Séminaire (Organisé par l’Equipe de recherche DI)


Universidad Carlos III, Madrid

Metrics for sequential data based on dynamic models

Jeudi 6 juin 2013 à 14h en salle RD134

Résumé :

In this talk I will present two metrics for sequential data and some results of their application in clustering and segmentation problems. Both metrics involve the use of Hidden Markov Models (HMM) to capture the dynamics of the sequences into probabilistic models. This way, each sequence is represented by a sort of proxy probability density and thus similarities between pairs of sequences are computed as the divergences between the corresponding proxy densities.

The first metric, is based on learning several HMMs (up to one per sequence) and using as proxy pdfs the posterior probability of each model having generated the sequence. The second metric, called Space-State Dynamics, learns a single and more complex (in terms of hidden states) HMM and uses as proxy the transition matrix induced by each single sequence. The talk concludes with some experimental results in UCI-like problems that give some insight about the preformances of these two metrics in comparison with alternatives found in the literature.

PDF - 851 ko


FR SHIC 3272

Collegium UTC/CNRS