| Structure of ACE |
|---|
Overview
The structure of ACE is designed such that all possible sources can access ACE in the same way; ACE's functionality is independent of the source that calls it. Code is divided based on the processing that it performs (code that trains is in a class called Trainer, code that classifies is in a class called InstanceClassifier, etc). Classes found in the ace.datatypes package are storage objects used for communication between the main processing classes. The main processing classes and datatypes that can be used by the graphic user interface, the command line interface, or external APIs are described below.
Main Processing Classes
- Coordinator - The class through which all sources can access ACE's training, classification, cross validation, and experimentation functionality. The following classes are intended to only be called by Coordinator. All other sources need only call the appropriate methods in Coordinator. Loading and preparation of instances as well as dimensionality reduction is performed in this class prior to passing instances to processing classes.
- Trainer - Trains the specified type of Weka Classifier based on the given training instances. The trained Classifier is stored and saved in a TrainedModel object.
- InstanceClassifier - Classifies a set of instances using a trained Weka Classifier. This class reads the TrainedModel object from the specified file and uses it to classify the given instances. The classifications are stored and returned in an array of SegmentedClassification objects. In the context of cross validation, a classified Instances object is returned. Classifications can be written to a Weka ARFF file or an ACE XML classifications file (depending on which format was used to load the initial instances).
- CrossValidator - Cross validates the given instances with the specified type of Weka Classifier and the specified number of folds. Instances are partitioned randomly into training and testing data for each fold (note that the original order of instances remains the same, the random assignment affect only whether an instance is used for training or testing in any one fold). CrossValidator makes calls to the Trainer and InstanceClassifier classes to evaluate the performance of a specific classification approach. The specified type of Weka Classifier is trained on the training data and tested on the testing data for each fold. Statistics are stored for each fold and used to calculate average statistics to represent the overall success of the classification approach being tested.
- Experimenter - Finds the best classification approach by making repeated calls to the CrossValidator class. Different types of Classifiers are tested with multiple types of dimensionality reduction. Dimensionality reduction is performed first. Experimenter calls DimensionalityReducer to get an array of Instances objects, each cell containing a different dimensionality reduced version of the original instances. Each type of Classifier is cross validated with each set of dimensionality reduced instances. Results are saved in multiple files. One file will be created for each set of dimensionality reduced instances. A summary of the results for each cross validation for each dimensionality reduction will be written in the file for the corresponding set of instances. The results for the best found Classifier for each dimensionality reduction are printed at the beginning of printed at the beginning of each file. The best classification overall is chosen by comparing the best results for each dimensionality reduction. A copy of the results summary for the best found classification approach will be written in a separate "best results overall" file. After the best classification has been chosen, validation is performed using a publication set that was set aside at the beginning of the experiment. A new Classifier of the chosen type is created and trained on the chosen type of dimensionality reduced instances (all instances are included except for the publication set). The newly trained Classifier is tested on the publication set and results are saved to the "best results overall" file.
- DimensionalityReducer - Reduces the dimensionality of a set of instances. This class is called by the Coordinator class to reduce the dimensionality of the training data prior to training and cross validating. This class is also called by the Experimenter class to create an array of multiple dimensionality reduced versions of an original set of instances.
ACE Datatypes
For a detailed description of how to use ACE from the command line, please see the ACE Command Line Manual.
- DataSet - Stores the feature values for a set of instances. This datatype is similar to the Weka Instances object. This object corresponds to the ACE XML feature vectors file.
- FeatureDefinition - Stores the possible features for which the instances of this ACE project could have values. This object corresponds to the ACE XML feature definitions file.
- SegmentedClassification - Stores the classifications for a set of instances. This object corresponds to the ACE XML model classifications file.
- Taxonomy - Stores the hierarchy of the classes into which an instance of this project may be classified.
- DataBoard - Holds reference to the main ACE datatypes (Taxonomy, DataSet, Feature Definition, SegmentedClassifiation). Constructors of this class enable the creation of these main ACE datatypes from either ACE XML files or Weka ARFF files.
- Project - Holds reference to the paths to each of the component ACE XML files of an ACE project. This object is associated with the ACE XML project file. This class contains methods for loading, parsing, and saving an ACE project file.
- TrainedModel - Serializable object that stores the Weka Objects associated with a Classification. Trainer will train the Weka Classifier, set the fields of this class, and save it to a file. InstanceClassifier will read this object from a file and access its fields to be used for classification. This object contains reference to a Weka Classifier, a Weka AttributeSelection, and a Weka Attribute object describing the possible classes to which an Instance may belong.
- CrossValidationResults - Holds the results of a cross validation. Statistics like error rate, standard deviation of error rates, and confusion matrix can be accessed through the public fields of this class. In experimentation, a 2D array of these objects is used to store the results for each classifier and each dimensionality reduction.
Questions and Comments