When causality matters for prediction: investigating the practical tradeoffs

Robert E. Tillman, Peter Spirtes
Proceedings of Workshop on Causality: Objectives and Assessment at NIPS 2008, PMLR 6:137-146, 2010.

Abstract

Recent evaluations have indicated that in practice, general methods for prediction which do not account for changes in the conditional distribution of a target variable given feature values in some cases outperform causal discovery based methods for prediction which can account for such changes. We investigate some possibilities which may explain these findings. We give theoretical conditions, which are confirmed experimentally, for when particular manipulations of variables should not affect predictions for a target. We then consider the tradeoff between errors related to causality, i.e. not accounting for changes in a distribution after variables are manipulated, and errors resulting from sample bias, overfitting, and assuming specific parametric forms that do not fit the data, which most existing causal discovery based methods are particularly prone to making.

Cite this Paper

BibTeX
@InProceedings{pmlr-v6-tillman10a, title = {When causality matters for prediction: investigating the practical tradeoffs}, author = {Tillman, Robert E. and Spirtes, Peter}, booktitle = {Proceedings of Workshop on Causality: Objectives and Assessment at NIPS 2008}, pages = {137--146}, year = {2010}, editor = {Guyon, Isabelle and Janzing, Dominik and Schölkopf, Bernhard}, volume = {6}, series = {Proceedings of Machine Learning Research}, address = {Whistler, Canada}, month = {12 Dec}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v6/tillman10a/tillman10a.pdf}, url = {https://proceedings.mlr.press/v6/tillman10a.html}, abstract = {Recent evaluations have indicated that in practice, general methods for prediction which do not account for changes in the conditional distribution of a target variable given feature values in some cases outperform causal discovery based methods for prediction which can account for such changes. We investigate some possibilities which may explain these findings. We give theoretical conditions, which are confirmed experimentally, for when particular manipulations of variables should not affect predictions for a target. We then consider the tradeoff between errors related to causality, i.e. not accounting for changes in a distribution after variables are manipulated, and errors resulting from sample bias, overfitting, and assuming specific parametric forms that do not fit the data, which most existing causal discovery based methods are particularly prone to making.} }
Endnote
%0 Conference Paper %T When causality matters for prediction: investigating the practical tradeoffs %A Robert E. Tillman %A Peter Spirtes %B Proceedings of Workshop on Causality: Objectives and Assessment at NIPS 2008 %C Proceedings of Machine Learning Research %D 2010 %E Isabelle Guyon %E Dominik Janzing %E Bernhard Schölkopf %F pmlr-v6-tillman10a %I PMLR %P 137--146 %U https://proceedings.mlr.press/v6/tillman10a.html %V 6 %X Recent evaluations have indicated that in practice, general methods for prediction which do not account for changes in the conditional distribution of a target variable given feature values in some cases outperform causal discovery based methods for prediction which can account for such changes. We investigate some possibilities which may explain these findings. We give theoretical conditions, which are confirmed experimentally, for when particular manipulations of variables should not affect predictions for a target. We then consider the tradeoff between errors related to causality, i.e. not accounting for changes in a distribution after variables are manipulated, and errors resulting from sample bias, overfitting, and assuming specific parametric forms that do not fit the data, which most existing causal discovery based methods are particularly prone to making.
RIS
TY - CPAPER TI - When causality matters for prediction: investigating the practical tradeoffs AU - Robert E. Tillman AU - Peter Spirtes BT - Proceedings of Workshop on Causality: Objectives and Assessment at NIPS 2008 DA - 2010/02/18 ED - Isabelle Guyon ED - Dominik Janzing ED - Bernhard Schölkopf ID - pmlr-v6-tillman10a PB - PMLR DP - Proceedings of Machine Learning Research VL - 6 SP - 137 EP - 146 L1 - http://proceedings.mlr.press/v6/tillman10a/tillman10a.pdf UR - https://proceedings.mlr.press/v6/tillman10a.html AB - Recent evaluations have indicated that in practice, general methods for prediction which do not account for changes in the conditional distribution of a target variable given feature values in some cases outperform causal discovery based methods for prediction which can account for such changes. We investigate some possibilities which may explain these findings. We give theoretical conditions, which are confirmed experimentally, for when particular manipulations of variables should not affect predictions for a target. We then consider the tradeoff between errors related to causality, i.e. not accounting for changes in a distribution after variables are manipulated, and errors resulting from sample bias, overfitting, and assuming specific parametric forms that do not fit the data, which most existing causal discovery based methods are particularly prone to making. ER -
APA
Tillman, R.E. & Spirtes, P.. (2010). When causality matters for prediction: investigating the practical tradeoffs. Proceedings of Workshop on Causality: Objectives and Assessment at NIPS 2008, in Proceedings of Machine Learning Research 6:137-146 Available from https://proceedings.mlr.press/v6/tillman10a.html.

Related Material