Name Classifier

Usage

The program consists of two modules, a RDF_processor and a model. It also contains two demonstration scripts, build_model.py, which can be used to create a model from serialized hash arrays or RDF triples and predict.py which provides predictions for a stored model.

The model file contains a prebuilt model which can be used by the predict script

RDF_processor

RDF_processor.parse_identifiers(ident_file, object)

Constructs and stores a set of <subjects> from ident_file with RDF:type == object using the Redland Python bindings. The format of the RDF file should be:

<subject> RDF:type <object>.

• ident_file: Name of the RDF turtle file to be parsed.
• object: URI of the parsing <object>

RDF_processor.map(map_file, balance=True)

Constructs and stores an array of <object> strings for FOAF:Name predicates, and a corresponding identifier array describing the <subjects>'s presence in the stored set of subjects.

The format of the RDF file should be:

<subject> FOAF:Name <object>.

• map_file: Name of the RDF turtle file to be parsed & mapped
• balance: If True, balances the arrays by downsampling the more prevelant category.

RDF_processor.hash(mapping_size=1000)

Tokenises the array of object strings and hashes them using mmh3 to create and store a scipy_dok sparse matrix.

• mapping_size: The range of the hashes, between [-mapping_size, +mapping_size]

RDF_processor.shuffle()

Shuffles the subject, features and identifier arrays.

RDF_processor.get_features()

Returns the current feature array.

RDF_processor.get_identifiers()

Returns the array of identifiers.

RDF_processor.get_subject()

Returns the array of subject strings.

log_reg

log_reg.log_reg(size, batch_size=1000, alpha=0.2, C=0.0)

• size: Number of features in the array to be modelled
• batch_size: The maximum size of each batch
• alpha: The learning rate for batch SGD
• C: The L2 regularization term

log_reg.fit(X, Y)

Fits dataset X to target Y by minimizing the logistic cost function using Mini-batch Gradient Descent with L2 regularization.

• X: The array to be fitted. Of shape (n_samples, n_features)
• Y: The target array for X. Of shape (n_samples)

log_reg.predict(X)

Predicts the value of X using the fitted model.

• X: Value to be fitted

log_reg.score(X, Y)

Returns the mean successful prediction rate for X against targets Y on the fitted model.

• X: The array to be predicted
• Y: The targets to be compared against

Requirements

• Python 2.7
• numpy
• scipy
• mmh3
• Redland Python bindings
• cPickle