Data

pendigits

active
ARFF
Publicly available Visibility: public Uploaded 06-04-2014 by Jan van Rijn

0 likes downloaded by 0 people , 0 total downloads 0 issues 0 downvotes

0 likes downloaded by 0 people , 0 total downloads 0 issues 0 downvotes

Issue | #Downvotes for this reason | By |
---|

Loading wiki

Help us complete this description
Edit

Author:
Source: Unknown -
Please cite:
1. Title of Database: Pen-Based Recognition of Handwritten Digits
2. Source:
E. Alpaydin, F. Alimoglu
Department of Computer Engineering
Bogazici University, 80815 Istanbul Turkey
alpaydin@boun.edu.tr
July 1998
3. Past Usage:
F. Alimoglu (1996) Combining Multiple Classifiers for Pen-Based
Handwritten Digit Recognition,
MSc Thesis, Institute of Graduate Studies in Science and
Engineering, Bogazici University.
http://www.cmpe.boun.edu.tr/~alimoglu/alimoglu.ps.gz
F. Alimoglu, E. Alpaydin, "Methods of Combining Multiple Classifiers
Based on Different Representations for Pen-based Handwriting
Recognition," Proceedings of the Fifth Turkish Artificial
Intelligence and Artificial Neural Networks Symposium (TAINN 96),
June 1996, Istanbul, Turkey.
http://www.cmpe.boun.edu.tr/~alimoglu/tainn96.ps.gz
4. Relevant Information:
We create a digit database by collecting 250 samples from 44 writers.
The samples written by 30 writers are used for training,
cross-validation and writer dependent testing, and the digits
written by the other 14 are used for writer independent testing. This
database is also available in the UNIPEN format.
We use a WACOM PL-100V pressure sensitive tablet with an integrated
LCD display and a cordless stylus. The input and display areas are
located in the same place. Attached to the serial port of an Intel
486 based PC, it allows us to collect handwriting samples. The tablet
sends $x$ and $y$ tablet coordinates and pressure level values of the
pen at fixed time intervals (sampling rate) of 100 miliseconds.
These writers are asked to write 250 digits in random order inside
boxes of 500 by 500 tablet pixel resolution. Subject are monitored
only during the first entry screens. Each screen contains five boxes
with the digits to be written displayed above. Subjects are told to
write only inside these boxes. If they make a mistake or are unhappy
with their writing, they are instructed to clear the content of a box
by using an on-screen button. The first ten digits are ignored
because most writers are not familiar with this type of input devices,
but subjects are not aware of this.
In our study, we use only ($x, y$) coordinate information. The stylus
pressure level values are ignored. First we apply normalization to
make our representation invariant to translations and scale
distortions. The raw data that we capture from the tablet consist of
integer values between 0 and 500 (tablet input box resolution). The
new coordinates are such that the coordinate which has the maximum
range varies between 0 and 100. Usually $x$ stays in this range, since
most characters are taller than they are wide.
In order to train and test our classifiers, we need to represent
digits as constant length feature vectors. A commonly used technique
leading to good results is resampling the ( x_t, y_t) points.
Temporal resampling (points regularly spaced in time) or spatial
resampling (points regularly spaced in arc length) can be used here.
Raw point data are already regularly spaced in time but the distance
between them is variable. Previous research showed that spatial
resampling to obtain a constant number of regularly spaced points
on the trajectory yields much better performance, because it provides
a better alignment between points. Our resampling algorithm uses
simple linear interpolation between pairs of points. The resampled
digits are represented as a sequence of T points ( x_t, y_t )_{t=1}^T,
regularly spaced in arc length, as opposed to the input sequence,
which is regularly spaced in time.
So, the input vector size is 2*T, two times the number of points
resampled. We considered spatial resampling to T=8,12,16 points in our
experiments and found that T=8 gave the best trade-off between
accuracy and complexity.
5. Number of Instances
pendigits.tra Training 7494
pendigits.tes Testing 3498
The way we used the dataset was to use first half of training for
actual training, one-fourth for validation and one-fourth
for writer-dependent testing. The test set was used for
writer-independent testing and is the actual quality measure.
6. Number of Attributes
16 input+1 class attribute
7. For Each Attribute:
All input attributes are integers in the range 0..100.
The last attribute is the class code 0..9
8. Missing Attribute Values
None
9. Class Distribution
Class: No of examples in training set
0: 780
1: 779
2: 780
3: 719
4: 780
5: 720
6: 720
7: 778
8: 719
9: 719
Class: No of examples in testing set
0: 363
1: 364
2: 364
3: 336
4: 364
5: 335
6: 336
7: 364
8: 336
9: 336
Accuracy on the testing set with k-nn
using Euclidean distance as the metric
k = 1 : 97.74
k = 2 : 97.37
k = 3 : 97.80
k = 4 : 97.66
k = 5 : 97.60
k = 6 : 97.57
k = 7 : 97.54
k = 8 : 97.54
k = 9 : 97.46
k = 10 : 97.48
k = 11 : 97.34

class (target) | nominal | 10 unique values 0 missing | |

input9 | numeric | 101 unique values 0 missing | |

input16 | numeric | 101 unique values 0 missing | |

input15 | numeric | 101 unique values 0 missing | |

input14 | numeric | 101 unique values 0 missing | |

input13 | numeric | 101 unique values 0 missing | |

input12 | numeric | 101 unique values 0 missing | |

input11 | numeric | 101 unique values 0 missing | |

input10 | numeric | 101 unique values 0 missing | |

input1 | numeric | 101 unique values 0 missing | |

input8 | numeric | 101 unique values 0 missing | |

input7 | numeric | 101 unique values 0 missing | |

input6 | numeric | 101 unique values 0 missing | |

input5 | numeric | 101 unique values 0 missing | |

input4 | numeric | 98 unique values 0 missing | |

input3 | numeric | 101 unique values 0 missing | |

input2 | numeric | 96 unique values 0 missing |

0.98

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.DecisionStump -E "weka.attributeSelection.CfsSubsetEval -P 1 -E 1" -S "weka.attributeSelection.BestFirst -D 1 -N 5" -W

0.05

Error rate achieved by the landmarker weka.classifiers.trees.DecisionStump -E "weka.attributeSelection.CfsSubsetEval -P 1 -E 1" -S "weka.attributeSelection.BestFirst -D 1 -N 5" -W

0.94

Kappa coefficient achieved by the landmarker weka.classifiers.trees.DecisionStump -E "weka.attributeSelection.CfsSubsetEval -P 1 -E 1" -S "weka.attributeSelection.BestFirst -D 1 -N 5" -W

0.98

Area Under the ROC Curve achieved by the landmarker weka.classifiers.bayes.NaiveBayes -E "weka.attributeSelection.CfsSubsetEval -P 1 -E 1" -S "weka.attributeSelection.BestFirst -D 1 -N 5" -W

0.05

Error rate achieved by the landmarker weka.classifiers.bayes.NaiveBayes -E "weka.attributeSelection.CfsSubsetEval -P 1 -E 1" -S "weka.attributeSelection.BestFirst -D 1 -N 5" -W

0.94

Kappa coefficient achieved by the landmarker weka.classifiers.bayes.NaiveBayes -E "weka.attributeSelection.CfsSubsetEval -P 1 -E 1" -S "weka.attributeSelection.BestFirst -D 1 -N 5" -W

0.98

Area Under the ROC Curve achieved by the landmarker weka.classifiers.lazy.IBk -E "weka.attributeSelection.CfsSubsetEval -P 1 -E 1" -S "weka.attributeSelection.BestFirst -D 1 -N 5" -W

0.05

Error rate achieved by the landmarker weka.classifiers.lazy.IBk -E "weka.attributeSelection.CfsSubsetEval -P 1 -E 1" -S "weka.attributeSelection.BestFirst -D 1 -N 5" -W

0.94

Kappa coefficient achieved by the landmarker weka.classifiers.lazy.IBk -E "weka.attributeSelection.CfsSubsetEval -P 1 -E 1" -S "weka.attributeSelection.BestFirst -D 1 -N 5" -W

0.72

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.DecisionStump

0.11

Kappa coefficient achieved by the landmarker weka.classifiers.trees.DecisionStump

Number of attributes needed to optimally describe the class (under the assumption of independence among attributes). Equals ClassEntropy divided by MeanMutualInformation.

0.98

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.J48 -C .00001

0.95

Kappa coefficient achieved by the landmarker weka.classifiers.trees.J48 -C .00001

0.98

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.J48 -C .0001

0.95

Kappa coefficient achieved by the landmarker weka.classifiers.trees.J48 -C .0001

0.98

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.J48 -C .001

Maximum mutual information between the nominal attributes and the target attribute.

10

The maximum number of distinct values among attributes of the nominal type.

Average mutual information between the nominal attributes and the target attribute.

An estimate of the amount of irrelevant information in the attributes regarding the class. Equals (MeanAttributeEntropy - MeanMutualInformation) divided by MeanMutualInformation.

10

Average number of distinct values among the attributes of the nominal type.

Minimal mutual information between the nominal attributes and the target attribute.

10

The minimal number of distinct values among attributes of the nominal type.

0.98

Area Under the ROC Curve achieved by the landmarker weka.classifiers.bayes.NaiveBayes

-1.2

First quartile of kurtosis among attributes of the numeric type.

First quartile of mutual information between the nominal attributes and the target attribute.

-0.43

First quartile of skewness among attributes of the numeric type.

26.51

First quartile of standard deviation of attributes of the numeric type.

-0.78

Second quartile (Median) of kurtosis among attributes of the numeric type.

48.53

Second quartile (Median) of means among attributes of the numeric type.

Second quartile (Median) of mutual information between the nominal attributes and the target attribute.

0.09

Second quartile (Median) of skewness among attributes of the numeric type.

30.24

Second quartile (Median) of standard deviation of attributes of the numeric type.

-0.26

Third quartile of kurtosis among attributes of the numeric type.

Third quartile of mutual information between the nominal attributes and the target attribute.

0.45

Third quartile of skewness among attributes of the numeric type.

34.23

Third quartile of standard deviation of attributes of the numeric type.

0.99

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.REPTree -L 1

0.92

Kappa coefficient achieved by the landmarker weka.classifiers.trees.REPTree -L 1

0.99

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.REPTree -L 2

0.92

Kappa coefficient achieved by the landmarker weka.classifiers.trees.REPTree -L 2

0.99

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.REPTree -L 3

0.92

Kappa coefficient achieved by the landmarker weka.classifiers.trees.REPTree -L 3

0.97

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.RandomTree -depth 1

0.05

Error rate achieved by the landmarker weka.classifiers.trees.RandomTree -depth 1

0.94

Kappa coefficient achieved by the landmarker weka.classifiers.trees.RandomTree -depth 1

0.97

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.RandomTree -depth 2

0.05

Error rate achieved by the landmarker weka.classifiers.trees.RandomTree -depth 2

0.94

Kappa coefficient achieved by the landmarker weka.classifiers.trees.RandomTree -depth 2

0.97

Area Under the ROC Curve achieved by the landmarker weka.classifiers.trees.RandomTree -depth 3

0.05

Error rate achieved by the landmarker weka.classifiers.trees.RandomTree -depth 3

0.94

Kappa coefficient achieved by the landmarker weka.classifiers.trees.RandomTree -depth 3

0

Standard deviation of the number of distinct values among attributes of the nominal type.