blablahaha / concept-drift

Algorithms for detecting changes from a data stream.

AdWin: ADaptive sliding WINdow algorithm

Based on paper:

Bifet and R. Gavalda. 2007. Learning from Time-Changing Data with Adaptive Windowing

class concept_drift.adwin.AdWin(
    delta=0.002, max_buckets=5, min_clock=32, min_win_len=10, min_sub_win_len=5
)

Parameters
delta	Confidence value
max_buckets	Max number of buckets within one bucket row
min_clock	Min number of new data for starting to reduce window and detect change
min_window_len	Min window length for starting to reduce window and detect change
min_sub_window_len	Min sub-window length, which is split from whole window

Methods

set_input(value)

Set input value to the drift detector - ADWIN.

Type	Input - Output
Parameters	value: Input value
Return	Boolean: Whether has drift

Example

from concept_drift.adwin import AdWin

adwin = AdWin()
for i in range(1000):
    if adwin.set_input(i):
        print("Here is a drift")

Page-Hinckley Test

Based on Chapter 5.1:

Gama, J., Sebastião, R., & Rodrigues, P. P. (2013). On evaluating stream learning algorithms

class concept_drift.page_hinkley.PageHinkley(
    delta_=0.005, lambda_=50, alpha_=1 - 0.0001
)

Parameters
delta_	Magnitude of changes that are allowed
lambda_	Difference threshold
alpha_	Adaption variable for lambda adjustment

Methods

set_input(value)

Set input value to the drift detector - PageHinkley.

Type	Input - Output
Parameters	value: Input value
Return	Boolean: Whether has drift

Test

Used the elecNormNew dataset;
Used GaussianNB as based classification;
Used accuracy_score as input date for change detector;
Used the prequential evaluation;

GaussianNB:
Mean acc within the window 1000: 0.7289912189511405

Page-Hinkley:
Drift detection: 8
Mean acc within the window 1000: 0.7342849306584686

AdWin:
Drift detection: 139
Mean acc within the window 1000: 0.7496421003738032

concept-drift's Issues

Thanks for your code!

Thanks for your code!
But i want a regressor version,can you upload it？
I‘m very grateful.

Epsilon cut calculation

Hello,
while going through your code, I tried to understand the calculation of epsilon_cut. This is a code snippet from __reduce_expressions() function:

m = 1 / (n0 - self.min_length_sub_window + 1) + 1 / (n1 - self.min_length_sub_window + 1)
d = log(2 * log(self.width) / self.delta)
variance = self.variance / self.width
epsilon = sqrt(2 * m * variance * d) + 2 / 3 * m * d

I do not really understand how you come to the d value. In the original publication it is ln (2 / delta^' ). Maybe I am just not getting some easy part, but I don't see you switching ln to log and adding log to self.width.

(BTW, I agree on everything else)