with links to early neural network
approaches, like Oja PCA [47].
In high-dimensional data, using all attributes is often not feasible, and we
may need to preprocess the data to perform feature selection, or feature trans-
formation. This can be done by a streaming PCA [46], but also other dimension
reduction approaches are under research like multi dimensional scaling (MDS)
[48] to obtain low dimensional data representations which keeps some distance
preservation.
Another interesting approach is to systematically identify relevant input fea-
tures by means of a weighting or relevance scheme and a metric adaptation
concept [30].
The streaming domain shows also links to different types of the (contextual)
bandit problem. In the contextual bandit problem a learner chooses an action
among a set of available ones, based on the observation of action features or
contexts and then receives a reward that quantifies the quality of the chosen
action. This scenario can be used in vary different dedicated algorithms e.g. to
adapt models, to switch between different input streams or other ways [49, 50].
The majority of the proposed streaming analysis methods make still use of
linear or piecewise linear models due to the simplicity which is very desirable for
large scale and high throughput streaming data. To overcome the limitation of
linear methods also first kernelization strategies have been proposed, with some
initial work in [51].
5 Future trends
Future trends in streaming data analysis, are based on how to develop data
streaming methods that scale to Big Data like large deep neural networks, but
work well in all domains. In the future, the quantity of data generated in real-
time is going to continue growing, so there will be need to develop new methods
using large distributed systems.
Deep Learning has become a very extreme successful use case for Machine
Learning and Artificial Intelligence, due to the availability of massive quantities
426
of data to build data models, and large computational resources. How to imple-
ment powerful methods such as deep learning, in a more green, low-emissions,
sustainable way, is going to be an important scientific trend to fight against
climate change. Standard deep learning techniques needs to do several passes
over the data. How to build models only doing one pass over the data, without
storing the data, will be an important future area of research [52].
Finally, when dealing with large quantities of data, an important trend will be
how to do online learning using distributed streaming engines, as Apache Spark,
Apache Flink, Apache Storm and others. Algorithms have to be distributed in
an efficient way, so that the performance of the distributed algorithms does not
suffer from the network cost of distributing the data [53].
6 Conclusions
In this tutorial we briefly reviewed challenges and approaches common in the field
of streaming data analysis, concept drift and the analysis of dynamic data sets.
The more recent proposals in these domains provide sophisticated algorithms and
models to address the aforementioned challenges in streaming analysis and in
particular the handling of concept drift. Also a variety of classical supervised and
unsupervised analysis tasks like modeling non-linear decision planes or finding
relevant dimensions in the data streams now came in the focus of recent research.
Although the field has made much progress in preprocessing [42], concept drift
detection [24, 26] and by means of generic frameworks for streaming analysis
[54] there are still particular challenges as detailed before with a variety of open
research perspectives.
Acknowledgment
BH gratefully acknowledges funding in the frame of the CITEC centre of excel-

lence and the BMBF project ITS ML, grant number 01IS18041A. FMS is thank-
ful for support in the FuE program Informations- und Kommunikationstechnik
of the StMWi, project OBerA, grant number IUK-1709-0011// IUK530/010.
References
[1] RR Ade and PR Deshmukh. Methods for incremental learning: a survey. International
Journal of Data Mining & Knowledge Management Process, 3(4):119, 2013.
[2] Gregory Ditzler, Manuel Roveri, Cesare Alippi, and Robi Polikar. Learning in nonstation-
ary environments: A survey. Computational Intelligence Magazine, IEEE, 10(4):12–25,
2015.
[3] R. Elwell and R. Polikar. Incremental learning of concept drift in nonstationary environ-
ments. IEEE Transactions on Neural Networks, 22(10):1517–1531, Oct 2011.
[4] Alexander Gepperth and Barbara Hammer. Incremental learning algorithms and ap-
plications. In Proceedings of the European Sympoisum on Artificial Neural Networks
(ESANN), 2016.
[5] João Gama, Indre Žliobaite, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid
Bouchachia. A survey on concept drift adaptation. ACM Computing Surveys (CSUR),
46(4):44, 2014.
427
[6] Viktor Losing, Barbara Hammer, and Heiko Wersing. Incremental on-line learning: A
review and comparison of state of the art algorithms. Neurocomputing, 275:1261–1274,
2018.
[7] Indre Zliobaite. Learning under concept drift: an overview, 2010.
[8] Adriana Sayuri Iwashita and Joao Paulo Papa. An Overview on Concept Drift Learning.
IEEE Access, 7(Section III):1–1, 2018.
[9] C. P. Diehl and G. Cauwenberghs. Svm incremental learning, adaptation and optimiza-
tion. In Proceedings of the International Joint Conference on Neural Networks, 2003.,
volume 4, pages 2685–2690 vol.4, July 2003.
[10] Prachi Joshi and Parag Kulkarni. Incremental learning: areas and methods-a survey.
International Journal of Data Mining & Knowledge Management Process, 2(5):43, 2012.
[11] Gregory Ditzler and Robi Polikar. Incremental learning of concept drift from streaming
imbalanced data. ieee transactions on knowledge and data engineering, 25(10):2283–2301,
2013.
[12] Martial Mermillod, AurÃ liac Bugaiska, and Patrick Bonin. The stability-plasticity
dilemma: Investigating the continuum from catastrophic forgetting to age-limited learn-
ing effects. Frontiers in Psychology, 4(504), 2013.
[13] G. A. Carpenter, S. Grossberg, and J. Reynolds. Artmap: a self-organizing neural network
architecture for fast supervised learning and pattern recognition. In IJCNN-91-Seattle
International Joint Conference on Neural Networks, volume i, pages 863–868 vol.1, July
1991.
[14] Teuvo Kohonen. Self-organizing maps. Springer series in information sciences, 30.
Springer, Berlin, 3rd edition, December 2001.
[15] Albert Bifet and Ricard Gavalda. Learning from time-changing data with adaptive win-
dowing. In SIAM International Conference on Data Mining (SDM), volume 7, page 2007.
SIAM, 2007.
[16] Pedro Domingos and Geoff Hulten. Mining high-speed data streams. In Proceedings of the
sixth ACM SIGKDD international conference on Knowledge discovery and data mining,
pages 71–80. ACM, 2000.
[17] Pedro Domingos and Geoff Hulten. A general framework for mining massive data streams.
Journal of Computational and Graphical Statistics, 12(4):945–949, 2003.
[18] Mohamed Medhat Gaber, Arkady Zaslavsky, and Shonali Krishnaswamy. Mining data
streams: a review. ACM Sigmod Record, 34(2):18–26, 2005.
[19] Michiel Straat, Fthi Abadi, Christina Göpfert, Barbara Hammer, and Michael Biehl.
Statistical mechanics of on-line learning under concept drift. Entropy, 20(10):775, 2018.
[20] R. Polikar, L. Upda, S. S. Upda, and V. Honavar. Learn++: an incremental learning
algorithm for supervised neural networks. IEEE Transactions on Systems, Man, and
Cybernetics, Part C (Applications and Reviews), 31(4):497–508, Nov 2001.
[21] Aiping Wang, Guowei Wan, Zhiquan Cheng, and Sikun Li. An incremental extremely
random forest classifier for online learning and tracking. In 2009 16th IEEE International
Conference on Image Processing (ICIP), Nov 2009.
[22] W. Jamil and A. Bouchachia. Online bayesian shrinkage regression. In Michel Verleysen,
editor, Proceedings of the 27. European Symposium on Artificial Neural Networks ESANN
2019, page numbers to be obtained from ToC of this proceedings book, Evere, Belgium,
2019. D-Side Publications.
[23] Iain A D Gunn, Álvar Arnaiz-González, and Ludmila I Kuncheva. A taxonomic look at
instance-based stream classifiers. Neurocomputing, 286:167–178, 2018.
[24] Viktor Losing, Barbara Hammer, and Heiko Wersing. Self-adjusting memory: How to deal
with diverse drift types. IJCAI International Joint Conference on Artificial Intelligence,
pages 4899–4903, 2017.
428
[25] Joao Gama, Pedro Medas, Gladys Castillo, and Pedro Rodrigues. Learning with drift
detection. In Advances in artificial intelligence–SBIA 2004, pages 286–295. Springer,
2004.
[26] Roberto Souto Maior Barros and Silas Garrido T.Carvalho Santos. A large-scale compar-
ison of concept drift detectors. Information Sciences, 451-452:348–370, 2018.
[27] L. Fleckenstein, S. Kauschke, and J. Fürnkranz. Beta distribution drift detection for
adaptive classifiers. In Michel Verleysen, editor, Proceedings of the 27. European Sym-
posium on Artificial Neural Networks ESANN 2019, page numbers to be obtained from
ToC of this proceedings book, Evere, Belgium, 2019. D-Side Publications.
[28] Jeremias Knoblauch, Jack E Jewson, and Theodoros Damoulas. Doubly robust bayesian
inference for non-stationary streaming data with \beta-divergences. In S. Bengio, H. Wal-
lach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in
Neural Information Processing Systems 31, pages 64–75. Curran Associates, Inc., 2018.
[29] Cesare Alippi and Manuel Roveri. Just-in-time adaptive classifiers-part i: Detecting
nonstationary changes. IEEE Transactions on Neural Networks, 19(7):1145–1153, July
2008.
[30] Ilja Kuzborskij and Nicolò Cesa-Bianchi. Nonparametric Online Regression while Learn-
ing the Metric. In Advances in Neural Information Processing Systems 30: Annual
Conference on Neural Information Processing Systems 2017, 4-9 December 2017, Long
Beach, CA, USA, pages 667–676, 2017.
[31] Heitor Gomes, Albert Bifet, Jesse Read, Jean Paul Barddal, Fabrı́cio Enembreck, Bern-
hard Pfharinger, Geoff Holmes, and Talel Abdessalem. Adaptive random forests for
evolving data stream classification. Machine Learning, 106(9):1469–1495, Oct 2017.
[32] Junming Shao, Feng Huang, Qinli Yang, and Guangchun Luo. Robust Prototype-Based
Learning on Data Streams. IEEE Transactions on Knowledge and Data Engineering,
30(5):978–991, 2018.
[33] Viktor Losing, Barbara Hammer, and Heiko Wersing. Tackling heterogeneous concept
drift with the self-adjusting memory (SAM). Knowl. Inf. Syst., 54(1):171–201, 2018.
[34] Viktor Losing, Barbara Hammer, and Heiko Wersing. KNN classifier with self adjusting
memory for heterogeneous concept drift. In IEEE 16th International Conference on Data
Mining, ICDM 2016, December 12-15, 2016, Barcelona, Spain, pages 291–300, 2016.
[35] C. Raab, M. Heusinger, and F.-M. Schleif. Reactive soft prototype computing for frequent
reoccurring concept drift. In Michel Verleysen, editor, Proceedings of the 27. European
Symposium on Artificial Neural Networks ESANN 2019, page numbers to be obtained
from ToC of this proceedings book, Evere, Belgium, 2019. D-Side Publications.
[36] P. Siirtola, H. Koskimäki, and J. Röning. Importance of user inputs while using incremen-
tal learning to personalize human activity recognition models. In Michel Verleysen, editor,
Proceedings of the 27. European Symposium on Artificial Neural Networks ESANN 2019,
page numbers to be obtained from ToC of this proceedings book, Evere, Belgium, 2019.
D-Side Publications.
[37] Lijun Zhang, Shiyin Lu, and Zhi-Hua Zhou. Adaptive online learning in dynamic en-
vironments. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi,
and R. Garnett, editors, Advances in Neural Information Processing Systems 31, pages
1330–1340. Curran Associates, Inc., 2018.
[38] Rajasekar Venkatesan, Meng Joo Er, Shiqian Wu, and Mahardhika Pratama. A novel
online real-time classifier for multi-label data streams. CoRR, abs/1608.08905, 2016.
[39] Yanchao Li, Yongli Wang, Qi Liu, Cheng Bi, Xiaohui Jiang, and Shurong Sun. Incremental
semi-supervised learning on streaming data. Pattern Recognition, 88:383 – 396, 2019.
[40] Shuo Wang, Leandro L. Minku, and Xin Yao. A systematic study of online class imbalance
learning with concept drift. CoRR, abs/1703.06683, 2017.
429
[41] Daniele Zambon, Cesare Alippi, and Lorenzo Livi. Concept Drift and Anomaly Detec-
tion in Graph Streams. IEEE Transactions on Neural Networks and Learning Systems,
29(11):5592–5605, 2018.
[42] Sergio Ramı́rez-Gallego, Bartosz Krawczyk, Salvador Garcı́a, Michal Woźniak, and Fran-
cisco Herrera. A survey on data preprocessing for data stream mining: Current status
and future directions. Neurocomputing, 239:39–57, 2017.
[43] Yan Yan, Tianbao Yang, Yi Yang, and Jianhui Chen. A framework of online learning
with imbalanced streaming data, 2017.
[44] Swee Chuan Tan, Kai Ming Ting, and Fei Tony Liu. Fast anomaly detection for streaming
data. In IJCAI, pages 1511–1516. IJCAI/AAAI, 2011.
[45] Alice Marascu and Florent Masseglia. Parameterless outlier detection in data streams.
In Proceedings of the 2009 ACM Symposium on Applied Computing (SAC), Honolulu,
Hawaii, USA, March 9-12, 2009, pages 1491–1495, 2009.
[46] Ioannis Mitliagkas, Constantine Caramanis, and Prateek Jain. Memory Limited, Stream-
ing PCA. Advances in Neural Information Processing Systems 26: 27th Annual Confer-
ence on Neural Information Processing Systems, pages 1–9, 2013.
[47] Erkki Oja. Simplified neuron model as a principal component analyzer. Journal of
Mathematical Biology, 15(3):267–273, Nov 1982.
[48] Arvind Agarwal Jeff M. Phillips Hal Daume III Suresh Venkatasubramanian. Incremental
Multi-Dimensional Scaling. https://www.cs.utah.edu/~jeffp/papers/incrementalMDS.
pdf, 2010. Online; accessed 19 Feb. 2019.
[49] Thibault Gisselbrecht. Bandit algorithms for real-time data capture on large social medias.
CoRR, abs/1808.10725, 2018.
[50] Thibault Gisselbrecht, Sylvain Lamprier, and Patrick Gallinari. Dynamic data capture
from social media streams: A contextual bandit approach. In Proceedings of the Tenth
International Conference on Web and Social Media, Cologne, Germany, May 17-20,
2016., pages 131–140. AAAI Press, 2016.
[51] Joel A. Tropp, Alp Yurtsever, Madeleine Udell, and Volkan Cevher. Fixed-Rank Ap-
proximation of a Positive-Semidefinite Matrix from Streaming Data. dvances in Neural
Information Processing Systems 30: Annual Conference on Neural Information Process-
ing Systems 2017, 4-9 December, (Nips):1–10, 2017.
[52] Ahmad M Mustafa, Gbadebo Ayoade, Khaled Al-naami, Latifur Khan, Kevin W Hamlen,
Bhavani Thuraisingham, and Frederico Araujo. Unsupervised Deep Embedding for Novel
Class Detection over Data Stream. IEEE International Conference on Big Data (BIG-
DATA) Unsupervised, pages 1830–1839, 2017.
[53] Albert Bifet, Jesse Read, and Geoff Holmes. Efficient Online Evaluation of Big Data
Stream Classifiers Categories and Subject Descriptors. Proceedings of the 21th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, pages
59–68, 2015.
[54] Albert Bifet, Ricard Gavald, Geoff Holmes, and Bernhard Pfahringer. Machine Learning
for Data Streams: With Practical Examples in MOA. The MIT Press, 2018.
430
View publication stats

Es2019 3

Uploaded by

Copyright:

Available Formats

Es2019 3

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Es2019 3

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Recent trends in streaming data analysis, concept drift and analysis of

Conference Paper · August 2021

Albert Bifet Barbara Hammer

SEE PROFILE SEE PROFILE

scikit-multiflow: machine learning for data streams in Python View project

Statistical Physics of Machine Learning View project

The user has requested enhancement of the downloaded file.

1- LTCI, Telecom ParisTech,

2 Problem statement and challenges

(a) Rapid (b) Incremental (c) Gradual (d) Recurring

Challenge 3: How to deal with concept drift?

3 Supervised learning in the streaming context

4 Unsupervised data representation in the streaming con-

BH gratefully acknowledges funding in the frame of the CITEC centre of excel-

View publication stats

You might also like