Preferences Panel
Overview
This panel allows users to view and edit preference settings. This panel can
be seen in Figure 1.
Training Preferences
- Fraction of Recordings to use for Training: The fraction of candidate
recordings to use for testing during Randomly Partitioned Recs training
or cross-validation. Must be greater than 0 and less than 1.
- Train With Overide Features: Whether training should be performed
with the features marked as Selection Candidates in the Features
Panel as candidates for feature selection (unchecked) or with the features
marked as having Overide Status as candidates for feature selection
(checked).
- Train With One-D Features: Whether training is to be performed
with classifier ensembles that include a KNN classifier for all selected one-dimensional
features.
- Train With Multi-D Features: Whether training is to be performed
with classifier ensembles that include a neural net classifier for each selected
multi-dimensional feature.
- Find 1-D Features Selection: Whether or not feature selection should
be performed on the one-dimensional features, if any, during training. A value
of unchecked means that all one-dimensional features will automatically be
selected.
- Find 1-D Feature Weightings: Whether or not feature weighting should
be found for the selected one-dimensional features, if any, during training.
A value of unchecked means that all one-dimensional features are given equal
weighting.
- Find Classifier Selection: Whether or not classifier selection
should be performed on all classifiers in classifier ensembles during training.
A value of unchecked means that all available classifiers will automatically
be selected.
- Find Classifier Weightings: Whether or not classifier weighting
should be performed on all selected classifiers in classifier ensembles during
training. A value of unchecked means that half of the weighting is given to
the KNN classifier and the remaining weighting is divided among all available
neural networks.
- Selection/Weighting Training Fraction: The fraction of training
samples that will actually be used for training when feature selection and
weighting are calculated for one-dimensional features. The remaining recordings
are used for testing. Only has an effect of the feature selection/weighting
process, and nothing else. This value must be between 0 and 1.
- Classify Into Root Categories: Whether or not a classifier ensemble
should be trained that classifies all recordings into none or more root categories.
Note that this option should not be enabled if one wishes to classify recordings
using classifiers that were not trained with this option, but it may be disabled
even if the classifiers were trained with this option.
- Classify Into Direct Descendants of Parents: Whether or not hierarchal
classifier ensembles should be trained, which is to say an ensemble for each
category with children that classifies recordings into none or more of its
children. Note that this option cannot be enabled if one wishes to classify
recordings using classifiers that were not trained with this option, but it
may be disabled even if the classifiers were trained with this option.
- Classify Into All Leaf Categories: Whether or not a classifier
ensemble should be trained that classifies all recordings into none or more
leaf categories. Note that this option cannot be enabled if one wishes to
classify recordings using classifiers that were not trained with this option,
but it may be disabled even if the classifiers were trained with this option.
- Classify Using Round Robin Ensembles: Whether or not round robins
classifier ensembles should be trained, which is to say an ensemble for every
possible pair of leaf categories that classifies recordings into none or more
of the two categories. Note that this option cannot be enabled if one wishes
to classify recordings using classifiers that were not trained with this option,
but it may be disabled even if the classifiers were trained with this option.
Training Termination Preferences
- Auto End GA Training: Whether or not genetic algorithm training
should be terminated early (i.e. before the Max Number GA Epochs
has been reached) if the best fitness has not changed beyond a certain point
for more than a certain number of epochs.
- Max Number GA Epochs: The maximum number of epochs that should
be iterated through for feature (but not classifier) selections and weightings.
Must be 1 or greater.
- Classifier vs. Feature Iter. Multiplier: The multiplier applied
to the Max Number GA Epochs in order to get the corresponding value
for classifier selections and weightings. Is also used to multiply the population.
Must be 1 or greater.
- Max Change in GA Error for Training Abort: The maximum acceptable
change in fitness that will make it possible to end genetic algorithm evolution
before the Max Number GA Epochs has been reached. Must be greater
than 0.
- Min Consec GA Epochs for Training Abort: The minimum number of
epochs that must go by in genetic algorithm evolution where the fitness is
below the Max Change in GA Error for Training Abort in order for
training to end before the Max Number GA Epochs is reached. Must
be 0 or greater.
- Auto End NN Training: Whether or not neural network training should
be terminated early (i.e. before the Max Number NN Iterations have
been reached) if the sum of squares error has not changed beyond a certain
point for more than a certain number of epochs.
- Max Number NN Iterations: The maximum number of neural network
epochs that should be iterated through. Must be 1 or greater.
- Max Change in NN Error for Training Abort: The maximum acceptable
change in sum of squares error that will make it possible to end neural network
training before the Max Number NN Iterations has been reached. Must
be greater than 0.
- Min Consec NN Iters for Training Abort: The minimum number of epochs
that must go by in neural network training where the sum of squares error
is below the Max Change in NN Error for Training Abort in order for
training to end before the Max Number NN Iterations is reached. Must
be 0 or greater.
Classification Preferences
- Hierarchal Relative Weighting: The relative weighting to be given
to the hierarchal classifier ensembles when the results of multiple classifier
ensembles are combined. This weighting is relative to the All Leaf Category
Relative Weighting and Round Robin Relative Weightings weightings.
Must be 0 or greater.
- All Leaf Category Relative Weighting: The relative weighting to
be given to the flat all leaf classifier ensemble when the results of multiple
classifier ensembles are combined. This weighting is relative to the Hierarchal
Relative Weighting andRound Robin Relative Weightings weightings.
Must be 0 or greater.
- Round Robin Relative Weightings: The relative weighting to be given
to the hierarchal classifier ensembles when the results of multiple classifier
ensembles are combined. This weighting is relative to the All Leaf Category
Relative Weighting and Hierarchal Relative Weighting weightings.
Must be 0 or greater.
- Round Robin Score Divisor: The divisor that is used to divide
all round robin scores during classification. A higher number makes it less
likely that the round robin classifiers will select too many categories. Must
be 1 or greater.
- Should Use Winner Only For Round Robin: Whether or not all but
the highest round robin score should be set to 0 for each recording.
- Min Score for Auto 1st Choice: The minimum score that a category
must have to automatically be
counted as a first choice during classification, regardless of the scores
of other categories. Must be between 0 and 1.
- Min Score to be 1st Choice: The minimum score that a recording
must have to qualify as a first choice during classification. Must be between
0 and 1.
- % Below Highest to be 1st Choice: The percentage below the score
of the highest category that sets the minimum for a category in order for
it to be classified as a first choice (unless it has a score over Min
Score for Auto 1st Choice). Must be between 0 and 100.
- Min Score to be 2nd Choice: The minimum score that a recording
must have to qualify as a second choice during classification. Must be between
0 and 1.
- % Below Highest to be 2nd Choice: The percentage below the score
of the highest category that sets the minimum for a category in order for
it to be classified as a second choice. Must be between 0 and 100.
- Only Choose One Winner in Combined Classification: Whether or not
only one winning category is permitted per recording
Reporting Preferences
- Seconds Between Progress Bar Polls: How often, in seconds, that
the training progress bar (see the Classification
Panel) updates its status. Must be 0 or greater.
- Should Display Training Times in Raw Seconds: Whether training
times should be displayed as seconds (checked) or as hours, minutes and seconds,
as appropriate (unchecked).
- Report Classifier Settings: Whether or not classifier settings
should be included in classification reports.
- Report Ensemble Success Rates: Whether or not classification reports
should include overall classification statistics for the combined and individual
classifiers for each ensemble of classifiers. This option is only available
when recordings have been reserved for testing during training (see the Classification
Panel).
- Report Root Category Success Rates: Whether or not classification
reports should include how often the root categories of recordings were successfully
found. This option is only available when recordings have been reserved for
testing during training (see the Classification
Panel).
- Report Confusion Matrix: Whether or not classification reports
should include a confusion matrix. This option is only available when recordings
have been reserved for testing during training (see the Classification
Panel).
- Report Ensemble Eligible Categories: Whether or not classification
reports should include a list of the eligible categories into which recordings
could be classified by each ensemble of classifiers.
- Report Combined Ensemble Classifications: Whether or not classification
reports should include the combined classification results (i.e. name of final
categories) for each recording by each ensemble of classifiers.
- Report Individual Classifier Classifications: Whether or not classification
reports should include separately, for each classifier, the classification
results (i.e. name of final categories) for each recording for an ensemble
of classifiers.
- Include Classification Scores w/ Category Names: Whether or not
classification reports should include the associated scores along with the
names of the winning category(ies) for each recording.
- Include Secondary Classification Results: Whether or not secondary
classifications should be included in classification reports alongside primary
classifications.
- Include Model Results With Classifications: Whether or not classification
reports should include the names of model classifications, if available. This
option is only available when recordings have been reserved for testing during
training (see the Classification Panel).
- Report Combined Ensemble Category Scores: Whether or not classification
reports should include the combined classification scores for each category
of each recording for each ensemble of classifiers.
- Report Individual Ensemble Category Scores: Whether or not classification
reports should include, separately for each classifier, the classification
scores for each category of each recording for each ensemble of classifiers.
- Report Model Results in Separate Table: Whether or not classification
reports should include, as a separate table, a list of the eligible classification
categories for each ensemble of classifiers, along with their code numbers.
- Lines to Skip When Reporting Training Errors: The number of epochs
that are skipped between reportings of fitness or sum of squares error in
training reports. Must be 0 or greater.
- Report Recordings Used to Train Ensemble: Whether or not training
reports should include the names and model leaf categories of the recordings
used in training.
- Report Ensemble 1-D Feature Training Error: Whether or not training
reports should include the one-dimensional feature selection and/or weighting
training progress for ensembles of classifiers.
- Report Ensemble Multi-D Training Error: Whether or not training
reports should include training progress for the multi-dimensional classifiers
in ensemble of classifiers.
- Report Ensemble Classifier Selection Error: Whether or not training
reports should include classifier selection training progress in ensembles
of classifiers.
- Report Ensemble Classifier Weighting Error: Whether or not training
reports should include classifier weighting training progress for ensembles
of classifiers.
- Report Ensemble 1-D Features Selected: Whether or not selection/weighting
reports should include the one-dimensional features selected for each ensemble
of classifiers.
- Report Ensemble 1-D Feature Weightings: Whether or not selection/weighting
reports should include the one-dimensional feature weightings for each ensemble
of classifiers.
- Report Ensemble Classifiers Selected: Whether or not selection/weighting
reports should include the classifiers selected by each ensemble of classifiers.
- Report Ensemble Classifier Weightings: Whether or not selection/weighting
reports should include the classifier weightings for each ensemble of classifiers.
Buttons
- Neural Net Preferences: Brings up the Feedforward
Neural Network Initial Settings Dialog Box.
- Genetic Algorithm Preferences: Brings up the Genetic
Algorithm Initial Settings Diaolog Box.
- Load Preferences: Loads the preferences contained in a preferences_file
XML file on disk into memory and displays them. Allows the user to choose
the file using a file chooser dialog box.
- Save Preferences: Saves the current preferences into a preferences_file
XML file on disk. If a path is given in the Preferences file field
of the Configure File Locations
Dialog Box, then automatically saves the file to that path, overwriting
any existing file. If this field is blank, then allows the user to choose
the file location using a file chooser dialog box.
- Save Preferences As: Saves the current preferences into a preferences_file
XML file on disk. Allows the user to choose the file location using a file
chooser dialog box.
Status Bar
The status bar has no function on this panel.
Warnings
- Training preferences only affect behaviour at beginning of training, and
subsequent changes will have no effect on classifications performed using
existing trained classifiers. Exceptions are the Classify Into Root Categories,
Classify Into Direct Descendants of Parents, Classify Into All
Leaf Categories and Classify Using Round Robin Ensembles preferences,
which play a roll during classification as well as training. See the descriptions
of these preferences above for details.
Screen Shots
Figure 1: Preferences panel.