Dynamic link prediction and time prediction on Temporal Knowledge Graphs

What will you do?

The master student will work on prediction tasks on Temporal Knowledge Graphs, such as dynamic link prediction and time prediction. They will be employing Neural Networks (NNs) for this, more specifically Graph Neural Networks (GNNs).

Wait, what are Temporal Knowledge Graphs, GNNs and all that?

Temporal Knowledge Graphs:

• Knowledge Graphs (KGs) augmented with temporal information
• Edges are not only stating that facts are true, but also when they are true
  • Example:
    • Standard (static) KG fact (a.k.a. triple): (Obama, president, US)
    • Temporal KG fact: (Obama, president, US, [2009, 2017])

GNNs:

• Class of deep learning architectures that work on graph-structured data
• Constructed in a way that permits scaling up with the size of the graph and using transformations invariant to permutation of nodes
• Successfully applied to a series of tasks and one of the main architectures used for processing both static and temporal KGs

Dynamic Link Prediction:

• Addressed question: What will hold at a certain point in time?
• Inspired by the Link Prediction task in static KGs
• Can be framed as predicting the subject or the object in a potential temporal KG fact: (?, predicate, subject, time) or (subject, predicate, ?, time)

Time Prediction:

• Addressed question: When will a certain fact hold?
• Task framed as predicting a timestamp of a certain edge: (subject, predicate, object, ?)

Why would you choose this?

The student working on this project will get and improve important skills

• Learn more about neuro-symbolic AI and applying NNs to symbolic data
• Learn about processing graphs and temporal information with NNs
• Improve technical reading skills by surveying the literature in the area
• Improve their coding and experimentation skills
• Exercise their creativity (since there are not so many systems designed for the tasks in question)

Research value:

• Dynamic Link Prediction and Time Prediction on Temporal Knowledge Graphs are relatively under-researched, especially considering methods that work on continuous time.
•  Existing methods have their limitations; for example they cannot process edge removal: while they might predict that Obama became a president in 2009, they don’t have the expressivity to say that the presidency ended in 2017.
• Addressing such limitations would be highly valuable from a research perspective and might lead to publications in scientific venues.