Skip to content


Subversion checkout URL

You can clone with
Download ZIP
Large RDF hierarchies as vector spaces
Latest commit 216d105 @moustaki moustaki Bumping version number



This Python library helps generating a vector space from very large hierarchies encoded in RDF. An obvious example application is to generate a vector space from a SKOS hierarchy or an RDFS subclass hierarchy.

Getting started

Running the tests:

    $ nosetests

    $ python install

Example use

    $ wget
    $ bunzip2 skos_categories_en.nt.bz2
    $ python
    >>> from import Space
    >>> space = Space('skos_categories_en.nt')
    >>> space.similarity_uri(category1, category2)

Constructing a vector space for the entire DBpedia SKOS category
hierarchy (3M triples) takes a couple of minutes on a commodity laptop, 
and has a memory footprint of about 500M.

Alternatively, a subset of it is available in the examples/ directory.

How it works

For each topic t in the hierarchy, we consider the set of its parents
parents(t, k) at a level k. We construct a vector for each t in a space
where each dimension corresponds to a topic d in the hierarchy. The value
of t on dimension d is defined as follows:

    t_d = \sum_{k = 0}^{max_depth} \sum_{d \in parents(t, k)} decay^k

where max_depth and decay are two parameters, which can be used to influence
how much importance we attach to ancestors that are high in the
category hierarchy.

They can be specified as follows:

    >>> Space.max_depth = 8
    >>> Space.decay = 0.9

Licensing terms and authorship

See 'COPYING' and 'AUTHORS' files.
Something went wrong with that request. Please try again.