Event
Generating Doppelganger Graphs: Preserving both Privacy and Fine-Grained Data
- 20 December 2024
- Expired!
- 3:00 pm - 4:00 pm
Location
- Attendance: on site
- Language: EN
Event
Generating Doppelganger Graphs: Preserving both Privacy and Fine-Grained Data
In an era of data-driven insights, preserving privacy in sensitive networks (like financial ones) poses critical challenges. Even with appropriate agreements with companies and institutions, the data available for research activities is coarse-grained and therefore it is not possible to exploit the full potential of Big Data.
Is it then possible to provide companies and institutions with models that by themselves can learn on the original data and output fine-grained synthetic networks?
This work addresses indeed the generation, through Deep Learning models, of Doppelganger Graphs: synthetic, privacy-preserving graphs that retain node attributes and topology of real-world networks.
To do so, the pipeline includes a novel node embedding model called Gravity GraphSAGE, which optimizes directed link prediction, and a WGAN-GP (Wasserstein Generative Adversarial Network with Gradient Penalty) as a synthetic node generation model .
Benchmarking and Current Pitfalls:
Evaluations on benchmark datasets confirm state-of-the-art performance for our Gravity GraphSAGE model, but we will discuss the current pitfalls of Machine Learning literature on Link Prediction, especially when dealing with sparse graphs. Proposals for future research are shown.