- author = "Phil Truscott and Michael F. Korns",
- title = "Predicting Product Choice with Symbolic Regression and Classification",
- booktitle = "Genetic Programming Theory and Practice XIII",
- year = "2015",
- editor = "Rick Riolo and William P. Worzel and M. Kotanchek and A. Kordon",
- series = "Genetic and Evolutionary Computation",
- pages = "203--217",
- address = "Ann Arbor, USA",
- month = "14-16 " # may,
- publisher = "Springer",
- keywords = "genetic algorithms, genetic programming, Abstract regression grammars Genetic algorithms Symbolic regression Classification Non-linear regression",
- isbn13 = "978-3-319-34223-8",
-
URL = "
http://www.springer.com/us/book/9783319342214",
-
DOI = "
doi:10.1007/978-3-319-34223-8_12",
- abstract = "Market researchers often conduct surveys to measure how much value consumers place on the various features of a product. The resulting data should enable managers to combine these utility values in different ways to predict the market share of a product with a new configuration of features. Researchers assess the accuracy of these choice models by measuring the extent to which the summed utilities can predict actual market shares when respondents choose from sets of complete products. The current paper includes data from 201 consumers who gave ratings to 18 cell phone features and then ranked eight complete cell phones. A simple summing of the utility values predicted the correct product on the ranking task for 22.8 percent of respondents. Another accuracy measurement is to compare the market shares for each product using the ranking task and the estimated market shares based on summed utilities. This produced a mean absolute difference between ranked and estimated market shares of 7.8 percent. The current paper applied two broad strategies to improve these prediction methods. Various evolutionary search methods were used to classify the data for each respondent to predict one of eight discrete choices. The fitness measure of the classification approach seeks to reduce the Classification Error Percent (CEP) which minimizes the percent of incorrect classifications. This produced a significantly better fit with the hit rate rising from 22.8 to 35.8 percent. The mean absolute deviation between actual and estimated market shares declined from 7.8 to 6.1 percent (p. <0.01). A simple language specification will be illustrated to define symbolic regression and classification searches.",
-
notes = "
Part of \cite{Riolo:2015:GPTP} Published after the workshop in 2016",
