Fabien Torre's site, Lille university, France


JStumps: boosting of stumps in Java

JStumps implements in Java the boosting of stumps. Many options are available and described below.

Download the JStump archive here.

Examples

JStumps program

Learning in verbose mode

# java JStumps --file=datasets/breast-cancer/CrossValidation/shot_1 --boost=20
  . parameters used:
    help        = false
    verb        = true
    save        = false
    test        = false
    isoclass    = false
    file        = /home/torre/datasets/breast-cancer/shot_1
    insist      = 10
    onthefly    = false
    boost       = 20
    showst      = false
  . reading names file: /home/torre/datasets/breast-cancer/shot_1.names
  . reading data file:  /home/torre/datasets/breast-cancer/shot_1.data
  . 20 steps of boosting now
  . 18 stumps found
  . error rate on data file: 3,97 %

Learn and test, not verbose

# java JStumps --file=datasets/breast-cancer/CrossValidation/shot_1 --boost=20 --noverb --test
  . error rate on test file: 4,35 %

Learn and save classifier

# java JStumps --file=datasets/breast-cancer/CrossValidation/shot_1 --boost=20 --noverb --save

JSConsult program

Consult classifier

# java JSConsult --classifier=datasets/breast-cancer/CrossValidation/shot_1.JStumps --noverb --showst
if UniformityCellSize<=3.0 then class 2 else class 4 [weight=0.17779300366525003]
if ClumpThickness<=6.0 then class 2 else class 4 [weight=0.13548074159737228]
if BareNuclei<=1.0 then class 2 else class 4 [weight=0.10766897146164137]
if UniformityCellSize<=3.0 then class 4 else class 2 [weight=0.09351568652457329]
if BlandChromatin<=2.0 then class 2 else class 4 [weight=0.06902295191295167]
if NormalNucleoli<=3.0 then class 2 else class 4 [weight=0.04925353047208494]
if BlandChromatin<=4.0 then class 2 else class 4 [weight=0.047019527580647674]
if MarginalAdhesion<=1.0 then class 2 else class 4 [weight=0.0361683906745061]
if Mitoses<=1.0 then class 2 else class 4 [weight=0.034611098925205766]
if UniformityCellShape<=2.0 then class 2 else class 4 [weight=0.03309250317288488]
if ClumpThickness<=2.0 then class 2 else class 4 [weight=0.03203251573592544]
if ClumpThickness<=4.0 then class 2 else class 4 [weight=0.03196929631350762]
if NormalNucleoli<=1.0 then class 4 else class 2 [weight=0.029934655990969988]
if ClumpThickness<=5.0 then class 4 else class 2 [weight=0.028165092442751286]
if BareNuclei<=3.0 then class 2 else class 4 [weight=0.02730910998261713]
if SingleEpithelialCellSize<=1.0 then class 4 else class 2 [weight=0.023181867572012904]
if SingleEpithelialCellSize<=2.0 then class 2 else class 4 [weight=0.02212647398009039]
if MarginalAdhesion<=9.0 then class 2 else class 4 [weight=0.021654581995007294]

Classify test examples

# java JSConsult --classifier=datasets/breast-cancer/CrossValidation/shot_1.JStumps
                 --file=datasets/breast-cancer/CrossValidation/shot_1 --noverb --predictions
test1,4,0.41566580533981323,4
test2,4,0.5130572319030762,4
test3,4,0.09624868631362915,4
test4,4,0.3285263776779175,4
test5,2,0.24965348839759827,4
test6,4,0.3390907943248749,4
test7,4,0.1266273856163025,4
test8,4,0.12250885367393494,4
test9,4,0.607096254825592,4
test10,4,0.44383499026298523,4
test11,4,0.5130572319030762,4
test12,4,0.3924649953842163,4
test13,4,0.6504054665565491,4
test14,4,0.5563663840293884,4
test15,4,0.2863062620162964,4
.............................

Output confusion matrix

# java JSConsult --classifier=datasets/breast-cancer/CrossValidation/shot_1.JStumps
                 --file=datasets/breast-cancer/CrossValidation/shot_1 --noverb --confusion
;2,2,43;2,4,1;4,2,2;4,4,23

More informations

  • see the Javadoc
  • use java JStumps --help
  • use java JSConsult --help
  • works for learning problems with two classes to predict only
  • missing values are authorized in both training set and test set
  • JStumps uses C4.5 formats (.names, .data and .test)
  • requires Java 6 (downloads the latest JDK)

Bibliography

Definition of stumps (also called AttrTest, section 3.1):

Yoav Freund and Robert E. Schapire
Experiments with a New Boosting Algorithm
1996
CiteSeer page ]

Boosting algorithm used in this implementation (Adaboost, figure 1):

Yoav Freund and Robert E. Schapire
A short introduction to boosting
IJCAI 1999
CiteSeer page ]



Fabien Torre Valid HTML5! Valid CSS!
Accueil > Research > Softs > JStumps
(last update )
Fabien Torre's site, Lille university, France

Description

Survoler un lien de navigation pour lire sa description ici...