Skip to Main content Skip to Navigation

Estimation de modèles autorégressifs vectoriels à noyaux à valeur opérateur: application à l'inférence de réseaux

Abstract : In multivariate time series analysis, existing models are often used for forecasting, i.e. estimating future values of the observed system based on previously observed values. Another purpose is to find causal relationships among a set of state variables within a dynamical system. We focus on the latter and develop tools in order to address this problem. In this thesis, we define a new family of nonparametric vector autoregressive models based on operator-valued kernels. Assuming a sparse underlying structure, we control the model's sparsity by defining a loss function that includes sparsity-inducing penalties on the model parameters (which are basis vectors within a linear combination of kernels). The selected kernels sometimes involve hyperparameters that may need to be learned depending on the nature of the problem. On the one hand, when expert knowledge or working assumptions allow presetting the parameters of the kernel, the learning problem boils down to estimating only the model parameters. To optimize the corresponding loss function, we develop a proximal algorithm. On the other hand, when no prior knowledge is available, some other kernels may exhibit unknown parameters. Consequently, this leads to the joint learning of the kernel parameters in addition to the model parameters. We thus resort to an alternate optimization scheme which involves proximal methods. Subsequently, we propose to build an estimate of the adjacency matrix coding for the underlying causal network by computing a function of the instantaneous Jacobian matrices. In a high-dimensional setting, i.e. insufficient amount of data compared to the number of variables, we design an ensemble methodology that shares features of boosting and random forests. In order to emphasize the performance of the developed models, we apply them on two tracks: simulated data from gene regulatory networks and real climate data.
Complete list of metadatas

Cited literature [162 references]  Display  Hide  Download
Contributor : Néhémy Lim <>
Submitted on : Tuesday, April 14, 2015 - 1:08:10 AM
Last modification on : Tuesday, June 30, 2020 - 11:56:08 AM
Long-term archiving on: : Tuesday, April 18, 2017 - 6:15:16 PM


  • HAL Id : tel-01141855, version 1


Néhémy Lim. Estimation de modèles autorégressifs vectoriels à noyaux à valeur opérateur: application à l'inférence de réseaux. Apprentissage [cs.LG]. Université d'Evry Val-d'Essonne, 2015. Français. ⟨tel-01141855⟩



Record views


Files downloads