Arx Data Classification

[idhamari]

Arx uses the classification for two goals:

1. As an information loss metric to find the best transformation
2. As a data quality to evaluate the output anonymized data

To use the classification as a metric, select the Utility measure tab then Classification accuracy measure, and select one or more Target variables from Attribute metadata.

To evaluate the anonymized data using classification, select Analyze utility, Classification performance, Classification model then select one or more Target variables:

[idhamari]

Classification as an information loss metric:

The folder metric contains different metrics to find the best transformation. By tracing classification metric:

• Multiple parts:
  • Core: ARXConfiguration result.metric
  • GUI: Model,java config.setMetric
  • GUI: ViewPropertiesInput.java metric = model.getMetricDescription()
  • GUI: WorkerAnonymize.java config.setQualityModel
• public Metric<?> createInstance(MetricConfiguration config)
• public static Metric createClassificationMetric(double gsFactor). gsFactor: generalization and suppression factor which can be set via coding model:

• public static Metric createClassificationMetric(double gsFactor)
• MetricSDClassification(double gsFactor)
• AbstractMetricSingleDimensional
• Metric(final boolean monotonicWithGeneralization, final boolean monotonicWithSuppression, final boolean independent, final double gsFactor)

[idhamari]

Classification as a Data analysis:

All classifier implementations and data preparation can be found in arx/aggregates/classification. An example is the Linear Regression classifier.

The functions encodeFeatures and encodeClass do the data preparations.

Currently, there are only 3 classifiers supported:

1. Linear Regression from APACHE MAHOUT.
2. Naive Bayes from SMILE.
3. Random forests from SMILE.

Note: SVM was integrated in the past using SMILE but removed due to lack of performance.

How Arx classification works:

1. Each input and output data is separated using 10-fold cross-validation into training and validation datasets. Note the number of folds can be modified:

   • from GUI menus: Edit>Settings>Utility analysis>Number of folds

   • from API:

        ARXClassificationConfiguration configLR = ARXClassificationConfiguration.createLogisticRegression();
        configLR.setNumFolds(5); // 5 folds
     

2. StatisticsClassification loop through the rows of the training data and passes the row to:

   • input classifier for the input data.
   • input zero classifier (just compute the frequency of each class, no training is happening)
   • output classifier for the output data

3. In the classifier, e.g. Linear Regression, the data are preprocessed using encodeFeatures and encodeClass, then call lr.train() function of the classification library

4. After the classifier is trained, it is used to validate the validation dataset and obtain the classification results e.g. accuracy 1 2, BrierScore 1 2, ..etc. Note that the relative accuracy is computed in the GUI part in ViewStatisticsClassificationOutput

[idhamari]

Data pre-processing for classification:

In encodeFeatures and encodeClass:

1. It gets a row from the training data and prepares a DenseVector, a packed vector defined in the classification library,

Check what is the input and the output at each row for inputdata
Check what is the input and the output at each row for outputdata

Check other classifiers preprocessing
.