RandomTree is now much faster. Code taken originally from REPTree to avoid re-sorting of data has been removed: it is not beneficial in this case because it sorts *all* attributes before the tree is built; in the new version, only the local data for the randomly selected attributes is sorted. This also means that RandomForest is much faster (it can be as much as an order of magnitude faster on UCI datasets). The memory footprint has also been reduced. RandomTree now also has an option to perform backfitting, so that unbiased probability estimates can be obtained by backfitting a hold-out set.
The metaheuristic neighborhood search method applied to feature selection (weka.attributeSelection.TabuSearch). See:
Abdel-Rahman Hedar, Jue Wang, Masao Fukushima: Tabu search for attribute reduction in rough set theory. Soft Comput. 12(9): 909-918 (2008).
Thanks to Adrian Pino for this contribution.
The Wrapper (weka.attributeSelection.WrapperSubsetEval) subset evaluator now supports other evaluation metrics aside from error rate (classification) and RMSE (regression). Supported metrics now include: MAE, F-measure, AUC, RMSE (probabilities), MAE (probabilities).
Apriori (weka.associations.Apriori)can now make use of market basket-type data in sparse instances format. In this case, zeros (which are not stored explicitly in the sparse format) are used to represent absence of items from baskets. Previously, market basket data was encoded by using Weka's missing value indicator to indicate absence of items. Sparse data allows larger data sets to be loaded and processed by Apriori.
EMImputation (weka.filters.unsupervised.attribute.EMImputation) replaces missing missing values in a data set by using Expectation Maximization with a multi-variate normal model. This is a sophisticated alternative to Weka's standard imputation using means/modes. See:
Schafer, J.L. Analysis of Incomplete Multivariate Data, New York: Chapman and Hall, 1997.
Thanks to Amri Napolitano for this contribution.
Sort the labels of a nominal attribute (weka.filters.unsupervised.attribute.SortLabels.)
Import of PMML TreeModelis now supported.
Support has been added in Weka 3.7 for reading and writing MatLab's ASCII file format (single matrix per file only).
The Knowledge Flow now has GUI support for environment variables. Both system and Java variables can be used in file paths and other settings for all data sources and data sinks (including the serialized model saver component).
Weka now has support for the basic editing and execution of Groovyscripts. Groovy is a dynamic language, which allows you to quickly experiment with using Weka's core classes programatically. See also the Groovy scripting plugin for the Knowledge Flow.
Since instance weights can now be specified in standard ARFF and XML-based ARFF formats, the Preprocess panel of the Explorer has been upgraded to reflect information on weights and take weights into account for displayed statistics and histograms.