readme

来自「LIBSVM源码。LIBSVM 是台湾大学林智仁(Chih-Jen Lin)博士」· 代码 · 共 52 行

This code uses dense vectors to store data instances. If most feature
values are non-zeros, the training/testing time is faster than the
standard libsvm, which implement sparse vectors.


Experimental Setting:

We select four concepts, animal, people, sky, and weather, from
MediaMill Challenge Problem for this experiment. All instances are
120-dimension dense vectors. We compare the standard libsvm and this
modification.

The experimental procedure performs parameter selection (30 parameter
sets), model training (best parameters), test data predictions. There
are 30,993 and 12,914 records in training (used in parameter selection
and model training stages) and testing (used in prediction stage)
sets, respectively. The following table shows the results.

(1) parameter-selection execution time (sec.):
          original    dense        change 
          libsvm      repr.        ratio
animal     2483.23     1483.02     -40.3%
people    22844.26    13893.21     -39.2%
sky       13765.39     8460.06     -38.5%
weather    2240.01     1325.32     -40.1%
AVERAGE   10083.17     6540.46     -39.1%

(2) model-training execution time (sec.):
          original    dense        change 
          libsvm      repr.        ratio
animal     113.238      70.085     -38.1% 
people     725.995     451.244     -37.8%
sky       1234.881     784.071     -36.5%
weather    123.179      76.532     -37.9%
AVERAGE    549.323     345.483     -37.1%

(3) prediction execution time (sec.):
          original    dense       change
          libsvm      repr..      ratio
animal      12.226        6.895    -43.6%
people      99.268       54.855    -44.7%
sky         78.069       42.417    -45.7%
weather     10.843        6.495    -44.9%
AVERAGE     50.102       27.666    -44.8% 

Overall, dense-representation libsvm saves on average 39.1%, 37.1%,
and 44.8% of the execution time in parameter selection, model training
and prediction stages, respectively. This modification is thus useful
for dense data sets.