Distinguishing Cause and Effect: Inferring causal direction in two variables

Lohit Jain and Balram Meena

Abstract:

The task of attributing cause and effect helps us in finding factors affecting economy, human health, global warming and can help us in solving many problems. The standard way to detect causal direction is to perform random experiments, which can be expensive, unethical or even impossible to perform. Inferring the same from already collected data can solve many problems. In this project we dealt with finding causal direction among two variables, given their distribution. This was part of the Cause Effect pair challenge. We used functional noise model and deterministic relational models to find the cost involved in causal directions for the two variables . We determined the existence of a relation using a SVM classifier on features extracted from the distribution of data.

Are you thinking: what is so tough about this problem??

Well then, guess for this image, is the causal direction X->Y or Y->X.

Ans Here:X=>Y , X: Altitude, Y: Precipitation

Our Results

The final Area under the ROC score was found to be 0.658. This easily surpassed the python benchmark of 0.570 by the challenge authorities. The accuracy of the classifier was about 73%.

Links

Our Final report

Proposal

Poster

Code

External links and/or Interesting reads

Causality challenge 2013
Causality Inference group at Max Planck