P. Chan and S. Stolfo. Learning with non-uniform distributions: Effects and a multiclassifier approach. Submitted to Machine Learning Journal, 1999.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....a set of experiments using the credit card fraud data from Chase. We used transactions from the first 8 months (10 95 5 96) for training, the ninth month (6 96) for validating, and the twelfth month (9 96) for testing 3 . Based on the empirical results from the effects of class distributions [2], the desired distribution is 50:50. Since the natural distribution is 20:80, four subsets are generated from each month for a total of 32 subsets. We applied four learning algorithms (C4.5, CART, RIPPER, and BAYES) to each subset and generated 128 base classifiers. Based on our experience with ....

....This is different from the results obtained from previous experiments when meta learning was not used. Note that the desired distribution is empirically determined based on the cost model and does not have to be 50 ; for instance with this data set, it is 30 when the given distribution is 10:90 [2]. Furthermore, to investigate if our approach is indeed fruitful, we ran experiments on the class combiner strategy directly applied to the original data sets from the first 8 months 3 Since credit card transactions have a natural two month business cycle (the time to bill a customer is one ....

[Article contains additional citation context not shown here]

P. Chan and S. Stolfo. Learning with non-uniform distributions: Effects and a multiclassifier approach. Submitted to Machine Learning Journal, 1999.

....and excluded the low cost transactions (those under the overhead amount) These cost based training distributions were used in training base models, and meta classifiers. Unfortunately, the results indicated that the resultant classifiers did not consistently improve their cost performance [7] over varying cost distributions. Other experiments were performed to directly bias the internal strategy of the learning algorithm. One algorithm we have proposed and studied is a close variant of Singer and Schapire s [22] AdaBoost algorithm. AdaBoost is an algorithm that starts with a set of ....

P. Chan and S. Stolfo. Learning with non-uniform distributions: Effects and a multi-classifier approach. Submitted to Machine Learning Journal, 1999.

....and excluded the low cost transactions (those under the overhead amount) These cost based training distributions were used in training base models, and meta classifiers. Unfortunately, the results indicated that the resultant classifiers did not consistently improve their cost performance [7] over varying cost distributions. Other experiments were performed to directly bias the internal strategy of the learning algorithm. One algorithm we have proposed and studied is a close variant of Singer and Schapire s [22] AdaBoost algorithm. AdaBoost is an algorithm that starts with a set of ....

P. Chan and S. Stolfo. Learning with non-uniform distributions: Effects and a multi-classifier approach. Submitted to Machine Learning Journal, 1999.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC