liblinear gives very different regression results compared with libsvm

[CooledCoffee]

Procedures

1. Using the simple regression dataset provided by libsvm http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/regression.html#eunite2001.
2. Train & predict using both libsvm & liblinear

svm-train -s 3 -t 0 -c 1 -p 0.1 -e 0.001 -h 0 eunite2001 model.1 && svm-predict eunite2001.t model.1 prediction.1

liblinear-train -s 11 -c 1 -p 0.1 -e 0.001 eunite2001 model.11 && liblinear-predict eunite2001.t model.11 prediction.11

liblinear-train -s 12 -c 1 -p 0.1 -e 0.001 eunite2001 model.12 && liblinear-predict eunite2001.t model.12 prediction.12

liblinear-train -s 13 -c 1 -p 0.1 -e 0.001 eunite2001 model.13 && liblinear-predict eunite2001.t model.13 prediction.13 3. The results are here:

libsvm	liblinear -s 11	liblinear -s 12	liblinear -s 13
754.219	711.818	714.293	655.209
735.951	695.675	703.196	651.262
745.716	606.048	601.496	628.192
756.885	721.134	721.481	652.914
758.048	704.657	705.966	644.363
758.296	703.099	703.878	644.147
756.88	680.706	688.226	629.164
753.174	681.003	682.531	631.114
733.147	666.063	668.37	617.042
743.909	606.234	599.665	605.601
...	...	...	...

Questions

1. To my understanding, "liblinear-train -s 13" is the best match for "svm-train -s 3 -t 0". Is that correct?
2. Why are the results so different? In general, which tool gives better result?

[cjlin1]

-s 11 and -s 12 should give you same (or almost the same) results if you use a smaller -e. Regarding libsvm and liblinear their formulations are slightly different. In liblinear by default there is no bias term. Also for libsvm you want to make sure you are using linear kernel for this comparison CooledCoffee writes:

Procedures

1. Using the simple regression dataset provided by libsvm http:// www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/regression.html# eunite2001.
2. Train & predict using both libsvm & liblinear

svm-train -s 3 -t 0 -c 1 -p 0.1 -e 0.001 -h 0 eunite2001 model.1 && svm-predict eunite2001.t model.1 prediction.1

liblinear-train -s 11 -c 1 -p 0.1 -e 0.001 eunite2001 model.11 && liblinear-predict eunite2001.t model.11 prediction.11

liblinear-train -s 12 -c 1 -p 0.1 -e 0.001 eunite2001 model.12 && liblinear-predict eunite2001.t model.12 prediction.12

liblinear-train -s 13 -c 1 -p 0.1 -e 0.001 eunite2001 model.13 && liblinear-predict eunite2001.t model.13 prediction.13 3. The results are here:

libsvm liblinear -s 11 liblinear -s 12 liblinear -s 13 754.219 711.818 714.293 655.209
735.951 695.675 703.196 651.262
745.716 606.048 601.496 628.192
756.885 721.134 721.481 652.914
758.048 704.657 705.966 644.363
758.296 703.099 703.878 644.147
756.88 680.706 688.226 629.164
753.174 681.003 682.531 631.114
733.147 666.063 668.37 617.042
743.909 606.234 599.665 605.601
... ... ... ...

Questions

1. To my understanding, "liblinear-train -s 13" is the best match for "svm-train -s 3 -t 0". Is that correct?
2. Why are the results so different? In general, which tool gives better result?

— Reply to this email directly or view it on GitHub.*

[simsong]

This issue was moved to angleto/liblinear#20