1. Title: meta-data
2. Sources:
(a) Creator:
LIACC - University of Porto
R.Campo Alegre 823
4150 PORTO
(b) Donor: P.B.Brazdil or J.Gama Tel.: +351 600 1672
LIACC, University of Porto Fax.: +351 600 3654
Rua Campo Alegre 823 Email: statlog-adm@ncc.up.pt
4150 Porto, Portugal
(c) Date: March, 1996
(d) Acknowlegements:
LIACC wishes to thank Commission of European Communities
for their support. Also, we wish to thank the following partners
for providing the individual test results:
- Dept. of Statistics, University of Strathclyde, Glasgow, UK
- Dept. of Statistics, University of Leeds, UK
- Aston University, Birmingham, UK
- Forschungszentrum Ulm, Daimler-Benz AG, Germany
- Brainware GmbH, Berlin, Germany
- Frauenhofer Gesellschaft IITB-EPO, Berlin, Germany
- Institut fuer Kybernetik, Bochum, Germany
- ISoft, Gif sur Yvette, France
- Dept. of CS and AI, University of Granada, Spain
3. Past Usage:
Meta-Data was used in order to give advice about which classification
method is appropriate for a particular dataset.
This work is described in:
-"Machine Learning, Neural and Statistical Learning"
Eds. D.Michie,D.J.Spiegelhalter and C.Taylor
Ellis Horwood-1994
- "Characterizing the Applicability of
Classification Algorithms Using Meta-Level Learning",
P. Brazdil, J.Gama and B.Henery:
in Proc. of Machine Learning - ECML-94,
ed. F.Bergadano and L.de Raedt,LNAI Vol.784 Springer-Verlag.
-"Characterization of Classification Algorithms"
J.Gama, P.Brazdil
in Proc. of EPIA 95, LNAI Vol.990
Springer-Verlag, 1995
4. Relevant Information:n
This DataSet is about the results of Statlog project.
The project performed a comparative study between Statistical, Neural
and Symbolic learning algorithms.
Project StatLog (Esprit Project 5170) was concerned with comparative
studies of different machine learning, neural and statistical
classification algorithms. About 20 different algorithms were
evaluated on more than 20 different datasets. The tests carried out
under project produced many interesting results.
Algorithms DataSets
------------------------- --------------------------
C4.5 NewId Credit_Austr Belgian
AC2 CART Chromosome Credit_Man
IndCART Cal5 CUT DNA
CN2 ITRule Diabetes Digits44
Discrim QuaDisc Credit_German Faults
LogDisc ALLOC80 Head Heart
kNN SMART KLDigits Letters
BayesTree CASTLE New_Belgian Sat_Image
DIPLO92 RBF Segment Shuttle
LVQ Backprop Technical TseTse
Kohonen Vehicle
The results of these tests are comprehensively described in a book
(D.Michie et.al, 1994).
5. Number of Instances: 528
6. Number of Attributes: 22 (including an Id#) plus the class attribute
-- all but two attributes are continuously valued
7. Attribute Information:
1. DS_Name categorical Name of DataSet
2. T continuous Number of examples in test set
3. N continuous Number of examples
4. p continuous Number of attributes
5. k continuous Number of classes
6. Bin continuous Number of binary Attributes
7. Cost continuous Cost (1=yes,0=no)
8. SDratio continuous Standard deviation ratio
9. correl continuous Mean correlation between attributes
10. cancor1 continuous First canonical correlation
11. cancor2 continuous Second canonical correlation
12. fract1 continuous First eigenvalue
13. fract2 continuous Second eigenvalue
14. skewness continuous Mean of |E(X-Mean)|^3/STD^3
15. kurtosis continuous Mean of |E(X-Mean)|^4/STD^4
16. Hc continuous Mean entropy of attributes
17. Hx continuous Entropy of classes
18. MCx continuous Mean mutual entropy of class and attributes
19. EnAtr continuous Equivalent number of attributes
20. NSRatio continuous Noise-signal ratio
21. Alg_Name categorical Name of Algorithm
22. Norm_error continuous Normalized Error (continuous class)
8. Missing Attribute Values:
Note that fract2 and cancor2 only apply to datasets with more than
2 classes. When they appear as '?' this means a don't care value.
Summary Statistics:
Attribute Min Max Mean Std
T 270 20000 4569.05 5704.01
N 270 58000 10734.2 14568.8
p 6 180 29.5455 36.8533
k 2 91 9.72727 19.3568
Bin 0 43 3.18182 9.29227
Cost 0 1 0.13636 0.35125
SdRatio 1.0273 4.0014 1.4791 0.65827
Correl 0.0456 0.751 0.23684 0.1861
Cancor1 0.5044 0.9884 0.79484 0.15639
Cancor2 0.1057 0.9623 0.74106 0.269
Fract1 0.1505 1 0.70067 0.3454
Fract2 0.2807 1 0.70004 0.29405
Skewness 0.1802 6.7156 1.78422 1.79022
Kurtosis 0.9866 160.311 22.6672 41.8496
Hc 0.2893 4.8787 1.87158 1.44665
Hx 0.3672 6.5452 3.34502 1.80383
Mcx 0.0187 1.3149 0.31681 0.33548
EnAtr 1.56006 160.644 20.6641 35.6614
NsRatio 1.02314 159.644 28.873 37.925