Hello,
I'm having a problem with XVPrediction and the LinearRegression operator. What I want to achieve is to learn a linear regression model and then apply the model to some test data. Initially, I use cross-validation to evaluate my approach and as I am interested in the actual predictions I am using XVPrediction rather than XValidation. While I can see a new attribute "Prediction(...)" in the resulting data set, all values in this column are "unknown".
My process looks like this:
<?xml version="1.0" encoding="MacRoman"?>
<process version="4.3">
<operator name="Root" class="Process" expanded="yes">
<operator name="CSVExampleSource" class="CSVExampleSource">
<parameter key="filename" value="/Users/bfranke/HPCA/rapidminer/ARM.csv"/>
<parameter key="id_name" value="Filename"/>
<parameter key="label_name" value="CNT_cycles"/>
<parameter key="trim_lines" value="true"/>
</operator>
<operator name="FeatureNameFilter" class="FeatureNameFilter">
<parameter key="skip_features_with_name" value="CNT_alloc_OSMs|CNT_retired_OSMs"/>
</operator>
<operator name="Numerical2Real" class="Numerical2Real">
</operator>
<operator name="XVPrediction" class="XVPrediction" expanded="yes">
<parameter key="leave_one_out" value="true"/>
<parameter key="sampling_type" value="linear sampling"/>
<operator name="LinearRegression" class="LinearRegression">
<parameter key="feature_selection" value="greedy"/>
<parameter key="keep_example_set" value="true"/>
</operator>
<operator name="ModelApplier" class="ModelApplier">
<list key="application_parameters">
</list>
<parameter key="create_view" value="true"/>
<parameter key="keep_model" value="true"/>
</operator>
</operator>
<operator name="ResultWriter" class="ResultWriter">
<parameter key="result_file" value="/Users/bfranke/HPCA/rapidminer/results.res"/>
</operator>
</operator>
</process>
I have a data set with 293 examples and 45 attributes, of which one is the textual/nominal ID and another one the numerical label. All other attributes are also numerical, two of them I filter out using the FeatureNameFilter. I also make sure all regular attributes are converted to real values as some of the attributes are identified as int and others as real. This conversion using the Numerical2Real operator makes sure all attributes are represented using the same real type. Up to this point everything seems to be ok (I set a breakpoint and inspected the data). Hence, the problem seems to be related to XVPrediction and LinearRegression.
Checking the older posts to this forum I haven't found any known issues with either XVPrediction or LinearRegression (the same problem shows up with PolynomialRegression and also GPLearner), so I guess there's something wrong with my process. I've already "experimented" with explicit feature selection to select fewer attributes, but this didn't solve the problem. Any ideas?
Thanks!
Cheers,
Bjoern
