golearn

Table of Contents

Histogram binning
Chi-Merge binning
Support status

Golearn supports discretisation of FloatAttributes to CategoricalAttributes through fixed-width histogram binning and variable-width Chi-Merge binning.

Histogram binning

Code excerpt: Histogram binning

inst, _ := base.ParseCSVToInstances("../examples/datasets/iris_headers.csv", true)
binAttr := inst.AllAttributes()[0]
filt := NewBinningFilter(inst, 10)
filt.AddAttribute(binAttr)
filt.Train()
instf := base.NewLazilyFilteredInstances(inst, filt)

NewBinningFilter creates the filter, using the structure and data from the given training Instances. The second parameter controls the number of bins used for each Attribute.
Each FloatAttribute must be added using AddAttribute. All of the FloatAttributes can be added at once using the AddAllNumericAttributes function.
Build searches for the maximum and minimum values of each FloatAttribute and determines the bin-width.
Run discretises the Instances in place (IMPORTANT: this means that some data may be lost).

Chi-Merge binning

Chi-Merge binning is a supervised technique which iteratively merges initial bins so long as the merge doesn't affect the class distribution in the combined bin in a manner that's statistically significant. Bramer gives a comprehensive overview of the algorithm.

Code excerpt: Chi-Merge binning

inst, _ := base.ParseCSVToInstances("../examples/datasets/iris_headers.csv", true)
attrs := make([]int, 1)
attrs[0] = 0
inst.Sort(base.Ascending, attrs)
filt := NewChiMergeFilter(inst, 0.90)
filt.AddAttribute(inst.GetAttr(0))
filt.Train()
instf := base.NewLazilyFilteredInstances(inst, filt)

Instances must be sorted by the Attribute used for ChiMerge before Build() is called. This effectively limits ChiMerge to operating on one Attribute at a time. This is a bug. Use multiple ChiMerge filters if further discretisation is needed.
NewChiMergeFilter is used to generate the filter. The second argument is the significance level (in this case 90%). Chi-Merge won't merge adjacent bins if the Chi-Merge statistic is outside of the confidence level indicated by this significance threshold.
Build computes the Chi-Merge bins
Run discretises a set of Instances in place. If a value is lower than or higher than the lowest or highest training value seen, it's assigned the lowest or highest bin.

Support status

Operating Systems	Mac OS X 10.8 Ubuntu 14.04
Go version	1.2
GoLearn version	0.1
Support status	Current
Next revision	On version upgrade