![](bg2.png)
at stage 2. Here the unsupervised drift warning
indicator is independent with target and only rely on
distribution changes. Then the supervised indicator
confirms the drift if there is a significant variation in
the performances. The learning and adaption of the
stream with proposed drift detector is through
querying the uncertainty samples by batch based
active learning.
The paper organization is as follows. Section 2
presents the work related to drift detection and
adaption, Section 3 describes the proposed batch
based drift detection and adaption. Data stream
description, experimental setup results and
discussion are presented in Section 4. Finally, this
paper concludes in section 5.
2 RELATED WORK
The drift detection methods are categorized into two:
(i) active (ii) passive. The active methods works
based on drift detection and adaption. The drift can
be detected using
Hypothesis Tests: Validates the NULL
hypothesis, i.e., the two samples are derived
from same distribution (Patist, 2007 and
Nishida and Yamauchi, 2007).
Change-point Method: Tracking the point of
change of the behaviour of distribution
function (Hawkins, Qiu and Wook kang,
2003).
Sequential Hypothesis Test: Constantly
monitoring the stream until it attains enough
confidence to accept or reject the hypothesis
test (Wald, 1945).
Change Detection Test: Identifies the drift
based on a threshold on a classification error or
on a feature value (Bifet and Gavalda, 2007).
The use of Hellinger distance and adaptive
Cumulative SUM test for change detection
between data chunks is also studied in (Ditzler
and Polikar, 2011 )
Once the drift is detected in an evolving stream,
the learning framework adapts it by learning a new
model on current knowledge and forgetting the old.
The forgetting mechanisms are of selecting random
samples to filter, or weighing the samples based on
their age so that the sample with largest age is
forgotten. Another method is of windowing, once
the change is detected, the samples which are
relevant to current learner only retained in the
window. But the size of the window is critical here,
the adaptive window size mechanisms based on
Intersection of Confidence Interval (ICI) are
proposed in (Alippi, Boracchi and Roveri, 2011).
Unlike the drift detection and adaption methods the
passive approaches, constantly update the model to
adapt the change with new evolving data. The model
updation is carried out by resetting the parameters
(single classifier adaption) or add/remove/update a
classifier in an ensemble.
So far the drift detection methods for supervised
learning are intended for balanced classes and used
supervised performance estimates such as error,
accuracy and four rates such as TPR, NPR, PPV and
NPV. However, recently, few drift detection
methods are proposed for imbalanced streaming
distributions. The Drift Detection Method Online
Class Imbalance (DDM_OCI) (Wang et al, 2013) is
a modification to DDM (Gama et al, 2004). Unlike
DDM, whose focus is on the change detection in
over all error rate, DDM_OCI tracks changes in TPR
assuming that the drift in the distribution leads to
significant changes when there is an imbalance in
the stream. But, DDM_OCI is quite sensitive to the
dynamic imbalance rate of change than the real
concept drift which results in many false positives.
Instead of tracking the changes only in TPR (Wang
and Abraham, 2015) proposed a Linear Four rates
tracking mechanism for drift detection. If significant
change is detected in any of the performance rates
such as TPR, FPR, PPV and NPV then the drift
signal is alarmed. In (Brzezinski and Stefanowski,
2017) proposed a Prequential AUC based drift
detection mechanism which identifies the drift in
Prequential AUC by Page-Hinkley test. In (Yu et al,
2019) proposed a two-layer drift detection method
where layer 1 adapts LFR and layer 2 is based on
permutation test and both layers are of supervised.
All these methods mainly based on tracking the
changes in supervised performance estimators and
can prone to false positives due to the sensitivity of
TPR towards dynamic imbalance rather than drift at
high degree of imbalance cases. We propose two-
stage drift detection based on unsupervised and
supervised change tracking.
3 PROPOSED METHOD
Figure 1 depicts the flow diagram for proposed drift
detection and adaption method. Here the Learning of
the stream as well as the adaption to the drift is
handled based on batch based active learning. The
drift detection is carried out in two stages, named it
as Kolmogorov-Smirnov_Area under Curve
(KS_AUC) method. This drift detector assumes
initial training set is labelled and the rest of the
stream evolved as unlabelled.
DATA 2019 - 8th International Conference on Data Science, Technology and Applications
226