Modéliser les données multi-relationnelles comme des séquences / Modeling multi-relational data as sequences

The objective of the project is to automatically learn missing information in knowledge bases (KBs), which are becoming essential tools to deal with big data, since they provide means to organize, manage and retrieve all this knowledge. These databases are huge directed multi-relational graphs, whose nodes correspond to entities connected by edges representing a certain relationships between them. These databases are far from being complete, so new tools are needed to complete them by adding new facts; this is termed link prediction. This project proposes to tackle this problem by learning representations for each of the elements of the graph. However up to now, all proposed solutions do that by modeling of triples in isolation. We have

observed several limitations in these models that we attribute to this fact, so we aim at modeling sequences of triples, that could capture longerterm interactions. We propose to do such modeling using Recurrent Neural Networks, which have recently shown to be very powerful for dealing with sequences.

Faculty Supervisor:

Yoshua Bengio

Student:

Alberto Garcia-Duran

Partner:

Discipline:

Computer science

Sector:

University:

Program:

Globalink