1:8 J. Gama et al.
but still abrupt switch to burning. One challenge for learning is that the feedback (the
ground truth of mass flow) is not available at all, it can only be approximately esti-
mated by retrospectively inspecting the historical data. An additional challenge is to
deal with specific one-sided outliers that can be easily mistaken for changes.
Traditional approaches (such as ADWIN) for explicit change detection based on the
monitoring of the raw sensor signal or streaming error of the regressors give reason-
able results. They can be improved by considering the peculiarities of the application.
2.5.2. Management and strategic planning. The Smart Grid (SG) is an electric system that
uses two-way digital information, cyber-secure communication technologies, and com-
putational intelligence in an integrated fashion across heterogeneous and distributed
electricity generation, transmission, distribution and consumption to achieve energy
efficiency. A key and novel characteristic of SG’s is the intelligent layer that analyzes
the data produced by smart meters allowing companies to develop powerful new capa-
bilities in terms of grid management, planning and customer services for energy effi-
ciency. The advent of SG’s has changed the way energy is produced, priced and billed.
The key aspect of SG’s is distributed energy production, namely renewable energies.
The penetration of renewable energies (solar, wind, etc.) is increasing fast and power
forecasting becomes an important factor in defining the operation planning policies to
be adopted by a Transmission System Operator.
When observing the literature in wind power prediction [Monteiro et al. 2009],
one realizes that most proposals are based on an off-line training mode, building a
static model that is then used to produce predictions. This option rely in assump-
tions of stationarity of the wind electric power model, which must be strongly ques-
tioned [Bremnes 2004; Bessa et al. 2009]. Using real data from three distinct wind
parks, [Bessa et al. 2009] presents the merits of on-line training against off-line train-
ing of neural networks. The authors point out the evolving nature of data and the
presence of concept drift in wind pattern behavior.
2.5.3. Personal assistance and information. Text classification has been a popular topic in
machine learning for decades. However, interesting applications related to the problem
of concept drift appeared relatively recently. Examples of text stream applications in-
clude e-mail classification [Carmona-Cejudo et al. 2010], e-mail spam detection [Lind-
strom et al. 2010] and sentiment classification [Bifet and Frank 2010]. Sentiment clas-
sification is a popular task in social media monitoring, customer feedback analysis and
other applications.
The main source of concept drift in e-mail classification and spam filtering are due to
changing e-mail content and presentation (virtual drift), as well as adaptive behaviour
of spammers trying to overcome spam filters (may be virtual or real). Besides, users
may change their attitude towards particular categories of e-mails starting or stopping
to consider them spam (real drift). In sentiment classification the vocabulary used to
express positive and negative sentiments may change over time. Since the collection
of documents is not static (virtual drift, novelties), the feature space representing the
current collection is dynamic that may require specific updates of the models.
Various adaptive learning strategies have been used in this domain, including indi-
vidual methods like case-based reasoning, and ensembles, either evolving or with an
explicit detection of changes by means of change detectors (Section 3.2).
Availability of feedback is a serious challenge in personal assistance and informa-
tion. The dilemma is that if feedback is easily available, that implies no need for au-
tomated predictions. In e-mail classification we can hope that from time to time we
will receive feedback from the user in case of misclassifications or can design an active
learning system (e.g. [Zliobaite et al. 2013]), which from time to time asks the user to
ACM Computing Surveys, Vol. 1, No. 1, Article 1, Publication date: January 2013.