(such as a sensor’s gradual failure) and there are more or less similarities among the
concepts before and after the concept drift occurs.
At present, many researchers in related fields have proposed some effective algorithms to detect concept drifts of data streams. A constraint penalty regression combiner is used to track concept drifts in Document Ref. [1]. It is the first time that
integrated classification technology has been introduced into data stream classification
and SEA algorithm has been proposed in Ref. [2]. Gama et al. proposed the DDM
algorithm to detect whether the concept drift occurred in the data through the current
model error rate and standard deviation statistics, and statistically divide the data state
into three states: normal, warning and conceptual drift [3]. However, the DDM method
does not have a good effect on the detection of the gradient drift. In order to solve this
problem, Baena-Garcia et al. proposed the EDDM [4] algorithm to improve the
detection effect of the gradient concept drift. Information entropy is also often used as a
detection indicator for drift detection algorithms [5]. Ross et al. applied the exponentially weighted moving average charts (EWMA) to the detection of concept drift, and
proposed the ECDD algorithm [6], which uses the exponentially weighted moving
average control chart to monitor the error rate when the error rate when a certain
threshold is exceeded, it is considered that the concept drift is detected. The drift
detection method of comparing data distribution is also a very common method.
Recently, Frias. Blarco et al. used moving averages and weighted moving averages to
measure changes in the data stream, and experimentally proved that these two statistics
are applicable to abrupt drift and gradual drift, respectively [7], Farid proposed
Changes in data distribution such as CUSUM cumulative sum and exponential
weighted moving average method emphasize the problem of metric determination
during detection [8]. Literature [9] compares two data distributions that can continuously expand the window, and then selects different window strategies for detection
[9]. Sakthithasan proposed the SeqDrift detection algorithm, which uses reservoir
sampling for memory management and improves the sensitivity of conceptual drift
detection under slowly varying conditions [10, 11].
This paper proposed a detection method to concept drift by detecting the differences
of frequency distribution between data blocks, which is introduced in Sect. 2.
According to the method, In Sect. 3 we did related experiments and results analysis.
Finally, the conclusions are given according to the results analysis.
2 Concept Drift Detection Based on Kolmogorov–Smirnov
Test
2.1 Relevant Definitions
In order to facilitate the elaboration of relevant contents of this paper, the following
relevant definitions are given at first:
Definition 1 Concept drift: setting that X
t
n is a sample and y
t
n is sample’s class label; the
joint probability distribution of the sample generated by the data source changes with time.
Supposing that every time t arrives at a data block D t , D t ¼ fðX
t
1 ; y
t
1 Þ; ðX
t
2 ; y
t
2 Þ;
274
Z. Wang and W. Wang
concepts before and after the concept drift occurs.
At present, many researchers in related fields have proposed some effective algorithms to detect concept drifts of data streams. A constraint penalty regression combiner is used to track concept drifts in Document Ref. [1]. It is the first time that
integrated classification technology has been introduced into data stream classification
and SEA algorithm has been proposed in Ref. [2]. Gama et al. proposed the DDM
algorithm to detect whether the concept drift occurred in the data through the current
model error rate and standard deviation statistics, and statistically divide the data state
into three states: normal, warning and conceptual drift [3]. However, the DDM method
does not have a good effect on the detection of the gradient drift. In order to solve this
problem, Baena-Garcia et al. proposed the EDDM [4] algorithm to improve the
detection effect of the gradient concept drift. Information entropy is also often used as a
detection indicator for drift detection algorithms [5]. Ross et al. applied the exponentially weighted moving average charts (EWMA) to the detection of concept drift, and
proposed the ECDD algorithm [6], which uses the exponentially weighted moving
average control chart to monitor the error rate when the error rate when a certain
threshold is exceeded, it is considered that the concept drift is detected. The drift
detection method of comparing data distribution is also a very common method.
Recently, Frias. Blarco et al. used moving averages and weighted moving averages to
measure changes in the data stream, and experimentally proved that these two statistics
are applicable to abrupt drift and gradual drift, respectively [7], Farid proposed
Changes in data distribution such as CUSUM cumulative sum and exponential
weighted moving average method emphasize the problem of metric determination
during detection [8]. Literature [9] compares two data distributions that can continuously expand the window, and then selects different window strategies for detection
[9]. Sakthithasan proposed the SeqDrift detection algorithm, which uses reservoir
sampling for memory management and improves the sensitivity of conceptual drift
detection under slowly varying conditions [10, 11].
This paper proposed a detection method to concept drift by detecting the differences
of frequency distribution between data blocks, which is introduced in Sect. 2.
According to the method, In Sect. 3 we did related experiments and results analysis.
Finally, the conclusions are given according to the results analysis.
2 Concept Drift Detection Based on Kolmogorov–Smirnov
Test
2.1 Relevant Definitions
In order to facilitate the elaboration of relevant contents of this paper, the following
relevant definitions are given at first:
Definition 1 Concept drift: setting that X
t
n is a sample and y
t
n is sample’s class label; the
joint probability distribution of the sample generated by the data source changes with time.
Supposing that every time t arrives at a data block D t , D t ¼ fðX
t
1 ; y
t
1 Þ; ðX
t
2 ; y
t
2 Þ;
274
Z. Wang and W. Wang