A rule-based calibration method for models with multiple rank-ordered choices

摘要

A model of multiple-ranked shopping center destination choice was developed using a rule-based expert systems format. The motivation behind this approach was twofold. First, this model avoids the need to collect data regarding unchosen alternatives that the consumer faces, as is typically required by ranked or “exploded” multinomial logit or probit models. The rule-based model requires only a random sample of decision makers where their actual stated preference are observed. The second motivation is to avoid misallocation of demand to alternatives, as is often the case with conventional models that ignore ranked data. This model is also different from others since all of the observed data are chosen alternatives but chosen with unknown frequency.The model that was developed is based on fuzzily defined clusters of survey respondents that make similar-ranked choices for similar reasons. Once clusters are defined through conventional cluster analysis, rules are created to classify respondents into the groups. No parameters are estimated by this method, but the overall meaning of the rules can be discerned through sensitivity analysis where variables are selectively changed and forecast results are recorded.Calibration of the model was based on a sample of 813 home-based shoppers in St. Catharines, Ontario. It was found that about 50 rules were needed in order to correctly predict the top three-ranked choices with accuracy levels in the mid 70% range. Sensitivity analysis indicated that the model was making sensible predictions, given changes in the input data.