Éditeur(s) : EUTIC : le réseau international et interdisciplinaire pour les Enjeux et Usages des Technologies de l'Information et de la Communication EUTIC : le réseau international et interdisciplinaire pour les Enjeux et Usages des Technologies de l'Information et de la Communication
Extrait de : "Les écosystèmes numériques et la démocratisation informationnelle" : colloque, les 3 et 4 novembre 2015. Université des Antilles
Description : Dans cette communication, Dominique Cardon, sociologue au Laboratoire des usages d'Orange Labs et professeur associé à l'université de Marne la Vallée, propose une réflexion sur les propriétés organisationnelles, et notamment algorithmiques, de l'espace public numérique. Comment les calculateurs du web produisent-ils la visibilité numérique ? A partir de quels principes le PageRank de Google, les métriques du web social ou les outils de recommandation décident-ils de donner la prééminence à telle information plutôt qu'à telle autre ? On cherchera notamment à tirer des conséquences des transformations des épistémologies statistiques mises en oeuvre par les algorithmes du web avec la généralisation des techniques d'apprentissage (machine learning). Le web que nous connaissons ne serait pas le même s'il n'avait pas été profondément nourri, domestiqué et organisé par l'algorithme qui le classe, le PageRank de Google. C'est parce que ces documents s'adressent entre eux des liens de reconnaissance que la métrique du PageRank a pu construire une représentation de l'information proche de celles qui prévalaient dans l'espace public traditionnel. Cependant, d'autres formes de métriques de l'information se sont aujourd'hui généralisées avec les compteurs réputationnels du web social et les techniques de recommandation personnalisées. Aussi est-il utile de dégager les différentes formes du calcul qui sont aujourd'hui au principe de la construction des métriques du web afin de proposer une lecture de la manière dont elles enferment des représentations différentes des individus et de leur place dans nos sociétés.

Droits : CC-BY-NC-ND - Attribution - Pas d'utilisation commerciale - Pas de modification
Permalien : http://www.manioc.org/fichiers/V15436
V15436
V15436

Éditeur(s) : HAL CCSD
Résumé : 6 pages, 4 figures. In ICML 2013 Workshop on Machine Learning for Bioacoustics
International audience
The work presented in this paper is part of a global framework which long term goal is to design a wireless sensor network able to support the observation of a population of endangered birds. We present the first stage for which we have conducted a knowledge discovery approach on a sample of acoustical data. We use MFCC features extracted from bird songs and we exploit two knowledge discovery techniques. One that relies on clustering-based approaches, that highlights the homogeneity in the songs of the species. The other, based on predictive modeling, that demonstrates the good performances of various machine learning techniques for the identification process. The knowledge elicited provides promising results to consider a widespread study and to elicit guidelines for designing a first version of the automatic approach for data collection based on acoustic sensors.
ICML 2013 Workshop on Machine Learning for Bioacoustics
Atlanta, United States

hal-00840741
https://hal.archives-ouvertes.fr/hal-00840741
ARXIV : 1306.5349

Éditeur(s) : HAL CCSD Elsevier
Résumé : International audience
Voting rules relying on k-nearest neighbors (k-NN) are an effective tool in countless many machine learning techniques. Thanks to its simplicity, k-NN classification is very attractive to practitioners, as it enables very good performances in several practical applications. However, it suffers from various drawbacks, like sensitivity to "noisy" instances and poor generalization properties when dealing with sparse high-dimensional data. In this paper, we tackle the k-NN classification problem at its core by providing a novel k-NN boosting approach. Namely, we propose a supervised learning algorithm, called Universal Nearest Neighbors (UNN), that induces a leveraged k-NN rule by globally minimizing a surrogate risk upper bounding the empirical misclassification rate over training data. Interestingly, this surrogate risk can be arbitrary chosen from a class of Bregman loss functions, including the familiar exponential, logistic and squared losses. Furthermore, we show that UNN allows to efficiently filter a dataset of instances by keeping only a small fraction of data. Experimental results on the synthetic Ripley's dataset show that such a filtering strategy is able to reject "noisy" examples, and yields a classification error close to the optimal Bayes error. Experiments on standard UCI datasets show significant improvements over the current state of the art.
ISSN: 0925-2312

hal-00664462
https://hal.inria.fr/hal-00664462

Éditeur(s) : HAL CCSD Elsevier
Résumé : International audience
In this paper, we propose a benchmarking of supervised machine learning techniques (neural networks, Gaussian processes and support vector machines) in order to forecast the Global Horizontal solar Irradiance (GHI). We also include in this benchmark a simple linear autoregressive (AR) model as well as two naive models based on persistence of the GHI and persistence of the clear sky index (denoted herein scaled persistence model). The models are calibrated and validated with data from three French islands: Corsica (41.91°N; 8.73°E), Guadeloupe (16.26°N; 61.51°W) and Reunion (21.34°S ; 55.49°E). The main findings of this work are, that for hour ahead solar forecasting, the machine learning techniques slightly improve the performances exhibited by the linear AR and the scaled persistence model. However, the improvement appears to be more pronounced in case of unstable sky conditions. These nonlinear techniques start to outperform their simple counterparts for forecasting horizons greater than one hour.
ISSN: 0038-092X

Droits : info:eu-repo/semantics/OpenAccess
hal-01101564
https://hal.archives-ouvertes.fr/hal-01101564
https://hal.archives-ouvertes.fr/hal-01101564/document
https://hal.archives-ouvertes.fr/hal-01101564/file/Manuscript_Lauretetal_Rev_unmarked.pdf
DOI : 10.1016/j.solener.2014.12.014

Éditeur(s) : HAL CCSD
Résumé : International audience
In this work, we have led an analysis of the error of different global solar radiation prediction models according to the global solar radiation variability. Different predictions models where performed such as machine learning techniques (Neural Networks, Gaussian processes and support vector machines) in order to forecast the Global Horizontal solar Irradiance (GHI). We also include in this study a simple linear autoregressive (AR) model as well as two naive models based on persistence of the GHI and persistence of the clear sky index (denoted herein scaled persistence model). The models are calibrated and tested with data from three 3 French islands: Corsica (42.15°N ; 9.08°E), Guadeloupe (16.25°N ; 61.58°W) and Reunion (21.15°S ; 55.5°E). Guadeloupe and Reunion are located in a subtropical climatic zone whereas Corsica is in a tempered climatic zone. Hence the global solar radiation variation differs significantly. The output error of the different models was quantified by the nRSME. In order to quantify the influence of the global solar radiation variability on the forecasting models error we performed a classification of typical days according to different typical days. Each class of typical day is defined by a variation of global solar radiation rate. For each class and for each location, the selected forecasting models where performed and the error was quantified. With this analysis a global solar radiation forecasting models can be selected according to the location, the global solar radiation fluctuations and hence the meteorological conditions. INTRODUCTION Large and frequent variations of solar radiation can be observed in tropical climates with amplitudes reaching 800 W/m² and occurring within a short time interval, from few seconds to few minutes, according to the geographical location. Such fluctuations can be due for example to the dynamic of clouds which can be very complex and depend on cloud type, size, speed and spatial distribution and, more generally, due to some specific local meteorological conditions. Thus, the solar energy forecasting, a process used to predict the amount of solar energy available in the current and near terms, might be a difficult task. Some of the best predictors found in literature are Autoregressive and moving average (ARMA) [5,7,8], Bayesian inferences [9,10], Markov chains [11], k-Nearest-Neighbors predictors [12] or artificial intelligence techniques as the Artificial Neural Network (ANN) [9-11]. Although these methodologies are potentially good in many areas, we observed in our previous studies on global radiation prediction [9,13,14] that the simple model based on the persistence of the clear sky index gives often very good results with acceptable errors [15] for short term forecasting time horizon (<= 1 hour). The goal of this paper is to determinate the influence of solar radiation variability regarding different classes of days on the expected error provided by different forecasting methods that the modeller can possibly implement. The paper is organized as follow: Section 2 describes the data we have used. Section 3 exposes the classification methodology and the results obtained for the three studied locations. In the two following sections, the forecasting methods are exposed and then 3 the errors on the forecasting results for each location and for each class are exposed.
The Third Southern African Solar Energy Conference (SASEC2015)
Kruger National Park, South Africa

Droits : info:eu-repo/semantics/OpenAccess
hal-01099487
https://hal.archives-ouvertes.fr/hal-01099487
https://hal.archives-ouvertes.fr/hal-01099487/document
https://hal.archives-ouvertes.fr/hal-01099487/file/article%20SASEC.pdf

Éditeur(s) : HAL CCSD Elsevier
Résumé : International audience
In this paper, we propose a benchmarking of supervised machine learning techniques (neural networks, Gaussian processes and support vector machines) in order to forecast the Global Horizontal solar Irradiance (GHI). We also include in this benchmark a simple linear autoregressive (AR) model as well as two naive models based on persistence of the GHI and persistence of the clear sky index (denoted herein scaled persistence model). The models are calibrated and validatedwith data from three French islands: Corsica (41.91°N; 8.73°E), Guadeloupe (16.26°N; 61.51°W) and Reunion (21.34°S ; 55.49°E). The main findings of this work are, that for hour ahead solar forecasting,the machine learning techniques slightly improve the performances exhibited by the linear AR and the scaled persistence model. However, the improvement appears to be more pronounced in case of unstable sky conditions.These nonlinear techniques start to outperform their simple counterparts for forecasting horizons greater than one hour.
ISSN: 0038-092X

Droits : info:eu-repo/semantics/OpenAccess
hal-01099211
https://hal.archives-ouvertes.fr/hal-01099211
https://hal.archives-ouvertes.fr/hal-01099211/document
https://hal.archives-ouvertes.fr/hal-01099211/file/Manuscript_Lauretetal_Rev_unmarked.pdf

Éditeur(s) : HAL CCSD Elsevier
Résumé : International audience
These last years, many analysis methods have been proposed to extract knowledge from social networks. As for the traditional data mining domain, these network-based approaches can be classified according to two main families. The approaches based on predictive modelling, which encompass the techniques that analyse current and historical facts to make predictive assumptions about future or unknown events. The approaches based on descriptive modelling, which cover the set of techniques that aim to summarize the data by identifying some relevant features in order to describe how things organize and actually work. In this paper, we review the main descriptive modelling methods of social networks and show for each of them the resulting useful knowledge on a running example. We particularly emphasize on the most recent methods that combine information available on both the network structure and the node attributes in order to provide original description models taking into account the context.
EISSN: 1877-0509

hal-01312383
https://hal.univ-antilles.fr/hal-01312383
DOI : 10.1016/j.procs.2015.05.505

Éditeur(s) : HAL CCSD
Résumé : International audience
Modelling techniques for river hydrologic forecasting systems have taken advantage of machine learning methods especially for flood prediction. But current solutions, mostly based on artificial neural networks do not always meet end users requirements on the readability and the understandability of predictive models. In this paper, we present a new version of our original solution based on the concept of aggregate variables in order to predict flash flood events from observed water level and/or rain measurements, particularly in the context of caribbean watersheds in which flash flood are much uncertain. We combine aggregate variables in juries. Juries of aggregate variables are trained and tested using a typical 10-fold cross validation scheme. Best juries are searched through an evolutionary approach that is optimized. Different parameters are set up like aggregation periods and jury sizes to prove the efficiency of the proposed approach compared to classical solutions.
Congress on Evolutionary Computation (CEC) 2013
Cancun, Mexico

hal-00847582
https://hal.archives-ouvertes.fr/hal-00847582
DOI : 10.1109/CEC.2013.6557656