Networks and AI seminar series 2025/2026 edition: 11th September

[Liubov]

Dear all,

After a long break we are restarting the seminar series. 

This year we will change the format, making it more lively: 20 min main idea + 10 mins hands-on session of the seminar + active discussions. 

Hope to see many of you there!

When?

11th of September 4pm CET

Where? 

Online/register to get Teams meeting link We will make sure the seminar stays hybrid with onsite discussions.

What?
Presentation title: Hypergraph Motif Representation Learning

Abstract: Hypergraphs provide a powerful abstraction for modeling high-order interactions in complex systems, extending beyond simple pairwise relations captured by traditional graphs. Within these structures, hypergraph motifs (h-motifs) represent recurrent local connectivity patterns that are critical for understanding the dynamics of domains ranging from biology to social networks. While motif analysis has been extensively explored in graphs, the prediction and representation of h-motifs in hypergraphs remain largely uncharted. In this seminar, we present recent advances on hypergraph motif representation learning, focusing on the novel task of h-motif prediction. We formalize this problem and introduce an end-to-end architecture that combines hypergraph neural networks (HGCNs) with graph convolutions to effectively capture correlations among hyperedges. To address the challenge of robust training, we design an innovative negative sampling strategy that generates non-trivial, close-to-positive motif samples. Through extensive experiments on real-world datasets, our approach demonstrates superior predictive performance over a wide range of baselines, highlighting its effectiveness and robustness.

Biography: Valerio Di Pasquale is a computer science graduate with several years of professional experience. After completing his Bachelor's degree, he collaborated on research activities at ISISLab, Università degli Studi di Salerno, where he also earned his Master's degree. His research interests include Artificial Intelligence and Generative AI, with a focus on complex networks, Graph Neural Networks, and hypergraphs for the modeling and analysis of high-dimensional and relational data.

Link to the paper/github from ACM Digital Library: https://dl.acm.org/doi/10.1145/3690624.3709274

Here is the general google group registration form if you are interested to receive other reminders with it for the future 

https://forms.gle/SnEz7twwzY536RNk8

Thank you,

All the best,

Network seminar organisers

If you are no longer willing to receive the notifications, press unsubscribe from google groups below

https://projects.learningplanetinstitute.org/projects/network-seminars-at-cri/summary 

[Liubov]

Dear all,

We are continuing our seminar series. This week we will have an exciting talk from Mike Tamm (TLU, Estonia). Hope to see many of you there!

When?

9th of October 4pm CET

Where? 

Online/register to get Teams meeting link We will make sure the seminar stays hybrid with onsite discussions.

What?
City representation in the Soviet propaganda: quantifying biases of the Soviet worldview

Representation of geographical locations in cultural data is almost always biased, creating a distorted representation of the world. Identifying and measuring such biases is crucial to understand both the data and the socio-cultural processes that have produced them. We suggest measuring geographical biases in a historical news media corpora by studying the representation of cities. Leveraging ideas of quantitative urban science, we develop a mixed quantitative-qualitative procedure, which allows us to get robust quantitative estimates of the biases. The procedure includes  (i)formulation of a hypothesis based on qualitative analysis of the data, (ii) parameter extraction by optimization of an explicit loss function and (iii) model selection and pruning to avoid overfitting. These three steps form a feedback loop: the output of the last step (and, in particular, the outliers, which the model is unable to explain) can be used as an input for refining the hypothesis.  We apply this procedure to a corpus of Soviet newsreel series 'Novosti Dnya' (News of the Day) and show that city representation grows super-linearly with city size and is further biased by city specialization and geographical location. This allows systematical identification of the geographical regions which are explicitly or covertly emphasized by Soviet propaganda and quantify their importance.

In the talk I will pay attention both to the fine methodological aspects, such as choice of loss function to allow for scarceness of the dataset and information-theoretical criteria of model selection, and to the discussion of the actual result and what they tell us about the underlying Soviet world view.

All the best,

Liubov

[Liubov]

Dear all, 

We are continuing our seminar series. We are really happy to invite you to the next lecture with Tiago Peixoto, which will happen on 13th March at 5 pm CET. 

Please feel free to register here.

Title: Graphs are maximally expressive for higher-order interactions
Abstract:
In this talk, I discuss a recent work [1] where we demonstrate that graph-based models are fully capable of representing higher-order (multibody) interactions, challenging the common claim that only hypergraphs can capture such dependencies. We emphasize that graphs are not limited to pairwise interactions and that hypergraphs are actually special, more constrained cases of general graph-based representations. We show that phenomena often attributed uniquely to hypergraphs, such as abrupt transitions, can also be reproduced with standard graph models, often even locally tree-like ones. Overall, we suggest that the perceived need for hypergraphs stems from misconceptions about the expressive power of graph-based models and advocate for a clearer distinction between multivariate interactions, parametrized by graphs, and the functions that define them.

Biography of the speaker https://skewed.de/lab/tiago.html#short-bio

[1] https://arxiv.org/abs/2602.16937

Figure source: https://skewed.de/lab/posts/higher-standards/diagram_func.png