Learning with Temporal Knowledge Graphs
Yunpu Maa , Zhen Hanb and Volker Trespc
a Ludwig Maximilian University of Munich
b Ludwig Maximilian University of Munich & Siemens CT
c Ludwig Maximilian University of Munich & Siemens CT



                Abstract
                Temporal knowledge graphs, also known as episodic or time-dependent knowledge graphs, are large-
                scale event databases that describe temporally evolving multi-relational data. An episodic knowledge
                graph can be regarded as a sequence of semantic knowledge graphs incorporated with timestamps.
                In this talk, we review recently developed learning-based algorithms for temporal knowledge graphs
                completion and forecasting.

                Keywords
                Temporal Knowledge Graphs, Representation Learning


1. Learning with                                attention in the community. In the pioneer-
                                                ing work [3, 4], we investigate representa-
   Knowledge Graphs                             tion learning models of episodic knowledge
If political relations between two countries graphs. To generalize the semantic models
becomes more tense, will it affect the inter- for knowledge graphs to temporal knowledge
national trades between them? If yes, which graphs, we introduce unique latent represen-
industries will bear the brunt? Modeling the tations for each timestamp. The deep con-
relevant events that can be temporarily af- nections between temporal knowledge graphs
fected by international relations is the key and cognitive functions, e.g., semantic and
to answer this question. However, the issue episodic memory, will be elaborated in this
of how to model these complicated temporal talk [5].
events is an intriguing question. A possible       Besides, we will introduce a non-parametric
way is to embed events in a temporal knowl-     Graph   Hawkes process for dynamic events
edge graph, which is a graph-structured multi-  forecasting in temporal knowledge graphs [6],
relational database that stores an event in the and recent  developments   of explainable rea-
form of a quadruple.                            soning  and  forecasting  on temporal knowl-
   For instance, Global Database of Events,     edge  graphs.
Language, and Tone (GDELT) [1] and Inte-           Acknowledgement This work is supported
grated Crisis Early Warning System (ICEWS)      by the Munich Center for Machine Learning
[2] are two available event-based temporal      Center  (MCML) funding project.
knowledge graphs that have been drawing

Proceedings of the CIKM 2020 Workshops, October 19-20,
2020, Galway, Ireland
" cognitive.yunpu@gmail.com (Y. Ma)                                    References

 © 2020 Copyright for this paper by its authors.
 Use permitted under Creative Commons License Attribution 4.0 Inter-   [1] K. Leetaru, P. A. Schrodt, Gdelt: Global
 national (CC BY 4.0).
 CEUR Workshop                    Proceedings           (CEUR-             data on events, location, and tone, 1979–
 WS.org)
    2012, in: ISA annual convention, vol-
    ume 2, Citeseer, 2013, pp. 1–49.
[2] E. Boschee, J. Lautenschlager, S. O’Brien,
    S. Shellman, J. Starz, M. Ward, Icews
    coded event data, Harvard Dataverse 12
    (2015).
[3] Y. Ma, V. Tresp, E. A. Daxberger, Em-
    bedding models for episodic knowledge
    graphs, Journal of Web Semantics 59
    (2019) 100490.
[4] V. Tresp, Y. Ma, S. Baier, Y. Yang, Embed-
    ding learning for declarative memories,
    in: European Semantic Web Conference,
    Springer, 2017, pp. 202–216.
[5] V. Tresp, S. Sharifzadeh, D. Konopatzki,
    Y. Ma, The tensor brain: Semantic decod-
    ing for perception and memory, arXiv
    preprint arXiv:2001.11027 (2020).
[6] Z. Han, Y. Ma, Y. Wang, S. Günnemann,
    V. Tresp, Graph hawkes neural network
    for forecasting on temporal knowledge
    graphs, in: 8th Automated Knowledge
    Base Construction (AKBC), 2020.