[4Xin2] Poster session 2

[4Xin2-01] An analysis of computational goals in observational causal induction among human multivalued variables.

〇Sota Aoki¹, Hiroto Ichino², Tatsuji Takahashi¹, Kohki Higuchi³ (1.Tokyo Denki University, 2.Graduate of Tokyo Denki University, 3.Chubu University)

Keywords:Causal Induction, Dual process theory, Observational Causal Induction

Cognitive science studies have explored mathematical models for understanding how humans learn causal relationships from observational data.Almost all of them use a framework in which the variables of interest are limited to binary variables. This is an oversimplification that is contrary to the complexity of the real-world problems that humans face.To understand human causal induction and apply cognitive science findings to AI and real-world decision-making, we must extend the theoretical framework to multivalued variables.Here, many previous studies assume that the initial stage of human causal induction is based on correlation detection. In a binary variable framework, association is indistinguishable from correlation, eliminating the need to separately consider association.However, when the variable of interest can take on three or more values, correlation and association are by definition distinct, and it becomes necessary to verify whether humans are specifically detecting correlation or association in causal induction.In our study, we conducted a cognitive experiment on inferring causal relationships among multivalued variables and verified the results.