Lecture
LCN2 seminar November 2024
- Date
- Friday 29 November 2024
- Time
- Address
- Gorlaeus
- Room
- BM. 2.26 (second floor)
70th LCN2 seminar
Speaker: Riccardo Milocco (IMT Lucca)
Title: Multi-Scale Node Embeddings for Graph Modeling and Generation
Abstract:
Networks have successfully represented a variety of complex systems, from economic interactions to ecosystem activities. Yet, in general, they do not provide a natural "embedding space'', from which a set of node coordinates can be extracted.
Since machine-learning applications, e.g. link prediction, require node features as inputs, node-embedding methods have become increasingly important to generate these attributes. However, typical embedding schemes regard the network only at the scale where the data were tracked down. If nodes were merged into a community, then the embedding for the community would have to be recomputed; there is no way to obtain it from the embeddings of its members. To address this limitation, we enhanced the Multi-Scale Model by equipping it with node embeddings. We tested our new multi-scale embedding model against a single-scale model using two multi-scale datasets: the ING Bank Input-Output Network and the World Trade Web. The results demonstrate that single-scale models are affected by "single-scale overfitting'', providing descriptions of economic phenomena that are not generalizable at coarser scales. In contrast, the multi-scale model remains self-consistent at every lower-resolution level, hereby offering a more comprehensive and robust representation of the economic system. Although we will focus on the problem of network modeling, the problem of robust embeddings is also crucial for many other applications such as link prediction, community detection, and network visualization.
Snacks and drinks will be available after the seminar at the Fusiebar.