Skip to content. | Skip to navigation

Informatik 5
Information Systems
Prof. Dr. M. Jarke
Sections
Personal tools
You are here: Home Theses Dynamic Embeddings of Evolving Knowledge Graphs

Contact

Prof. Dr. M. Jarke
RWTH Aachen
Informatik 5
Ahornstr. 55
D-52056 Aachen
Tel +49/241/8021501
Fax +49/241/8022321

How to find us

Annual Reports

Disclaimer

Webmaster

 

 

Dynamic Embeddings of Evolving Knowledge Graphs

Thesis type
  • Bachelor
Status Running
Proposal on 30. Apr 2019 00:00
Proposal room Library I5
Add proposal to calendar vCal
iCal
Supervisor(s)
Advisor(s)

The goal of this Bachelor thesis is the research of updating KG embeddings with new information in order to obtain a dynamic and stable embedding of the fast-evolving KG while reducing the computational effort.

In recent years, the growth of the Internet generated many new ways of gathering structured knowledge about general topics through automatic knowledge extraction from different sources like the World Wide Web as well as about social interaction between users of social networks like Twitter. A popular and common technique of storing this knowledge is the construction of a knowledge graph (KG), representing entities as nodes and relationships between these nodes as edges in a directed graph.

 

To make efficient use of the information stored in KGs for different kinds of learning tasks, which often require vectors of a fixed dimension as inputs, vector space embeddings, i.e. the transformation of the graph structure into a vector space by means of vector assignments to each entity and relationship in the KG, have been explored. While these graph embeddings perform well in various downstream tasks like question answering or recommender systems, the computation of an embedding for a large-scale KG requires significant computing power and a lot of time. The computational effort of graph embeddings is problematic, especially in the case of a fast-evolving KGs gathered e.g. from data streams like Wikipedia click rates or social network analysis, if the embedding is required to stay up to date with the current state of the KG.

Hence, the idea of simple recalculations of the entire embedding after discrete time steps in order to keep an updated embedding of a fast-evolving KG is impractical. Moreover, recalculated embeddings are very distinct from one another and cannot be used as updates of the training data in learning tasks. To continue and not restart the training of models using the KG embedding, it is necessary to guarantee stability of the embedding. Therefore, changes in the KG must lead to smooth changes in the updated embedding.

 

The goal of this Bachelor thesis is the research of updating KG embeddings with new information in order to obtain a dynamic and stable embedding of the fast-evolving KG while reducing the computational effort.

Document Actions