A tool to predict a suitable place for a product in an existing taxonomy.

Using data extracted from a product database, taxonomy-predict fits a tree of classifiers to the already categorized products.

This tree can then be used to classify un-categoried products.

This is a work in progress, preliminary results below.

See setup.txt

Training on a dataset of 55K products, where 5% is reserved for validation, achieved the following :

The classifier to be used, file locations and database settings can be configured at etc/config_openfoodfacts.

Results were achieved with the following classifier configurations:

classifier_module='sklearn.linear_model',
classifier_name='LogisticRegression',
classifier_params={'C':1,'class_weight':'balanced'}

classifier_module='sklearn.svm',
classifier_name='LinearSVC',
classifier_params={'C':1,'class_weight':'balanced'}

classifier_module='sklearn.ensemble',
classifier_name='RandomForestClassifier',
classifier_params={'n_estimators':10}

classifier_module='sklearn.naive_bayes',
classifier_name='MultinomialNB',
classifier_params={'alpha':1}

• results obtained on a Open Food Facts mongodb data dump
• only product category hierarchies of length of at least 5 are considered
• LogisticRegression requires about 10Gb of RAM on OFF data
  • however others may use considerably more

No warranties, provided 'AS-IS'.