Disclosed is a feature for estimating a causal relation with which it is possible to solve problems due to prior art, and which does not require preliminary settings of a regression model. One embodiment of the present invention pertains to a causal relation learning device having: a feature quantity calculation unit for accepting as inputs a correct answer label, which is a classification label that pertains to a causal relation of time-series data and is grouped into three or more classes, and time-series data that corresponds to the correct answer label, and calculating a feature quantity of the time-series data; and a classifier learning unit for learning a classifier using a set of the feature quantity and the correct answer label so that the output of a classifier for the feature quantity equals the maximum value of output value of the correct answer label.
展开▼