timbl(1) Tilburg Memory Based Learner

SYNOPSYS

timbl [options]

timbl -f data-file -t test-file

DESCRIPTION

TiMBL is an open source software package implementing several memory-based learning algorithms, among which IB1-IG, an implementation of k-nearest neighbor classification with feature weighting suitable for symbolic feature spaces, and IGTree, a decision-tree approximation of IB1-IG. All implemented algorithms have in common that they store some representation of the training set explicitly in memory. During testing, new cases are classified by extrapolation from the most similar stored cases.

OPTIONS

-a <n> or -a <string>
determines the classification algorithm.

Possible values are:

0 or IB
 the IB1 (k-NN) algorithm (default)

1 or IGTREE
 a decision-tree-based approximation of IB1

2 or TRIBL
 a hybrid of IB1 and IGTREE

3 or IB2
 an incremental editing version of IB1

4 or TRIBL2
 a non-parameteric version of TRIBL

-b n

number of lines used for bootstrapping (IB2 only)

-B n

number of bins used for discretization of numeric feature values

--Beam=<n>

limit +v db output to n highest-vote classes

--clones=<n>

number f threads to use for parallel testing

-c n

clipping frequency for prestoring MVDM matrices

+D

store distributions on all nodes (necessary for using +v db with IGTree, but wastes memory otherwise)

--Diversify

rescale weight (see docs)

-d val

weigh neighbors as function of their distance:
 Z      : equal weights to all (default)
 ID     : Inverse Distance
 IL     : Inverse Linear
 ED:a   : Exponential Decay with factor a (no whitespace!)
 ED:a:b : Exponential Decay with factor a and b (no whitespace!)

-e n

estimate time until n patterns tested

-f file

read from data file 'file' OR use filenames from 'file' for cross validation test

-F format

assume the specified input format (Compact, C4.5, ARFF, Columns, Binary, Sparse )

-G normalization

normalize distibutions (+v db option only)

Supported normalizations are:

Probability or 0

normalize between 0 and 1

addFactor:<f> or 1:<f>

add f to all possible targets, then normalize between 0 and 1 (default f=1.0).

logProbability or 2

Add 1 to the target Weight, take the 10Log and then normalize between 0 and 1

+H or -H

write hashed trees (default +H)

-i file

read the InstanceBase from 'file' (skips phase 1 & 2 )

-I file

dump the InstanceBase in 'file'

-k n

search 'n' nearest neighbors (default n = 1)

-L n

set value frequency threshold to back off from MVDM to Overlap at level n

-l n

fixed feature value length (Compact format only)

-m string

use feature metrics as specified in' string':
 The format is : GlobalMetric:MetricRange:MetricRange
           e.g.: mO:N3:I2,5-7


 C: cosine distance. (Global only. numeric features implied)
 D: dot product. (Global only. numeric features implied)
 DC: Dice coefficient
 O: weighted overlap (default)
 E: Euclidian distance
 L: Levenshtein distance
 M: modified value difference
 J: Jeffrey divergence
 S: Jensen-Shannon divergence
 N: numeric values
 I: Ignore named  values

--matrixin=file

read ValueDifference Matrices from file 'file'

--matrixout=file

store ValueDifference Matrices in 'file'

-n file

create a C4.5-style names file 'file'

-M n

size of MaxBests Array

-N n

number of features (default 2500)

-o s

use s as output filename

--occurences=<value>

The input file contains occurrence counts (at the last position) value can be one of: train , test or both

-O path

save output using 'path'

-p n

show progress every n lines (default p = 100,000)

-P path

read data using 'path'

-q n

set TRIBL threshold at level n

-R n

solve ties at random with seed n

-s

use the exemplar weights from the input file

-s0

ignore the exemplar weights from the input file

-T n

use feature n as the class label. (default: the last feature)

-t file

test using 'file'

-t leave_one_out

test with the leave-one-out testing regimen (IB1 only). you may add --sloppy to speed up leave-one-out testing (but see docs)

-t cross_validate

perform cross-validation test (IB1 only)

-t @file

test using files and options described in 'file' Supported options: d e F k m o p q R t u v w x % -

--Treeorder =value n

ordering of the Tree:
 DO: none
 GRO: using GainRatio
 IGO: using InformationGain
 1/V: using 1/# of Values
 G/V: using GainRatio/# of Valuess
 I/V: using InfoGain/# of Valuess
 X2O: using X-square
 X/V: using X-square/# of Values
 SVO: using Shared Variance
 S/V: using Shared Variance/# of Values
 GxE: using GainRatio * SplitInfo
 IxE: using InformationGain * SplitInfo
 1/S: using 1/SplitInfo

-u file

read value-class probabilities from 'file'

-U file

save value-class probabilities in 'file'

-V

Show VERSION

+v level or -v level

set or unset verbosity level, where level is:


 s:  work silently
 o:  show all options set
 b:  show node/branch count and branching factor
 f:  show calculated feature weights (default)
 p:  show value difference matrices
 e:  show exact matches
 as: show advanced statistics (memory consuming)
 cm: show confusion matrix (implies +vas)
 cs: show per-class statistics (implies +vas)
 cf: add confidence to output file (needs -G)
 di: add distance to output file
 db: add distribution of best matched to output file
 md: add matching depth to output file.
 k:  add a summary for all k neigbors to output file (sets -x)
 n:  add nearest neigbors to output file (sets -x)


  You may combine levels using '+' e.g. +v p+db or -v o+di

-w n

weighting
 0 or nw: no weighting
 1 or gr: weigh using gain ratio (default)
 2 or ig: weigh using information gain
 3 or x2: weigh using the chi-square statistic
 4 or sv: weigh using the shared variance statistic
 5 or sd: weigh using standard deviation. (all features must be numeric)

-w file

read weights from 'file'

-w file:n

read weight n from 'file'

-W file

calculate and save all weights in 'file'

+% or -%

do or don't save test result (%) to file

+x or -x

do or don't use the exact match shortcut
   (IB1 and IB2 only, default is -x)

-X file

dump the InstanceBase as XML in 'file'

BUGS

possibly

AUTHORS

Ko van der Sloot [email protected]

Antal van den Bosch [email protected]