A Betting Function for addressing Concept Drift with Conformal Martingales

Charalambos Eliades, Harris Papadopoulos
Proceedings of the Eleventh Symposium on Conformal and Probabilistic Prediction with Applications, PMLR 179:219-238, 2022.

Abstract

An important issue that appears when using Conformal Martingales (CM) for detecting Concept Drift (CD), is that martingale values get very close to zero when the data generating mechanism remains the same for a large number of instances. In such cases, the martingale takes a long time to recover, resulting in detection delays or even totally failing to detect the occurrence of a CD. To address this issue we propose a new betting function we call Cautious, that avoids betting when there is no evidence that any change is taking place, therefore preventing the continuous reduction of the martingale value. The proposed betting function can be built on top of any existing betting function to mitigate the aforementioned problem. In this work, we combine it with the kernel and histogram betting functions and compare its performance with that of the two original betting functions as well as that of existing methods for addressing CD on five datasets.

Cite this Paper


BibTeX
@InProceedings{pmlr-v179-eliades22a, title = {A Betting Function for addressing Concept Drift with Conformal Martingales}, author = {Eliades, Charalambos and Papadopoulos, Harris}, booktitle = {Proceedings of the Eleventh Symposium on Conformal and Probabilistic Prediction with Applications}, pages = {219--238}, year = {2022}, editor = {Johansson, Ulf and Boström, Henrik and An Nguyen, Khuong and Luo, Zhiyuan and Carlsson, Lars}, volume = {179}, series = {Proceedings of Machine Learning Research}, month = {24--26 Aug}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v179/eliades22a/eliades22a.pdf}, url = {https://proceedings.mlr.press/v179/eliades22a.html}, abstract = {An important issue that appears when using Conformal Martingales (CM) for detecting Concept Drift (CD), is that martingale values get very close to zero when the data generating mechanism remains the same for a large number of instances. In such cases, the martingale takes a long time to recover, resulting in detection delays or even totally failing to detect the occurrence of a CD. To address this issue we propose a new betting function we call Cautious, that avoids betting when there is no evidence that any change is taking place, therefore preventing the continuous reduction of the martingale value. The proposed betting function can be built on top of any existing betting function to mitigate the aforementioned problem. In this work, we combine it with the kernel and histogram betting functions and compare its performance with that of the two original betting functions as well as that of existing methods for addressing CD on five datasets. } }
Endnote
%0 Conference Paper %T A Betting Function for addressing Concept Drift with Conformal Martingales %A Charalambos Eliades %A Harris Papadopoulos %B Proceedings of the Eleventh Symposium on Conformal and Probabilistic Prediction with Applications %C Proceedings of Machine Learning Research %D 2022 %E Ulf Johansson %E Henrik Boström %E Khuong An Nguyen %E Zhiyuan Luo %E Lars Carlsson %F pmlr-v179-eliades22a %I PMLR %P 219--238 %U https://proceedings.mlr.press/v179/eliades22a.html %V 179 %X An important issue that appears when using Conformal Martingales (CM) for detecting Concept Drift (CD), is that martingale values get very close to zero when the data generating mechanism remains the same for a large number of instances. In such cases, the martingale takes a long time to recover, resulting in detection delays or even totally failing to detect the occurrence of a CD. To address this issue we propose a new betting function we call Cautious, that avoids betting when there is no evidence that any change is taking place, therefore preventing the continuous reduction of the martingale value. The proposed betting function can be built on top of any existing betting function to mitigate the aforementioned problem. In this work, we combine it with the kernel and histogram betting functions and compare its performance with that of the two original betting functions as well as that of existing methods for addressing CD on five datasets.
APA
Eliades, C. & Papadopoulos, H.. (2022). A Betting Function for addressing Concept Drift with Conformal Martingales. Proceedings of the Eleventh Symposium on Conformal and Probabilistic Prediction with Applications, in Proceedings of Machine Learning Research 179:219-238 Available from https://proceedings.mlr.press/v179/eliades22a.html.

Related Material