2007 IEEE International Conference on Information Reuse and Integration
We demonstrate the use of reconstructability analysis (RA) on the UCI Australian Credit dataset to reduce the number of input variables for two different analysis tools. Using 14 variables, an artificial neural net (NN) is able to predict whether or not credit was granted, with a 79.1% success rate. RA preprocessing allows us to reduce the number of independent variables from 14 to two different sets of three, which have success rates of 77.2% and 76.9% respectively. The difference between these rates and that of the 14-variable NN is not statistically significant. The three-variable rulesets given by RA achieve success rates of 77.8% and 79.7%. Again, the difference between those values and the 14-variable NN is not statistically significant, that is, our approach provides a three-variable model that is competitive with the 14-variable equivalent.
This is the post-print version (author's manuscript). The final published version, ©2007 IEEE, is available from the publisher:
Locate the Document
S. Shervais and M. Zwick, "Using Reconstructability Analysis for Input Variable Reduction: A Business Example," [Post-print] 2007. IEEE International Conference on Information Reuse and Integration, Las Vegas, IL, 2007, pp. 532-537.