Investigating the Contribution of Privileged Information in Knowledge Transfer LUPI by Explainable Machine Learning

Niharika Gauraha, Henrik Bostrom
Proceedings of the Twelfth Symposium on Conformal and Probabilistic Prediction with Applications, PMLR 204:470-484, 2023.

Abstract

Learning Under Privileged Information (LUPI) is a framework that exploits information that is available during training only, i.e., the privileged information (PI), to improve the classification of objects for which this information is not available. Knowledge transfer LUPI (KT-LUPI) extends the framework by inferring PI for the test objects through separate predictive models. Although the effectiveness of the framework has been thoroughly demonstrated, current investigations have provided limited insights only regarding what parts of the transferred PI contribute to the improved performance. A better understanding of this could not only lead to computational savings but potentially also to novel strategies for exploiting PI. We approach the problem by exploring the use of explainable machine learning through the state-of-the-art technique SHAP, to analyze the contribution of the transferred privileged information. We present results from experiments with five classification and three regression datasets, in which we compare the Shapley values of the PI computed in two different settings; one where the PI is assumed to be available during both training and testing, hence representing an ideal scenario, and a second setting, in which the PI is available during training only but is transferred to test objects, through KT-LUPI. The results indicate that explainable machine learning indeed has the potential as a tool to gain insights regarding the effectiveness of KT-LUPI.

Cite this Paper

BibTeX
@InProceedings{pmlr-v204-gauraha23a, title = {Investigating the Contribution of Privileged Information in Knowledge Transfer LUPI by Explainable Machine Learning}, author = {Gauraha, Niharika and Bostrom, Henrik}, booktitle = {Proceedings of the Twelfth Symposium on Conformal and Probabilistic Prediction with Applications}, pages = {470--484}, year = {2023}, editor = {Papadopoulos, Harris and Nguyen, Khuong An and Boström, Henrik and Carlsson, Lars}, volume = {204}, series = {Proceedings of Machine Learning Research}, month = {13--15 Sep}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v204/gauraha23a/gauraha23a.pdf}, url = {https://proceedings.mlr.press/v204/gauraha23a.html}, abstract = {Learning Under Privileged Information (LUPI) is a framework that exploits information that is available during training only, i.e., the privileged information (PI), to improve the classification of objects for which this information is not available. Knowledge transfer LUPI (KT-LUPI) extends the framework by inferring PI for the test objects through separate predictive models. Although the effectiveness of the framework has been thoroughly demonstrated, current investigations have provided limited insights only regarding what parts of the transferred PI contribute to the improved performance. A better understanding of this could not only lead to computational savings but potentially also to novel strategies for exploiting PI. We approach the problem by exploring the use of explainable machine learning through the state-of-the-art technique SHAP, to analyze the contribution of the transferred privileged information. We present results from experiments with five classification and three regression datasets, in which we compare the Shapley values of the PI computed in two different settings; one where the PI is assumed to be available during both training and testing, hence representing an ideal scenario, and a second setting, in which the PI is available during training only but is transferred to test objects, through KT-LUPI. The results indicate that explainable machine learning indeed has the potential as a tool to gain insights regarding the effectiveness of KT-LUPI.} }
Endnote
%0 Conference Paper %T Investigating the Contribution of Privileged Information in Knowledge Transfer LUPI by Explainable Machine Learning %A Niharika Gauraha %A Henrik Bostrom %B Proceedings of the Twelfth Symposium on Conformal and Probabilistic Prediction with Applications %C Proceedings of Machine Learning Research %D 2023 %E Harris Papadopoulos %E Khuong An Nguyen %E Henrik Boström %E Lars Carlsson %F pmlr-v204-gauraha23a %I PMLR %P 470--484 %U https://proceedings.mlr.press/v204/gauraha23a.html %V 204 %X Learning Under Privileged Information (LUPI) is a framework that exploits information that is available during training only, i.e., the privileged information (PI), to improve the classification of objects for which this information is not available. Knowledge transfer LUPI (KT-LUPI) extends the framework by inferring PI for the test objects through separate predictive models. Although the effectiveness of the framework has been thoroughly demonstrated, current investigations have provided limited insights only regarding what parts of the transferred PI contribute to the improved performance. A better understanding of this could not only lead to computational savings but potentially also to novel strategies for exploiting PI. We approach the problem by exploring the use of explainable machine learning through the state-of-the-art technique SHAP, to analyze the contribution of the transferred privileged information. We present results from experiments with five classification and three regression datasets, in which we compare the Shapley values of the PI computed in two different settings; one where the PI is assumed to be available during both training and testing, hence representing an ideal scenario, and a second setting, in which the PI is available during training only but is transferred to test objects, through KT-LUPI. The results indicate that explainable machine learning indeed has the potential as a tool to gain insights regarding the effectiveness of KT-LUPI.
APA
Gauraha, N. & Bostrom, H.. (2023). Investigating the Contribution of Privileged Information in Knowledge Transfer LUPI by Explainable Machine Learning. Proceedings of the Twelfth Symposium on Conformal and Probabilistic Prediction with Applications, in Proceedings of Machine Learning Research 204:470-484 Available from https://proceedings.mlr.press/v204/gauraha23a.html.

Related Material