[CONF] Apache Mahout > Breiman Example

[co...@apache.org]

Space: Apache Mahout (https://cwiki.apache.org/confluence/display/MAHOUT)
Page: Breiman Example (https://cwiki.apache.org/confluence/display/MAHOUT/Breiman+Example)


Edited by Frank Scholten:
---------------------------------------------------------------------
h1. Introduction

This quick start page shows how to run the Breiman example. It implements the test procedure described in Breiman's paper [1]. 
The basic algorithm is as follows :
* repeat I iterations
* foreach iteration do
 ** 10% of the dataset is kept apart as a testing set 
 ** build two forests using the training set, one with m=int(log2(M)+1) (called Random-Input) and one with m=1 (called Single-Input)
 ** choose the forest that gave the lowest oob error estimation to compute the test set error
 ** compute the test set error using the Single Input Forest (test error), this demonstrates that even with m=1, Decision Forests give comparable results to greater values of m
 ** compute the mean test set error using every tree of the chosen forest (tree error). This should indicate how well a single Decision Tree performs
* compute the mean test error for all iterations
* compute the mean tree error for all iterations

h1. Steps
h2. Download the data
* The current implementation is compatible with the UCI repository file format. Here are links to some of the datasets used in Breiman's paper:
 ** glass : http://archive.ics.uci.edu/ml/datasets/Glass+Identification
 ** breast cancer : http://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+(Diagnostic)
 ** diabetes : http://archive.ics.uci.edu/ml/datasets/Pima+Indians+Diabetes
 ** sonar : http://archive.ics.uci.edu/ml/datasets/Connectionist+Bench+(Sonar,+Mines+vs.+Rocks)
 ** ionosphere : http://archive.ics.uci.edu/ml/datasets/Ionosphere
 ** vehicle : http://archive.ics.uci.edu/ml/datasets/Statlog+(Vehicle+Silhouettes) 
 ** german : http://archive.ics.uci.edu/ml/datasets/Statlog+\(German+Credit+Data\)
* Put the data in HDFS: {code}$HADOOP_HOME/bin/hadoop fs -put <PATH TO DATA> testdata{code}

h2. Build the Job files
* In $MAHOUT_HOME/ run: {code}mvn install -DskipTests{code}

h2. Generate a file descriptor for the dataset: 
for the glass dataset (glass.data), run :
{code}
$HADOOP_HOME/bin/hadoop jar $MAHOUT_HOME/core/target/mahout-core-<VERSION>-job.jar org.apache.mahout.df.tools.Describe -p testdata/glass.data -f testdata/glass.info -d I 9 N L
{code}
The "I 9 N L" string indicates the nature of the variables. which means 1 ignored(I) attribute, followed by 9 numerical(N) attributes, followed by the label(L)
* you can also use C for categorical (nominal) attributes

h2. Run the example
{code}
$HADOOP_HOME/hadoop jar $MAHOUT_HOME/examples/target/mahout-examples-<VERSION>-job.jar org.apache.mahout.df.BreimanExample -d testdata/glass.data -ds testdata/glass.info -i 10 -t 100
{code}
which builds 100 trees (-t argument) and repeats the test 10 iterations (-i argument) 
* The example outputs the following results:
** Selection error : mean test error for the selected forest on all iterations
** Single Input error : mean test error for the single input forest on all iterations
** One Tree error : mean single tree error on all iterations
** Mean Random Input Time : mean build time for random input forests on all iterations
** Mean Single Input Time : mean build time for single input forests on all iterations


Change your notification preferences: https://cwiki.apache.org/confluence/users/viewnotifications.action