IEEE Computational Intelligence Magazine - August 2020 - 23

yet this approach is heavily dependent on the distance
between decision surfaces.
From a data-mining standpoint, prior probability shifts can
be tracked and estimated with parametric distribution estimation, with the assumption that data comprises mixtures of distributions [26], [27]. Still, this is a strong assumption with few
usable cases. From this perspective, "utilities," which are assessed
by the frequency an instance occurs in the dataset, can be
extracted from a data stream and tracked to detect drifts [66].
This method is similar to an unsupervised detection of drifts in
a stream based on the importance of each pattern within a
window. Importance measures, however, are difficult to define,
and poor choices may hinder the model.
More recently, regional concept drift detection strategies
were proposed to detect drifts that occur in local regions but
do not cause significant overall changes in data [67], [68]. A
local drift degree (LDD) measures regional drifts by determining whether the local density discrepancies of two i.i.d. distributions are similar [69]. Similarly, a shift detector based on
nearest neighbor-based density variation (NN-DVI) was proposed according to a k-nearest neighbor-based space-partitioning schema (NNPS) [68]. This schema improves the sensitivity
of regional drift detection by transforming unmeasurable discrete data instances into a set of shared subspaces for density
estimation. Due to the local metrics based on windows, they
are slower compared to other shift detectors, with marginal
improvements depending on shift characteristics.
B. Passive Approaches

Passive approaches assume that data will change; thus, the
model itself adapts for every new data it receives irrespective of
whether a drift occurred. Despite being generally robust to
gradual or incremental shifts, passive methods in off-line cases
might perform differently depending on the type and constraints of a shift. Some of the main passive approaches for offline problems are based on transductive and semi-supervised
learning models, described in Subsection VI.C.
The use of intrinsic local mode functions in a model can be
used to adapt to shifts in time [70]. This method is based on the
empirical decomposition of the data, allowing for a wellbehaved Hilbert transform for each function. As it concerns
local time scales of data, the model is redefined at each instant,

L2 Distance

However, multivariate statistical testing can be computationally
expensive, so dimensionality reduction along with maximum
mean discrepancy tests, Kolmogorov-Smirnov tests with Bonferroni correction, and chi-squared tests are possible approaches
[57]. Still, most of these statistical tests depend on the underlying distribution. In addition, in the case of the drift causing
small observable changes by the statistical method, the detector
tends to perform poorly [58]. A semi-supervised approach
monitors changes in the classifier's confidence to detect shift.
The approach proposed to detect multi-class novelty and concept shifts using dynamic windows to mitigate the trade-off
between performance during stable periods and delayed detection [59], [60]. The method suffered from low execution speed,
so dynamic programming and selective executions of the
detection module were implemented as an improvement [59].
The detection and classification framework proposed is an
ensemble of k-nearest neighbors (k-NN) classifiers. Each test
instance is classified, and the classifier's confidence score is
stored in a vector. Then, a sliding window approach is used to
detect significant changes. If a change is detected, the framework determines a new chunk boundary of scores and updates
the classifier. The dynamic implementation improves the framework speed; however, the method still estimates distributions
and implements a recursive calculation of change detection
scores. Despite the sporadic calls of the detector, burdensome
calculations may cause the method to be prohibitively slow in
streaming scenarios.
A series of methods employing exponentially weighted
moving average (EWMA) for the detection of some dataset shift
problems, particularly covariate shift, have been proposed in the
literature [61], [62]. These methods use statistical process control
charts to detect shifts in the input covariate. The shift detection
based on EWMA works in two stages: (1) a control chart is used
to detect the dataset shift in the data stream, and (2) a statistical
hypothesis test is used retrospectively in the testing phase to validate the shift detected by the first stage. During the test phase,
input data is continuously monitored by the EWMA chart and,
when control limits are exceeded, the method considers that a
shift has occurred. For example, in applications with EEG-based
BCIs, detection based on EWMA was successful in identifying
covariate shifts in motor imagery-based EEG with principal
component analysis [20], [21].
In unlabeled data streams, a possible approach is to use
unsupervised anomaly detection to define transitions of concepts [63]. However, this approach requires meticulous work
as search window sizes would need to be well tuned depending on the speed pattern and intensity of the shift, which
could make this approach unfeasible. The Plover [64] algorithm facilitates the detection of drift in unsupervised streams
using statistical moments to measure data stability, but it
requires independent and identically distributed (i.i.d.) input
data. For unlabeled data streams, a semi-supervised diversity
measure was proposed as a drift detection method [65],
which evaluates the diversity of a pair of classifiers. Drifts can
be detected when the disagreement between them increases,

0.6
0.4
0.2
0
A

FIGURE 8 Incremental shift represented by pictures of a person in her
infancy (initial state A), childhood, teenage years, and adulthood (final
state B). L2 distance to state A was calculated with OpenFace [50].

AUGUST 2020 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

IEEE Computational Intelligence Magazine - August 2020

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - August 2020