This repository contains the data for the experiments used in the paper: On Inductive Abilities of Latent Factor Models for Relational Learning, Théo Trouillon, Éric Gaussier, Christopher R. Dance, Guillaume Bouchard.
In the folder relation_properties are the generated matrices for each
of the 13 combination of interest between reflexivity, symmetry and transitivity,
as described in the paper.
Matrices are saved in the .mat format under the variable name y.
They can be loaded from matlab or from python with the scipy.io.loadmat function.
Ones represent positive samples, -1 negatives, and 0 missings (used for diagonal in
the non-[ir-]reflexive cases).
The families data is represented in text files in the folder families,
following a binary predicate syntax: [!]relation(subject_entity,object_entity).
Leading ! indicates a negative fact. Families are numbered from 1 to 5,
and the set of facts of each family is split between the four main relations
(father, mother, daughter, son) in files suffixed by _4main.db,
and the 13 other relations in files suffixed by _13other.db,
as described in the paper.