Releases: awojna/Rseslib
v3.5.0
This release adds the AQ15 rule classifier and Rseslib neural network to Weka and enables to save these two classifiers on disk.
CHANGES:
rules/AQ15 and functions/RseslibNN (neural network) added to Rseslib package in Weka.
AQ15 and neural network can be saved on disk (serialization added).
"classificiationByRuleVoting" in AQ15 renamed to "ruleVoting".
The "ruleVoting" flag in AQ15 can be toggled on a trained classifier.
Performance of AQ15 improved.
Default time for training the neural network changed from 120 to 30 seconds.
BUGS FIXED:
Infinite loop possible in AQ15.
v3.4.1
This release improves visualization of rough set classifier and includes important bug fixes.
CHANGES:
Text represention of rules improved.
Vote count for each decision class added in visualization of classification by rough set classifier in QMAK.
Filtering and sorting of rough set rules by decision values added in QMAK.
User can change the time limit while retraining neural network in QMAK.
Help in QMAK improved.
8 extra exemplary ARFF datasets added to the Rseslib package.
BUGS FIXED:
Stack overflow error in rough set classifier / local MD discretization.
No message and no effect of "Classify row" pop-up menu command in table view in QMAK if no row or no classifier is selected.
"Add classifier type" command in QMAK menu not working.
v3.4.0
This release includes new algorithms for lower and upper approximation, positive region and approximation accuracy, and a new program calculating attribute significance, improves QMAK performance and enhances KNN visualization.
CHANGES:
Lower and upper approximation of decision classes and positive region added (chapter 'Rough sets' in user guide).
Approximation accuracy and significance of subsets of attributes added (chapter 'Attribute evaluation' in user guide).
Command-line program calculating significance of subsets of attributes added (section 'Calculate significance of attributes' in user guide).
Names of the following classes implementing classifiers changed:
RoughSetRuleClassifier to RoughSetRules,
KnnClassifier to KNearestNeighbors,
LocalKnnClassifier to LocalKNearestNeighbors,
C45 to C45DecisionTree,
AQ15Classifier to AQ15,
NeuronNetwork to NeuralNetwork,
NaiveBayesClassifier to NaiveBayes,
SVM to SupportVectorMachine,
PcaClassifier to PrincipalComponentNetwork,
LocalPcaClassifier to LocalPrincipalComponentNetwork.
Names of the following classifiers in the Weka package changed:
RseslibKnn to RseslibKNN,
LocalKnn to LocalKNN.
Linear orders and sorters adjusted to standard Java interfaces.
The window for object rename opens on single click on an active icon in QMAK.
Prompt improved and the old name inserted while opening the window for object rename in QMAK.
User can specify the split ratio when dividing a data table in QMAK.
Data table view in QMAK improved.
Drop-down list with possible values available for each text and for each true/false option while setting options of classifiers in QMAK.
Vote count for each decision class added in visualization of classification by KNN in QMAK.
Information about voting weight of each nearest neighbor added in visualization of classification by KNN in QMAK.
Placement search status removed from KNN visualization in QMAK.
Help in QMAK improved.
Format of QMAK configuration file with list of classifiers simplified.
BUGS FIXED:
Unhandled exception while creating AQ15 model.
Editing of nominal attributes in data table in QMAK not working.
Data point not distinguished with cross in visualization of its KNN classification in QMAK if it is from the data table used to train the classifier.
Empty result window in QMAK after running experiment only with SVM and PCA classifiers on unsupported data.
v3.3.1
This release adds the full configuration to the programs computing reducts and rules, and improves presentation of rules and extends the report from experiments in QMAK.
CHANGES:
The full configuration made available in the command-line programs computing reducts and rules.
Text presentation of reduct based rules and AQ15 rules improved.
Rules sorted by support or accuracy displayed the greatest at the top in QMAK.
Presentation of rules in QMAK improved.
F-measure, G-mean and sensitivity added to the report from experiments in QMAK for datasets with binary decision.
Communication between the server and the worker nodes in Simple Grid Manager improved.
BUGS FIXED:
"Nominal value local integer code is out of range" error while using EqualFrequency discretization on data with few distinct values on a numerical attribute.
"Jumping text" effect in the panel with description of objects while hovering over the points representing the objects in k-NN visualization.
Missing value in drop down list of rule selection combo box in QMAK.
v3.3.0
This release includes the new classifier RIONIDA for imbalanaced data.
CHANGES:
The new classification algorithm RIONIDA dedicated to imbalanced data with two decision classes is implemented. It is available in WEKA, in QMAK and in Simple Grid Manager. The algorithm combines instance-based learning with rule induction and enables to differentiate the importance of the decisions and to control the impact of rules on the decision selection process. Different parametrisations correspond to different approaches, including a pure instance-based approach, a pure rule-based approach, and combination of both. The algorithm applies multi-dimensional optimization of classification measures relevant for imbalanced data.
Framework for classification is improved for imbalanced data: information about minority decision class added in data tables, test results extended with new classaification measures (G-mean, F-measure, sensitivity, specificity and precision).
Text presentation of test results and their presentation in QMAK is improved and extended with new classification measures (G-mean, F-measure, sensitivity).
BUGS FIXED:
Progress window in QMAK hanging if a classifier returns error on train.
v3.2.5
This release updates Weka library in QMAK and relocates packages for Weka to GitHub.
CHANGES:
Weka library used in QMAK is updated from the version 3.8.5 to the version 3.8.6.
Rseslib packages for Weka relocated to https://github.com/awojna/Rseslib/releases/.
v3.2.4
This release includes several bug fixes.
CHANGES:
Weka library used in QMAK is updated from the version 3.8.0 to the version 3.8.5.
BUGS FIXED:
Parsing data in Rseslib format: whitespaces following the last value in a line are added to this value.
The error message does not provide the line and column number if a data file in Rseslib format contains incorrect value format for a numerical attribute.
The error message provides no details while loading incorrect data format in QMAK.
Compatibility of the Rseslib classifiers capabilities with training data is not verified in Weka.
v3.2.3
This release improves the ROC area (AUC) measure of Rseslib classifiers in Weka by implementing the method computing class distribution.
CHANGES:
The distributionForInstance() method is implemented for all three Rseslib classifiers exposed in Weka: RoughSet, RseslibKnn and LocalKnn.
BUGS FIXED:
Rseslib classifiers throw exceptions when stopped by a user in Weka GUI.
v3.2.2
This release includes new implementation of Johnson's algorithm for reducts and new command-line programs computing reducts and rules.
CHANGES:
New implementation of Johnson's algorithm for reducts fixing defects of the previous implementation and incomparably faster.
Command-line programs computing and writing reducts and rules added as examples.
Command-line programs running tests on Rseslib classifiers simplified and extended with an example of a classifier with non-default parameters.
BUGS FIXED:
Algorithm computing Johnson's reducts:
- Incorrect set of attributes is returned as reduct.
Algorithm computing partial reducts:
- Algorithm adds the decision attribute to reduct when decision is not the last attribute.
Rough set based rule classifier:
- Typo in the names of discretization methods.
v3.2.1
This release improves visualization of classifiers in Qmak and debug mode of classifiers in Weka.
CHANGES:
Debug mode of Rseslib classifiers in Weka displays statistics of training on the console:
- Number of generated rules for RoughSet classifier,
- Optimized number of neighbors for RseslibKnn and LocalKnn.
KnnVis:
- Classified object and its neighbors represented with more visible bigger graphical objects.
C45Vis:
- Classification path represented with more visible thicker line.
VisualNeuronNetwork:
- Descriptive labels revised and improved.
BUGS FIXED:
KnnVis:
- Attribute values displayed after selection or hovering over objects are sometimes incorrect.