Hi Abhishek, I have also tried using WEKA SMO, however it take toooooo long (I waited for more than 6 days ) for training for set of more than million instances. However logistic regression could come out with model in 20 mins.
This is pretty fast! My problem is I can use model as is in classification rather want to understand and use weights generated from training. My question is how to use Coefficients and Odds Ratios that could be used for classification as mentioned in previous post? Hope people dont raise exception to ask hadoop only question :( Thanks, Rajesh On Tue, Oct 16, 2012 at 7:37 PM, Abhishek Shivkumar < abhisheksgum...@gmail.com> wrote: > As far as I know weka cannot be run on hadoop directly. > What can be done is if your algorithm first generats a model based on a > training data initially, then you can run your training offline on your > laptop and serialize, i.e. write the trained model in a file. Now, put this > model file on hdfs and read it inside your setup method of map reduce > programs. > > As and when you read your input in your mapper method, you can take the > trained model file to determine any decision such as a classification or > other supervised machine lerarning algorithm decisions. > > I did this for SVM and it did work. > I am interested to know if anyone else has tried any alternate method to > port weka algorithms on hadoop. > > Thanks! > With Regards, > Abhishek S > > On Oct 16, 2012, at 7:16 PM, Rajesh Nikam <rajeshni...@gmail.com> wrote: > > > Hi, > > > > I was looking for logistic regression algorithms on hadoop. > > mahout is one good package to use on hadoop, however I am not able to > get could results with my experiments. > > > > There are logistic regression algorithms supported with WEKA which I > have used on Windows. > > I guess I should be able to run these algos from JAR files as is on > linux. > > > > java -classpath weka.jar weka.classifiers.functions.Logistic -R 1.0E-8 > -M 6 -t lr.arff > > > > Have anyone ported them to take advantage of hadoop ? > > > > How to interpret the output generated from it like what is Coefficients > and Odds Ratios that could be used for classification ? > > > > > > Options: -R 1.0E-8 -M 6 > > > > Logistic Regression with ridge parameter of 1.0E-8 > > Coefficients... > > Class > > Variable class_1 > > ====================== > > a1 0 > > a2 0 > > a3 0 > > a4 0.0082 > > a5 0.0151 > > a6 -0.1034 > > a7 0 > > a8 0 > > a9 0 > > a10 -0.0397 > > a11 -0.0003 > > a13 -0.1195 > > a14 -0.1389 > > Intercept -21.487 > > > > > > Odds Ratios... > > Class > > Variable class_1 > > ====================== > > a1 1 > > a2 1 > > a3 1 > > a4 1.0083 > > a5 1.0152 > > a6 0.9018 > > a7 1 > > a8 1 > > a9 1 > > a10 0.961 > > a11 0.9997 > > a13 0.8873 > > a14 0.8703 > > > > Time taken to build model: 6.39 seconds > > Time taken to test model on training data: 1.86 seconds > > > > === Error on training data === > > > > Correctly Classified Instances 49528 99.9173 % > > Incorrectly Classified Instances 41 0.0827 % > > Kappa statistic 0.9983 > > Mean absolute error 0.0011 > > Root mean squared error 0.0244 > > Relative absolute error 0.2202 % > > Root relative squared error 4.895 % > > Total Number of Instances 49569 > > > > > > === Confusion Matrix === > > > > a b <-- classified as > > 26526 37 | a = class_1 > > 4 23002 | b = class_2 > > > > > > > > === Stratified cross-validation === > > > > Correctly Classified Instances 49492 99.8447 % > > Incorrectly Classified Instances 77 0.1553 % > > Kappa statistic 0.9969 > > Mean absolute error 0.0015 > > Root mean squared error 0.0358 > > Relative absolute error 0.3108 % > > Root relative squared error 7.1718 % > > Total Number of Instances 49569 > > > > > > === Confusion Matrix === > > > > a b <-- classified as > > 26532 31 | a = class_1 > > 46 22960 | b = class_2 > > > > Thanks in advance. > > Rajesh >
