Open WeiFoo opened 8 years ago
the AUCs report some large changes, particular for random forests. are they statistically significant? (Scott Knott!)
I didn't see any large changes in AUC, you can check the raw results here by SK test, F measure #13 Precision #14 AUC #15
what do you mean by predicting only large delay? is this a two class problems?
yes, only two classes here, large delay and non delay. Originally, there're high, medium, low delays and non-delay labels. I use high as positive class, all the others are negative class
is the large class rare? does training need smote? can smote be tuned as well as the learner?
percentage of large class are listed here:
apache: 373/5725=6.5%, duraspace: 99/2078=4.7%, jbos: 597/8281=7.2%, jira: 198/6193=3.2% moodel:61/1313=4.6% spring:91/5202=1.7% wso2: 118/4586=2.5%
Both duraspace and jira get very high(98+) scores when tuning goals are F and precision, separately. But the percentage of large class are not very high. I will try to smote on those data sets and see whether have any improvements. personally, I don't consider tuning smote, based on other results, like text mining results.
is this a cross-val? or is this past data predicting for future data?
no, it's not cross evaluation. the data from Wollongong was already split into training and testing data, I sent email to ask for details about how to generate them.
update: according to Morakot, he can guarantee that for each class, testing data is generated after training data.
Settings
Results