連続データと離散データの間の因果関係の同定

抄録

This paper considers causal discovery between discrete and continuous variables based on additive noise model. In many database, some fields are discrete while others continuous. However, the previous notion assumes that all the variables are either discrete or continuous. In this paper, we prove that for discrete (m values) and continuous variables X, Y , causality X ! Y cannot be identified for m = 2 under regular conditions, and conjecture that X ! Y can be identified for m · 3, and that Y ! X can be identified for any m. Several experiments support those properties successfully. Furthermore, using R, the program language, we implemented causal discovery between X ="month" and Y ="average temperature" in the data provided by the US National Weather Service Weather Forecast Office.