Active Semi-Supervised Learning for Diffusions on Graphs

Active Semi-Supervised Learning for Diffusions on Graphs

Das, Bishwadeep (Student TU Delft)
Isufi, E. (TU Delft Multimedia Computing)
Leus, G.J.T. (TU Delft Signal Processing Systems)

2020

Diffusion-based semi-supervised learning on graphs consists of diffusing labeled information of a few nodes to infer the labels on the remaining ones. The performance of these methods heavily relies on the initial labeled set, which is either generated randomly or using heuristics. The first sometimes leads to unsatisfactory results because random labeling has no guarantees to label all classes while heuristic methods only yield a good performance when multiple recursive training stages are possible. In this paper, we put forth a new paradigm for one-shot active semi-supervised learning for graph diffusions. We rephrase active learning as the problem of selecting the output labels from a label propagation model. Subsequently, we develop two methods to solve this problem and label the nodes. The first method assumes there are only a few starting labels and relies on projected compressive sensing to build the label set. The second method drops the assumption of a few starting labels and builds on sparse sensing techniques to label a few nodes. Both methods have solid mathematical grounds in signal processing and require a single training phase. Numerical results on three scenarios corroborate our findings and showcase the improved performance compared with the state of the art.

Active learning
compressed sensing
diffusion on graphs
random walks
semi-supervised learning
sparse sensing

http://resolver.tudelft.nl/uuid:1204e0db-d655-4bd5-a5ee-51384df249e3

https://doi.org/10.1109/ICASSP40776.2020.9054300

IEEE

2020-11-14

978-1-5090-6632-2

ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

ICASSP 2020, 2020-05-04 → 2020-05-08, Barcelona, Spain

Green Open Access added to TU Delft Institutional Repository ‘You share, we take care!’ – Taverne project https://www.openaccess.nl/en/you-share-we-take-care Otherwise as indicated in the copyright section: the publisher is the copyright holder of this work and the author uses the Dutch legislation to make this work public.

Institutional Repository

conference paper

© 2020 Bishwadeep Das, E. Isufi, G.J.T. Leus