[MADLIB-604] SVM Regression Performance : Several data sets handling is much slower than libsvm

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Abandoned
• Fix Version/s: v1.9
• Assignee: Rahul Iyer
• Reporter: Jiali Yao
• Created: 30/Jul/12 15:01
• Updated: 22/Mar/18 18:46
• Resolved: 22/Mar/18 18:46

Description

For several data sets, MADlib is slower than libsvm
1. Time differnce

Kernel is dot
Data Sets 	MADlib(Para=true)	MADlib(Para=false)	libsvm	MADlib/libsvm
cadata	        874.15	277.35	2	138.68
etfidf	        501.61	1844.26	32	15.68

kernel is Polymial
cadata	932.13	8979.85	2761	0.34
etfidf	2269.23	3175.87	33	68.76
space	139.12	238.26	1	139.12

kernel is Gaussian
cadata	900.83	9130.2	1	900.83
cpusmall	390.57	196.13	1	196.13

2. Test case example:
SELECT madlib.svm_regression
                        ( 'madlibtestdata.svm_cadata'::text     --input_table
                        , 'madlibtestresult.reg_model_table'::text    --model_table
                        , 'false'::boolean       --parallel
                        , 'madlibtestdata.svm_polynomial'::text    --kernel_func
                        , 'false'::boolean        --verbose
                        , '0.1'::float8            --eta
                        , '0.005'::float8             --nu
                        , '0.05'::float8        --slambda
                   ) AS q;

3. Data sets
4. Parameter seting
MADlib parameter: default value
R parameter:

svm-train  -s 4 -t 0 -c $cost -n 0.005 
eunite2001 0.0595
E2006 0.0012