Additional classification problem instances

[sforsten]

Adding additional classification problems to problem instances from ​http://archive.ics.uci.edu/ml/datasets.html

[sforsten]

r8595:

• renamed real world classification to UCI
• added wine classification problem
• deleted Iris class, which has been forgotten in r8254

[sforsten]

r8596: adjusted unit tests

[abeham]

I just received this error when building HL as Release and x86 using Build.cmd:

"c:\Windows\Microsoft.NET\Framework64\v4.0.30319\Microsoft.Common.targets(483,9): error : The OutputPath property is not set for project 'HeuristicLab.Problems.Instances.DataAnalysis.Views-3.3.csproj'. Please check to make sure that you have specified a valid combination of Configuration and Platform for this project. Configuration='Release' Platform='x86'. You may be seeing this message because you are trying to build a project without a solution file, and have specified a non-default Configuration or Platform that doesn't exist for this project. [D:\HL3\trunk\sources\HeuristicLab.Problems.Instances.DataAnalysis.Views\3.3\HeuristicLab.Problems.Instances.DataAnalysis.Views-3.3.csproj]"

[abeham]

r8781: added missing platforms in Problems.Instances.DataAnalysis.Views

[sforsten]

r8841: added Iris and Thyroid problems from UCI

[abeham]

I see a problem with Thyroid in that there are many versions and different databases. I think to avoid confusion, we should distinguish it by name. I would suggest something like "Thyroid, S. Aeberhard, 1992" (name, donor, year received), as this seems to be the version that you have added. Should be done for the other UCI datasets also.

[sforsten]

r8889: IUCIDataDescriptor to add more information about the datasets to the data, to distinguish between different version, as suggested by abeham

[abeham]

I looked through some other available datasets and found two that could be interesting to add as benchmarking:

• ​Vertebral Column (the 3-class dataset) has 310 instances and 6 attributes
• ​Parkinson (remove the first attribute (it's a string) and move the class attribute to the end) - seems to be suited for k-NN (after a quick analysis with weka), has few instances and a skewed class distribution

Please shuffle both datasets randomly before adding them.

[sforsten]

r8902: fixed CreateGpSymbolicClassificationSampleTest and therefore also RunGpSymbolicClassificationSampleTest

[sforsten]

r8903:

• added two additional classification instances as suggested by abeham
• fixed a bug in UCIInstanceProvider

[abeham]

r8908: Abbreviated first names as in the other datasets

[sforsten]

r9208:

• added wisconsin breast cancer problem instance
• corrected iris dataset
• changed classification data descriptors to be able to set training and test partition as well as input and target variables (in the same way as it is done in regression)

[gkronber]

Replying to abeham:

I looked through some other available datasets and found two that could be interesting to add as benchmarking:

• ​Vertebral Column (the 3-class dataset) has 310 instances and 6 attributes
• ​Parkinson (remove the first attribute (it's a string) and move the class attribute to the end) - seems to be suited for k-NN (after a quick analysis with weka), has little instances and a skewed class distribution

Please shuffle both datasets randomly before adding them.

I just checked the 'Parkinsons' data set and strongly believe that instances should not be shuffled. The data set contains multiple measurements for each patient. So, if the data set is shuffled it should be guaranteed that all measurements from a patient are either in the training partition or in the test partition only. I also verified this in HL using the original and the shuffled data set (leads to much better estimated values).

[gkronber]

r9453: removed shuffling for the Parkinson's data set from UCI.

[abeham]

You're right, thx!