machineJS

machineJS copied to clipboard

data-formatter is going to start taking the log of regression values in most cases. when making final predictions, un-log the values when making stage 0 predictions, modify the output column...

ClimbsRocks

this is a far future idea. when we have disagreement between different algorithms, that is a difficult case. when we have disagreement, mark it as such. aggregate all the samples...

ClimbsRocks

if true, we will modify the process to be simpler. the second round of machineJS, run by ensembler, will pick the algorithm with the best score, and just use that,...

ClimbsRocks

https://github.com/yandex/rep/ if it works as well as it says it does, it could provide a bunch of awesome new algos to include! and it appears to be under active development.

ClimbsRocks

related to this thread: http://stackoverflow.com/questions/18306416/adaboostclassifier-with-different-base-learners xgboost jumps immediately to mind...

ClimbsRocks

have it check the number of trained classifiers find the best score for each classifier make sure the best scores beat the benchmarks we've already set

ClimbsRocks

right now they're only for regression data

ClimbsRocks

tiny is just logistically- does the process obviously break down anywhere? this lets us quickly check changes for things like syntax errors and such robust trains against the full datasets,...

ClimbsRocks

idea: randomly shuffling the input data

1

slide 18 says vw, libffm and nn are order dependent, so shuffling the data will give you better results: http://www.slideshare.net/odsc/owen-zhangopen-sourcetoolsanddscompetitions1

ClimbsRocks

have it make a rough guess to shoot for 8 hours of training time. but make that training time variable super obvious, so people can tweak it themselves. obviously, this...

ClimbsRocks