Mercurial > hg > camir-aes2014

annotate toolboxes/FullBNT-1.0.7/bnt/examples/static/dtree/test_restaurants.m @ 0:e9a9cd732c1e tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 % Here the training data is adapted from Russell95 book. See restaurant.names for description.
wolffd@0 2 % (1) Use infomation-gain as the split testing score, we get the the same decision tree as the book Russell 95 (page 537),
wolffd@0 3 % and the Gain(Patrons) is 0.5409, equal to the result in Page 541 of Russell 95. (see below output trace)
wolffd@0 4 % (Note: the dtree in that book has small compilation error, the Type node is from YES of Hungry node, not NO.)
wolffd@0 5 % (2) Use gain-ratio (Quilan 93), the splitting defavorite attribute with more values. (e.g. the Type attribute here)
wolffd@0 6
wolffd@0 7 dtreeCPD=tree_CPD;
wolffd@0 8
wolffd@0 9 % load data
wolffd@0 10 fname = fullfile(BNT_HOME, 'examples', 'static', 'uci_data', 'restaurant', 'restaurant.data');
wolffd@0 11 data=load(fname);
wolffd@0 12 data=data';
wolffd@0 13
wolffd@0 14 %make the data be BNT compliant (values for discrete nodes are from 1-n, here n is the node size)
wolffd@0 15 % e.g. if the values are [0 1 6], they must be mapping to [1 2 3]
wolffd@0 16 %data=transform_data(data,'tmp.dat',[]); %here no cts nodes
wolffd@0 17
wolffd@0 18 % learn decision tree from data
wolffd@0 19 ns=2*ones(1,11);
wolffd@0 20 ns(5:6)=3;
wolffd@0 21 ns(9:10)=4;
wolffd@0 22 dtreeCPD1=learn_params(dtreeCPD,1:11,data,ns,[]);
wolffd@0 23
wolffd@0 24 % evaluate on data
wolffd@0 25 [score,outputs]=evaluate_tree_performance(dtreeCPD1,1:11,data,ns,[]);
wolffd@0 26 fprintf('Accuracy in training data %6.3f\n',score);
wolffd@0 27
wolffd@0 28 % show decision tree using graphpad
wolffd@0 29
wolffd@0 30
wolffd@0 31
wolffd@0 32 % --------------------------Output trace: using Information-Gain------------------------------
wolffd@0 33 % The splits are Patron, Hungry, Type, Fri/Sat
wolffd@0 34 % *********************************
wolffd@0 35 % Create node 1 split at 5 gain 0.5409 Th 0. Class 1 Cases 12 Error 6
wolffd@0 36 % Create leaf node(onecla) 2. Class 1 Cases 2 Error 0
wolffd@0 37 % Add subtree node 2 to 1. #nodes 2
wolffd@0 38 % Create leaf node(onecla) 3. Class 2 Cases 4 Error 0
wolffd@0 39 % Add subtree node 3 to 1. #nodes 3
wolffd@0 40 % Create node 4 split at 4 gain 0.2516 Th 0. Class 1 Cases 6 Error 2
wolffd@0 41 % Create leaf node(onecla) 5. Class 1 Cases 2 Error 0
wolffd@0 42 % Add subtree node 5 to 4. #nodes 5
wolffd@0 43 % Create node 6 split at 9 gain 0.5000 Th 0. Class 1 Cases 4 Error 2
wolffd@0 44 % Create leaf node(nullset) 7. Father 6 Class 1
wolffd@0 45 % Create node 8 split at 3 gain 1.0000 Th 0. Class 1 Cases 2 Error 1
wolffd@0 46 % Create leaf node(onecla) 9. Class 1 Cases 1 Error 0
wolffd@0 47 % Add subtree node 9 to 8. #nodes 9
wolffd@0 48 % Create leaf node(onecla) 10. Class 2 Cases 1 Error 0
wolffd@0 49 % Add subtree node 10 to 8. #nodes 10
wolffd@0 50 % Add subtree node 8 to 6. #nodes 10
wolffd@0 51 % Create leaf node(onecla) 11. Class 2 Cases 1 Error 0
wolffd@0 52 % Add subtree node 11 to 6. #nodes 11
wolffd@0 53 % Create leaf node(onecla) 12. Class 1 Cases 1 Error 0
wolffd@0 54 % Add subtree node 12 to 6. #nodes 12
wolffd@0 55 % Add subtree node 6 to 4. #nodes 12
wolffd@0 56 % Add subtree node 4 to 1. #nodes 12
wolffd@0 57 % ********************************
wolffd@0 58 %
wolffd@0 59 % Note:
wolffd@0 60 % ***Create node 4 split at 4 gain 0.2516 Th 0. Class 1 Cases 6 Error 2
wolffd@0 61 % This mean we create a new node number 4, it is splitting at the attribute 4, and info-gain is 0.2516,
wolffd@0 62 % "Th 0" means threshhold for splitting continous attribute, "Class 1" means the majority class at node 4 is 1,
wolffd@0 63 % and "Cases 6" means it has 6 cases attached to it, "Error 2" means it has two errors if changing the class lable of
wolffd@0 64 % all the cases in it to the majority class.
wolffd@0 65 % *** Add subtree node 12 to 6. #nodes 12
wolffd@0 66 % It means we add the child node 12 to node 6.
wolffd@0 67 % *** Create leaf node(onecla) 10. Class 2 Cases 1 Error 0
wolffd@0 68 % here 'onecla' means all cases in this node belong to one class, so no need to split further.
wolffd@0 69 % 'nullset' means no training cases belong to this node, we use its parent node majority class as its class
wolffd@0 70 %
wolffd@0 71 %
wolffd@0 72 %
wolffd@0 73 % ---------------Output trace: using GainRatio-----------------------
wolffd@0 74 % The splits are Patron, Hungry, Fri/Sat, Price
wolffd@0 75 %
wolffd@0 76 %
wolffd@0 77 % Create node 1 split at 5 gain 0.3707 Th 0. Class 1 Cases 12 Error 6
wolffd@0 78 % Create leaf node(onecla) 2. Class 1 Cases 2 Error 0
wolffd@0 79 % Add subtree node 2 to 1. #nodes 2
wolffd@0 80 % Create leaf node(onecla) 3. Class 2 Cases 4 Error 0
wolffd@0 81 % Add subtree node 3 to 1. #nodes 3
wolffd@0 82 % Create node 4 split at 4 gain 0.2740 Th 0. Class 1 Cases 6 Error 2
wolffd@0 83 % Create leaf node(onecla) 5. Class 1 Cases 2 Error 0
wolffd@0 84 % Add subtree node 5 to 4. #nodes 5
wolffd@0 85 % Create node 6 split at 3 gain 0.3837 Th 0. Class 1 Cases 4 Error 2
wolffd@0 86 % Create leaf node(onecla) 7. Class 1 Cases 1 Error 0
wolffd@0 87 % Add subtree node 7 to 6. #nodes 7
wolffd@0 88 % Create node 8 split at 6 gain 1.0000 Th 0. Class 2 Cases 3 Error 1
wolffd@0 89 % Create leaf node(onecla) 9. Class 2 Cases 2 Error 0
wolffd@0 90 % Add subtree node 9 to 8. #nodes 9
wolffd@0 91 % Create leaf node(nullset) 10. Father 8 Class 2
wolffd@0 92 % Create leaf node(onecla) 11. Class 1 Cases 1 Error 0
wolffd@0 93 % Add subtree node 11 to 8. #nodes 11
wolffd@0 94 % Add subtree node 8 to 6. #nodes 11
wolffd@0 95 % Add subtree node 6 to 4. #nodes 11
wolffd@0 96 % Add subtree node 4 to 1. #nodes 11
wolffd@0 97 %
wolffd@0 98 %