Events

Professor Fan Wang
Department of Economics
Assistant Professor of Economics

Association rules are conditional statements, suggesting a value for a predicted variable y if certain values of the predictor x= (x^1,…,x^m) occur. They are widely used in machine learning, where each rule’s weight depends on past predictive performance. We consider a simple evaluation-and-aggregation model, where the degree of credence of each rule is additive in its past successes. Given past observations and a new prediction problem, a reasoner can generate either (i) a binary relation over possible values of y – “at least as likely as” or (ii) a quantitative probability vector over them. We axiomatize both models, providing conditions on these presumably-observable data, that are equivalent to the corresponding association rule model. Generalizations and applications are discussed.