Section: New Results

Semi and non-parametric methods

Robust estimation for extremes

Participants : Clement Albert, Stephane Girard.

Joint work with: M. Stehlik (Johannes Kepler Universitat Linz, Austria and Universidad de Valparaiso, Chile) and A. Dutfoy (EDF R&D).

In the PhD thesis of Clément Albert (funded by EDF), we study the sensitivity of extreme-value methods to small changes in the data [46]. To reduce this sensitivity, robust methods are needed and, in [21], we proposed a novel method of heavy tails estimation based on a transformed score (the t-score). Based on a new score moment method, we derive the t-Hill estimator, which estimates the extreme value index of a distribution function with regularly varying tail. t-Hill estimator is distribution sensitive, thus it differs in e.g. Pareto and log-gamma case. Here, we study both forms of the estimator, i.e. t-Hill and t-lgHill. For both estimators we prove weak consistency in moving average settings as well as the asymptotic normality of t-lgHill estimator in the i.i.d. setting. In cases of contamination with heavier tails than the tail of original sample, t-Hill outperforms several robust tail estimators, especially in small sample situations. A simulation study emphasizes the fact that the level of contamination is playing a crucial role. We illustrate the developed methodology on a small sample data set of stake measurements from Guanaco glacier in Chile. This methodology is adapted to bounded distribution tails in [26] with an application to extreme snow loads in Slovakia.

Conditional extremal events

Participant : Stephane Girard.

Joint work with: L. Gardes (Univ. Strasbourg) and J. Elmethni (Univ. Paris 5)

The goal of the PhD theses of Alexandre Lekina and Jonathan El Methni was to contribute to the development of theoretical and algorithmic models to tackle conditional extreme value analysis, ie the situation where some covariate information $X$ is recorded simultaneously with a quantity of interest $Y$. In such a case, the tail heaviness of $Y$ depends on $X$, and thus the tail index as well as the extreme quantiles are also functions of the covariate. We combine nonparametric smoothing techniques  [77] with extreme-value methods in order to obtain efficient estimators of the conditional tail index and conditional extreme quantiles. When the covariate is functional and random (random design) we focus on kernel methods [18].

Conditional extremes are studied in climatology where one is interested in how climate change over years might affect extreme temperatures or rainfalls. In this case, the covariate is univariate (time). Bivariate examples include the study of extreme rainfalls as a function of the geographical location. The application part of the study is joint work with the LTHE (Laboratoire d'étude des Transferts en Hydrologie et Environnement) located in Grenoble [31], [32].

Estimation of extreme risk measures

Participant : Stephane Girard.

Joint work with: A. Daouia (Univ. Toulouse), L. Gardes (Univ. Strasbourg) and G. Stupfler (Univ. Aix-Marseille).

One of the most popular risk measures is the Value-at-Risk (VaR) introduced in the 1990's. In statistical terms, the VaR at level $\alpha \in (0,1)$ corresponds to the upper $\alpha$-quantile of the loss distribution. The Value-at-Risk however suffers from several weaknesses. First, it provides us only with a pointwise information: VaR($\alpha$) does not take into consideration what the loss will be beyond this quantile. Second, random loss variables with light-tailed distributions or heavy-tailed distributions may have the same Value-at-Risk. Finally, Value-at-Risk is not a coherent risk measure since it is not subadditive in general. A first coherent alternative risk measure is the Conditional Tail Expectation (CTE), also known as Tail-Value-at-Risk, Tail Conditional Expectation or Expected Shortfall in case of a continuous loss distribution. The CTE is defined as the expected loss given that the loss lies above the upper $\alpha$-quantile of the loss distribution. This risk measure thus takes into account the whole information contained in the upper tail of the distribution. In [64], we investigate the extreme properties of a new risk measure (called the Conditional Tail Moment) which encompasses various risk measures, such as the CTE, as particular cases. We study the situation where some covariate information is available under some general conditions on the distribution tail. We thus has to deal with conditional extremes (see paragraph 7.2.2).

A second possible coherent alternative risk measure is based on expectiles [63]. Compared to quantiles, the family of expectiles is based on squared rather than absolute error loss minimization. The flexibility and virtues of these least squares analogues of quantiles are now well established in actuarial science, econometrics and statistical finance. Both quantiles and expectiles were embedded in the more general class of M-quantiles as the minimizers of a generic asymmetric convex loss function. It has been proved very recently that the only M-quantiles that are coherent risk measures are the expectiles.

Multivariate extremal events

Participants : Stephane Girard, Florence Forbes.

Joint work with: F. Durante (Univ. Bolzen-Bolzano, Italy) and G. Mazo (Univ. Catholique de Louvain, Belgique).

Copulas are a useful tool to model multivariate distributions  [83]. However, while there exist various families of bivariate copulas, much fewer has been done when the dimension is higher. To this aim an interesting class of copulas based on products of transformed copulas has been proposed in the literature. The use of this class for practical high dimensional problems remains challenging. Constraints on the parameters and the product form render inference, and in particular the likelihood computation, difficult. As an alternative, we proposed a new class of copulas constructed by introducing a latent factor. Conditional independence with respect to this factor and the use of a nonparametric class of bivariate copulas lead to interesting properties like explicitness, flexibility and parsimony. In particular, various tail behaviours are exhibited, making possible the modeling of various extreme situations [17], [22].

Level sets estimation

Participant : Stephane Girard.

Joint work with: G. Stupfler (Univ. Aix-Marseille).

The boundary bounding the set of points is viewed as the larger level set of the points distribution. This is then an extreme quantile curve estimation problem. We proposed estimators based on projection as well as on kernel regression methods applied on the extreme values set, for particular set of points [10]. We also investigate the asymptotic properties of existing estimators when used in extreme situations. For instance, we have established in collaboration with G. Stupfler that the so-called geometric quantiles have very counter-intuitive properties in such situations [20] and thus should not be used to detect outliers.

Robust Sliced Inverse Regression.

Participants : Stephane Girard, Alessandro Chiancone, Florence Forbes.

This research theme was supported by a LabEx PERSYVAL-Lab project-team grant.

Sliced Inverse Regression (SIR) has been extensively used to reduce the dimension of the predictor space before performing regression. Recently it has been shown that this technique is, not surprisingly, sensitive to noise. Different approaches have thus been proposed to robustify SIR. In [14], we start considering an inverse problem proposed by R.D. Cook and we show that the framework can be extended to take into account a non-Gaussian noise. Generalized Student distributions are considered and all parameters are estimated via an EM algorithm. The algorithm is outlined and tested comparing the results with different approaches on simulated data. Results on a real dataset show the interest of this technique in presence of outliers.

Collaborative Sliced Inverse Regression.

Participants : Stephane Girard, Alessandro Chiancone.

This research theme was supported by a LabEx PERSYVAL-Lab project-team grant.

Joint work with: J. Chanussot (Gipsa-lab and Grenoble-INP).

In his PhD thesis work, Alessandro Chiancone studies the extension of the SIR method to different sub-populations. The idea is to assume that the dimension reduction subspace may not be the same for different clusters of the data [15]. One of the difficulty is that standard Sliced Inverse Regression (SIR) has requirements on the distribution of the predictors that are hard to check since they depend on unobserved variables. It has been shown that, if the distribution of the predictors is elliptical, then these requirements are satisfied. In case of mixture models, the ellipticity is violated and in addition there is no assurance of a single underlying regression model among the different components. Our approach clusterizes the predictors space to force the condition to hold on each cluster and includes a merging technique to look for different underlying models in the data. A study on simulated data as well as two real applications are provided. It appears that SIR, unsurprisingly, is not able to deal with a mixture of Gaussians involving different underlying models whereas our approach is able to correctly investigate the mixture.

Hapke's model parameter estimation from photometric measurements

Participants : Florence Forbes, Emeline Perthame.

Joint work with: Sylvain Douté (IPAG, Grenoble).

The Hapke's model is a widely used analytical model in planetology to describe the spectro-photometry of granular materials. It is a non linear model $F$ that links a set of parameters $x$ to a "theoretical" Bidirectional Reflectance Diffusion Function (BRDF). In practice, we assume that the observed BRDF $Y$ is a noisy version of the "theoretical" one

where $ϵ$ is a centered Gaussian noise with diagonal covariance matrix $\Sigma$. Then $x$ is also assumed to be random with some prior distribution to be specified, e.g. uniform on the parameters range in [84]. The overall goal is to estimate the posterior distribution $p\left(x\right|y)$ for some observed BRDF $y$. Equation (5) defines the likelihood of the model which is $p\left(y\right|x)=𝒩(y;F\left(x\right),\Sigma )$. Then since $F$ is non linear, it is not possible to obtain an analytical expression for $p\left(x\right|y)$. However, it is easy to simulate parameters $x$ that follows the posterior distribution $p\left(x\right|y)\propto p(y\left|x\right)p\left(x\right)$ for instance using MCMC techniques [84]. If only point estimate are desired, the MAP can be used and evolutionary algorithms can then be used also using $p\left(y\right|x\left)p\right(x)$ as a fitness function. But obtaining such simulations is time consuming and has to be done for each observed value of $y$. In this work, we propose to use a locally linear mapping approximation and an inverse regression strategy to provide an analytical expression of $p\left(x\right|y)$. The idea is that the non linear $F$ can be approximated by a number $K$ of locally linear functions and that each of this function is easy to inverse. It follows that the inverse of $F$ is also approximated as locally linear. Preliminary results were presented at the MultiPlaNet workshop in Orsay, December 14, 2016. They show that the proposed method does not fully reproduce the previous results obtained using MCMC techniques. Further investigations are required to understand the origin of the difference. Also ABC (approximate Bayes computation) methods will be considered as a subsequent step that may improved the current procedure while remaining computationally efficient.

Prediction intervals for inverse regression models in high dimension

Participant : Emeline Perthame.

Joint work with: Emilie Devijver (KU Leuven, Belgium).

Inverse regression, as a dimension reduction technique, is a reliable and efficient approach to handle large regression issues in high dimension, when the number of features exceeds the number of observations. Indeed, under some conditions, dealing with the inverse regression problem associated to a forward regression problem drastically reduces the number of parameters to estimate and make the problem tractable. However, regression models are often used to predict a new response from a new observed profile of covariates, and we may be interested in deriving confidence bands for the prediction to quantify the uncertainty around a predicted response. Theoretical results have already been derived for the well-known linear model, but recently, the curse of dimensionality has increased the interest of practitioners and theoreticians into generalization of those results on a high-dimension context. When both the responses and the covariates are multivariate, we derive in this work theoretical prediction bands for the inverse regression linear model and propose an analytical expression of these intervals. The feasibility, the confidence level and the accuracy of the proposed intervals are also analyzed through a simulation study.

Multi sensor fusion for acoustic surveillance and monitoring

Participants : Florence Forbes, Jean-Michel Becu.

Joint work with: Pascal Vouagner and Christophe Thirard from ACOEM company.

In the context of the DGA-rapid WIFUZ project with the ACOEM company, we addressed the issue of determining the localization of shots from multiple measurements coming from multiple sensors. We used Bayesian inversion and simulation techniques to recover multiple sources mimicking collaborative interaction between several vehicles. This project is at the intersection of data fusion, statistics, machine learning and acoustic signal processing. The general context is the surveillance and monitoring of a zone acoustic state from data acquired at a continuous rate by a set of sensors that are potentially mobile and of different nature. The overall objective is to develop a prototype for surveillance and monitoring that is able to combine multi sensor data coming from acoustic sensors (microphones and antennas) and optical sensors (infrared cameras) and to distribute the processing to multiple algorithmic blocs.