Bereiche | Tage | Auswahl | Suche | Downloads | Hilfe

AKSOE: Arbeitskreis Physik sozio-ökonomischer Systeme

AKSOE 15: Social-, Information-, and Production Networks II

AKSOE 15.4: Vortrag

Donnerstag, 28. Februar 2008, 11:45–12:15, EW 203

Limits of Unsupervised Learning in Networks — •Jörg Reichardt¹ and Michele Leone² — ¹Institute f. Theoretical Physics, University of Würzburg — ²ISI Foundation, Torino, Italy

Many systems in socio- and econophysics are abstracted as networks. Before we can build models for such systems, a careful data analysis is needed in order to select relevant features. The goal is to differentiate between those effects that arise from inherent randomness in the system and those that truly reflect structure in the data. Unsupervised learning algorithms can perform this task in an automated manner and the general experience from multi-variate data is that if the data set is only large enough, even the slightest deviation from randomness may be detected. The talk will show that this is not necessarily true for sparse networks. Even in the limit of infinite system size, sparse networks may not be differentiated from random networks despite them being generated by a non-random process. Equivalently, the fact that one cannot find deviations from randomness may not allow to rule out non-random data generating processes. The talk will discuss possible implications for the analysis of network data and limitations in our ability to forecast the evolution of the system.