Machine learning for nonparametric unsupervised fraud detection

Domingues, Rémi
Research report

The Dirichlet Process Mixture Model algorithm presented here aggregates the variational
inference method presented by Bishop in [1], the use of a Beta prior on
the Dirichlet process responsible for the mixing proportions in [2] and the use of a
Gamma prior on the concentration parameter of the Dirichlet process proposed by
[3].
The current variational inference algorithm approximates the posterior distribution
of the dataset by a mixture of multivariate Gaussians, inferring the mixing
proportions from a stick-breaking process which concentration is inferred from a
Gamma distribution.

Type:
Rapport
Date:
2018-09-26
Department:
Data Science
Eurecom Ref:
5694
Copyright:
© EURECOM. Personal use of this material is permitted. The definitive version of this paper was published in Research report and is available at :
See also:

PERMALINK : https://www.eurecom.fr/publication/5694