Search

In Sub-Saharan Africa, smallholder farms play a key role in agriculture, occupying most of the agricultural land. Design policies for increasing smallholder productivity remains a safe way to establish sustainable food systems and boost local economies. However, efforts are still needed in order to achieve accurate and timely monitoring in smallholder farming systems. With the advent of modern Earth Observation programmes such as the Sentinel satellites, which provide quasi-synchronous and high-resolution multi-source information over any area of the continental surfaces, new opportunities are opened up to accurately map crop yields in smallholder farming systems. This study intends to estimate and forecast millet yields in central Senegal, making the use of multi-source (synthetic-aperture radar (SAR) and optical) image time series and state-of-the-art machine learning models. A Random Forest (RF) model explained up to 50% of the millet yield variability, while deep learning models such as Convolutional Neural Network (CNN) showed promise results but performed lower. We also found that the concatenation of SAR polarizations and vegetation indices improved our crop yield modelling, but such improvement was tightly related to the modelling approach, namely RF and CNN. Using RF to forecast millet yields, we achieved stable and satisfactory accuracy 2 weeks before the harvest period.

International audience; This letter proposes a deep learning model to deal with the spatial transfer challenge for the mapping of irrigated areas through the analysis of Sentinel-1 data. First, a convolutional neural network (CNN) model called "Teacher Model" is trained on a source geographical area characterized by a huge volume of samples. Then, this model is transferred from the source area to a target area characterized by a limited number of samples. The transfer learning framework is based on a distill and refine strategy in which the teacher model is firstly distilled into a student model and, successively, refined by data samples coming from the target geographical area. The proposed strategy is compared to different approaches including a random forest (RF) classifier trained on the target dataset, a CNN trained on the source dataset and directly applied on the target area as well as several CNN classifiers trained on the target dataset. The evaluation of the performed transfer strategy shows that the "distill and refine" framework obtains the best performance compared to other competing approaches. The obtained findings represent a first step towards the understanding of the spatial transferability of deep learning models in the Earth Observation domain.

International audience; Nowadays, huge amount of data are being produced by a large and diverse family of sensors (e.g., remote sensors, biochemical sensors, wearable devices). These sensors typically measure multiple variables over time, resulting in data streams that can be profitably organized as multivariate time-series. In practical scenarios, the speed at which such information is collected often makes the data labeling a difficult task. This results in a low-data regime scenario where only a small set of labeled samples is available and standard supervised learning algorithms cannot be employed. To cope with the task of multi-variate time series classification in low-data regime scenarios, here, we propose a framework that combines convolutional neural networks (CNNs) with self-training (pseudo labeling) in a transductive setting (test data are already available at training time). Our framework, named ResNetIPL, wraps a CNN based classifier into an iterative procedure that, at each step, enlarges the training set with new samples and their associated pseudo labels. An experimental evaluation on several benchmarks, coming from different domains, has demonstrated the value of the proposed approach and, more generally, the ability of the deep learning approaches to effectively deal with scenarios characterized by low-data regimes.

International audience; Nowadays, huge volume of satellite images, via the different Earth Observation missions, are constantly acquired and they constitute a valuable source of information for the analysis of spatio-temporal phenomena. However, it can be challenging to obtain reference data associated to such images to deal with land use or land cover changes as often the nature of the phenomena under study is not known a priori. With the aim to deal with satellite image analysis, considering a real-world scenario where reference data cannot be available, in this paper, we present a novel end-to-end unsupervised approach for change detection and clustering for satellite image time series (SITS). In the proposed framework, we firstly create bi-temporal change masks for every couple of consecutive images using neural network autoencoders. Then, we associate the extracted changes to different spatial objects. The objects sharing the same geographical location are combined in spatio-temporal evolution graphs that are finally clustered accordingly to the type of change process with gated recurrent unit (GRU) autoencoder-based model. The proposed approach was assessed on two real-world SITS data supplying promising results.

The need for reliable and exhaustive data on land use is a major issue in planning policies and monitoring of land take. In this work, we evaluate urban building classification resulting from deep learning based approaches using as input SPOT 6/7 satellite images (at a spatial resolution of 1.5m) and national databases. In addition to assessing the classifier behaviour on urban land cover, the objective here is to compare the deep learning results at both pixel-based and object-based scale to qualify the overall results. Standard evaluation metrics (such as F-score) have shown better scores at the object-based assessment level (median F-score = 0.78) than the pixel ones (median F-score = 0.68). Identifying the impact of object characteristics using the object-based level of analysis has also revealed that beyond a surface area of 100 m2, objects are much better detected (median F-score > 0.91). It is the same with a high urban density (median F-score > 0.97). The accuracy of the intersections evaluated from the Intersection over Union also follows these trends as the entities area increases.

International audience; Semi-supervised learning is crucial in many applications where accessing class labels is unaffordable or costly. The most promising approaches are graph-based but they are transductive and they do not provide a generalized model working on inductive scenarios. To address this problem, we propose a generic framework, ESA☆, for inductive semi-supervised learning based on three components: an ensemble of semi-supervised autoencoders providing a new data representation that leverages the knowledge supplied by the reduced amount of available labels; a graph-based step that helps augmenting the training set with pseudo-labeled instances and, finally, a classifier trained with labeled and pseudo-labeled instances. Additionally, we also introduce two variants of our framework adopting different graph-based pseudo-labeling strategies: the first, ESALP, is based on a confidence-aware label propagation algorithm, while the second, ESAGAT, on a graph convolutional attention network. The experimental results show that our framework outperforms state-of-the-art inductive semi-supervised methods.

International audience; Semi-supervised learning is a family of classification methods conceived to reduce the amount of required labeled information in the training phase. Graph-based methods are among the most popular semi-supervised strategies: a nearest neighbor graph is built in such a way that the manifold of the data is captured and the labeled information is propagated to target samples along the structure of the manifold. Research in graph-based semi-supervised learning has mainly focused on two aspects: i) the construction of the k-nearest neighbors graph and/or ii) the propagation algorithm providing the classification. Differently from the previous literature, in this paper we focus on data representation with the aim of incorporating semi-supervision earlier in the process. To this end, we propose an algorithm that learns a new knowledge-aware data embedding via an ensemble of semi-supervised autoencoders to enhance a graph-based semi-supervised classification. The experiments carried out on different classification tasks demonstrate the benefit of our approach.

International audience; Monitoring the ecological status of natural habitats is crucial to the conservation process, as it enables the implementation of efficient conservation policies. Nowadays, it is increasingly possible to automate species identification, given the availability of very large image databases and state-of-the-art computational power which makes the training of automated machine learning-based classification models an increasingly viable tool for monitoring marine habitats. Coralligenous reefs are an underwater habitat of particular importance, found in the Mediterranean. This habitat is of a similar biocomplexity to coral reefs. They have been monitored in French waters since 2010 using manually annotated photo quadrats (RECOR monitoring network). Based on the large database of annotations accumulated therein, we have trained convolutional neural networks to automatically recognise coralligenous species using the data gathered from photo quadrats. Previous studies conducted on similar habitats performed well, but were only able to consider a limited number of classes, resulting in a very coarse description of these often-complex habitats. We therefore designed a custom network based on off-theshelf architectures which is able to discriminate between 61 classes with 72.59% accuracy. Our results showed that confusion errors were for the most part taxonomically coherent, showing accuracy performances of 84.47% when the task was simplified to 15 major categories, thereby outperforming the human accuracy previously recorded in a similar study. In light of this, we built a semi-automated tool to reject unsure results and reduce error risk, for when a higher level of accuracy is required. Finally, we used our model to assess the biodiversity and ecological status of coralligenous reefs with the Coralligenous Assemblage Index (CAI) and the Shannon Index. Our results showed that whilst the prediction of the CAI was only moderately accurate (pearson correlation between observed and predicted CAI = 0.61), the prediction of Shannon Index was more accurate (pearson correlation = 0.74). In conclusion, it will be argued that the approach outlined by this study offers a cost and time-effective tool for the analysis of coralligenous assemblages which is suitable for integration into a large-scale monitoring network of this habitat.

International audience; The huge amount of data currently produced by modern Earth Observation (EO) missions has allowed for the design of advanced machine learning techniques able to support complex Land Use/Land Cover (LULC) mapping tasks. The Copernicus programme developed by the European Space Agency provides, with missions such as Sentinel-1 (S1) and Sentinel-2 (S2), radar and optical (multi-spectral) imagery, respectively, at 10 m spatial resolution with revisit time around 5 days. Such high temporal resolution allows to collect Satellite Image Time Series (SITS) that support a plethora of Earth surface monitoring tasks. How to effectively combine the complementary information provided by such sensors remains an open problem in the remote sensing field. In this work, we propose a deep learning architecture to combine information coming from S1 and S2 time series, namely TWINNS (TWIn Neural Networks for Sentinel data), able to discover spatial and temporal dependencies in both types of SITS. The proposed architecture is devised to boost the land cover classification task by leveraging two levels of complementarity, i.e., the interplay between radar and optical SITS as well as the synergy between spatial and temporal dependencies. Experiments carried out on two study sites characterized by different land cover characteristics (i.e., the Koumbia site in Burkina Faso and Reunion Island, a overseas department of France in the Indian Ocean), demonstrate the significance of our proposal.

International audience; A major issue affecting optical imagery is the presence of clouds. The need of cloud-free scenes at specific date is crucial in a number of operational monitoring applications. On the other hand, the cloud-insensitive SAR sensors are a solid asset and they provide orthogonal information with respect to optical satellite, that enable the retrieval of information lost in optical images due to cloud cover. In the context of an increasing availability of both optical and SAR images, thank to the Sentinel constellation, we propose a deep learning method to reconstruct (gap-fill) optical data, polluted by cloud phenomena, exploiting multi-temporal SAR and optical images.

[Departement_IRSTEA]Territoires [TR1_IRSTEA]SYNERGIE [Axe_IRSTEA]TETIS-SISO [Axe_IRSTEA]TETIS-ATTOS [ADD1_IRSTEA]Dynamiques spatiales d'anthropisation; International audience; Nowadays, modern Earth Observation systems continuously generate huge amounts of data. A notable example is represented by the Sentinel-2 mission, which provides images at high spatial resolution (up to 10m) with high temporal revisit period (every 5 days), which can be organized in Satellite Image Time Series (SITS). While the use of SITS has been proved to be beneficial in the context of Land Use/Land Cover (LULC) map generation, unfortunately, machine learning approaches commonly leveraged in remote sensing field fail to take advantage of spatio-temporal dependencies present in such data. Recently, new generation deep learning methods allowed to significantly advance research in this field. These approaches have generally focused on a single type of neural network, i.e., Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs), which model different but complementary information: spatial autocorrelation (CNNs) and temporal dependencies (RNNs). In this work, we propose the first deep learning architecture for the analysis of SITS data, namely \method{} (DUal view Point deep Learning architecture for time series classificatiOn), that combines Convolutional and Recurrent neural networks to exploit their complementarity. Our hypothesis is that, since CNNs and RNNs capture different aspects of the data, a combination of both models would produce a more diverse and complete representation of the information for the underlying land cover classification task. Experiments carried out on two study sites characterized by different land cover characteristics (i.e., the \textit{Gard} site in France and the \textit{Reunion Island} in the Indian Ocean), demonstrate the significance of our proposal.

International audience; In most real world scenarios, experts dispose of limited background knowledge that they can exploit for guiding the analysis process. In this context, semi-supervised clustering can be employed to leverage such knowledge and enable the discovery of clusters that meet the analysts’ expectations. To this end, we propose a semi-supervised deep embedding clustering algorithm that exploits triplet constraints as background knowledge within the whole learning process. The latter consists in a two-stage approach where, initially, a low-dimensional data embedding is computed and, successively, cluster assignment is refined via the introduction of an auxiliary target distribution. Our algorithm is evaluated on real-world benchmarks in comparison with state-of-the-art unsupervised and semi-supervised clustering methods. Experimental results highlight the quality of the proposed framework as well as the added value of the new learnt data representation.

International audience; In this work we introduce and evaluate a deep learning model, mbCNN, that combines together satellite imagery and Volunteer Geographical Information (VGI) data to deal with different types of built-up surfaces. Differently from most of the previous works that only consider Urban/Non-Urban settings involving only one urban LULC class, here, we investigate the possibility to go a step further and distinguish among several urban land use classes: Residential, industrial, sport fields and non-urban. Experiments on a real-world dataset covering the City of Montpellier (South of France) site are reported. Such results demonstrate the quality of Deep Learning approaches to deal with several types of Urban LULC mapping as well as the positive influence to integrate VGI knowledge in the process.

Radar and Optical Satellite Image Time Series (SITS) are sources of information that are commonly employed to monitor earth surfaces for tasks related to ecology, agriculture, mobility, land management planning and land cover monitoring. Many studies have been conducted using one of the two sources, but how to smartly combine the complementary information provided by radar and optical SITS is still an open challenge. In this context, we propose a new neural architecture for the combination of Sentinel-1 (S1) and Sentinel-2 (S2) imagery at object level, applied to a real-world land cover classification task. Experiments carried out on the Reunion Island, a overseas department of France in the Indian Ocean, demonstrate the significance of our proposal.

International audience; Many real world data can be represented by a network with a set of nodes linked each other by multiple relations. Such a rich graph is called multilayer graph. In this demo, we present a tool for Visual Querying of Large Multilayer Graphs that allows to visually draw the query, retrieve result patterns and finally navigate and browse the results considering the original multilayer graph database. Our approach does not only provide a graphical user interface for the graph engine but the query processing is fully integrated.

International audience; In most real world clustering scenarios, experts generally dispose of limited background information, but such knowledge is valuable and may guide the analysis process. Semi-supervised clustering can be used to drive the algorithmic process with prior knowledge and to enable the discovery of clusters that meet the analyst's expectations. Usually, in the semi-supervised clustering setting, the background knowledge is converted to some kind of constraint and, successively, metric learning or constrained clustering are adopted to obtain the final data partition. Conversely, we propose a new semi-supervised clustering algorithm that directly exploits prior knowledge, under the form of labeled examples, avoiding the necessity to derive constraints. Our algorithm employs a multiresolution strategy to generate an ensemble of semi-supervised autoencoders that fit the data together with the background knowledge. Successively, the network models are employed to supply a new embedding representation on which clustering is performed. The proposed strategy is evaluated on a set of real-world benchmarks also in comparison with well-known state-of-the-art semi-supervised clustering methods. The experimental results highlight the benefit of directly leveraging the prior knowledge and show the quality of the representation learnt by the multiresolution schema.

International audience; Mapping winter vegetation quality is a challenging problem in remote sensing. This is due to cloud coverage in winter periods, leading to a more intensive use of radar rather than optical images. The aim of this letter is to provide a better understanding of the capabilities of Sentinel-1 radar images for winter vegetation quality mapping through the use of deep learning techniques. Analysis is carried out on a multitemporal Sentinel-1 data over an area around Charentes-Maritimes, France. This data set was processed in order to produce an intensity radar data stack from October 2016 to February 2017. Two deep recurrent neural network (RNN)-based classifiers were employed. Our work revealed that the results of the proposed RNN models clearly outperformed classical machine learning approaches (support vector machine and random forest).

Land cover is a fundamental variable for regional planning, as well as for the study and understanding of the environment. This work propose a multi-temporal approach relying on a fusion of radar multi-sensor data and information collected by the latest sensor (Sentinel-1) with a view to obtaining better results than traditional image processing techniques. The Danube Delta is the site for this work. The spatial approach relies on new spatial analysis technologies and methodologies: Deep Learning of multi-temporal Sentinel-1. We propose a deep learning network for image classification which exploits the multi-temporal characteristic of Sentinel-1 data. The model we employ is a Gated Recurrent Unit (GRU) Network, a recurrent neural network that explicitly takes into account the time dimension via a gated mechanism to perform the final prediction. The main quality of the GRU network is its ability to consider only the important part of the information coming from the temporal data discarding the irrelevant information via a forgetting mechanism. We propose to use such network structure to classify a series of images Sentinel-1 (20 Sentinel-1 images acquired between 9.10.2014 and 01.04.2016). The results are compared with results of the classification of Random Forest.

Nowadays, modern earth observation programs produce huge volumes of satellite images time series (SITS) that can be useful to monitor geographical areas through time. How to efficiently analyze such kind of information is still an open question in the remote sensing field. Recently, deep learning methods proved suitable to deal with remote sensing data mainly for scene classification (i.e. Convolutional Neural Networks - CNNs - on single images) while only very few studies exist involving temporal deep learning approaches (i.e Recurrent Neural Networks - RNNs) to deal with remote sensing time series. In this letter we evaluate the ability of Recurrent Neural Networks, in particular the Long-Short Term Memory (LSTM) model, to perform land cover classification considering multi-temporal spatial data derived from a time series of satellite images. We carried out experiments on two different datasets considering both pixel-based and object-based classification. The obtained results show that Recurrent Neural Networks are competitive compared to state-of-the-art classifiers, and may outperform classical approaches in presence of low represented and/or highly mixed classes. We also show that using the alternative feature representation generated by LSTM can improve the performances of standard classifiers.

International audience; The problem of local community detection in graphs refers to the identification of a community that is specific to a query node and relies on limited information about the network structure. Existing approaches for this problem are defined to work in dynamic network scenarios, however they are not designed to deal with complex real-world networks, in which multiple types of connectivity might be considered. In this work, we fill this gap in the literature by introducing the first framework for local community detection in multilayer networks (ML-LCD). We formalize the ML-LCD optimization problem and provide three definitions of the associated objective function, which correspond to different ways to incorporate within-layer and across-layer topo-logical features. We also exploit our framework to generate multilayer global community structures. We conduct an extensive experimentation using seven real-world multilayer networks, which also includes comparison with state-of-the-art methods for single-layer local community detection and for multilayer global community detection. Results show the significance of our proposed methods in discovering local communities over multiple layers, and also highlight their ability in producing global community structures that are better in modularity than those produced by native global community detection approaches .

International audience; In this paper we introduce a new approach of semi-supervised anomaly detection that deals with categorical data. Given a training set of instances (all belonging to the normal class), we analyze the relationships among features for the extraction of a discriminative characterization of the anomalous instances. Our key idea is to build a model characterizing the features of the normal instances and then use a set of distance-based techniques for the discrimination between the normal and the anomalous instances. We compare our approach with the state-of-the-art methods for semi-supervised anomaly detection. We empirically show that a specifically designed technique for the management of the categorical data outperforms the general-purpose approaches. We also show that, in contrast with other approaches that are opaque because their decision cannot be easily understood, our proposal produces a discriminative model that can be easily interpreted and used for the exploration of the data.

The problem of node-centric, or local, community detection in information networks refers to the identification of a community for a given input node, having limited information about the network topology. Existing methods for solving this problem, however, are not conceived to work on complex networks. In this paper, we propose a novel framework for local community detection based on the multilayer network model. Our approach relies on the maximization of the ratio between the community internal connection density and the external connection density, according to multilayer similarity-based community relations. We also define a biasing scheme that allows the discovery of local communities characterized by different degrees of layer-coverage diversification. Experimental evaluation conducted on real-world multilayer networks has shown the significance of our approach., Accepted at 8th International Conference on Complex Networks (CompleNet'17)

Since the BOSS competition, in 2010, most steganalysis approaches use a learning methodology involving two steps: feature extraction, such as the Rich Models (RM), for the image representation, and use of the Ensemble Classifier (EC) for the learning step. In 2015, Qian et al. have shown that the use of a deep learning approach that jointly learns and computes the features, is very promising for the steganalysis. In this paper, we follow-up the study of Qian et al., and show that, due to intrinsic joint minimization, the results obtained from a Convolutional Neural Network (CNN) or a Fully Connected Neural Network (FNN), if well parameterized, surpass the conventional use of a RM with an EC. First, numerous experiments were conducted in order to find the best " shape " of the CNN. Second, experiments were carried out in the clairvoyant scenario in order to compare the CNN and FNN to an RM with an EC. The results show more than 16% reduction in the classification error with our CNN or FNN. Third, experiments were also performed in a cover-source mismatch setting. The results show that the CNN and FNN are naturally robust to the mismatch problem. In Addition to the experiments, we provide discussions on the internal mechanisms of a CNN, and weave links with some previously stated ideas, in order to understand the impressive results we obtained., IS&T. Media Watermarking, Security, and Forensics, Part of IS&T International Symposium on Electronic Imaging, EI'2016, Feb 2015, San Fransisco, United States

International audience; In this letter, we propose a new active transductive learning (ATL) framework for object-based classification of satellite images. The framework couples graph-based label propagation with active learning (AL) to exploit positive aspects of the two learning settings. The transductive approach considers both labelled and unlabelled image objects to perform its classification as they are all available at training time while the AL strategy smartly guides the construction of the training set employed by the learner. The proposed framework was tested in the context of a land cover classification task using RapidEye optical imagery. A reference land cover map was elaborated over the whole study area in order to get reliable information about the performance of the ATL framework. The experimental evaluation under- lines that, with a reasonable amount of training data, our framework outperforms state of the art classification methods usually employed in the field of remote sensing.

Clustering data described by categorical attributes is a challenging task in data mining applications. Unlike numerical attributes, it is difficult to define a distance between pairs of values of a categorical attribute, since the values are not ordered. In this article, we propose a framework to learn a context-based distance for categorical attributes. The key intuition of this work is that the distance between two values of a categorical attribute A i can be determined by the way in which the values of the other attributes A j are distributed in the dataset objects: if they are similarly distributed in the groups of objects in correspondence of the distinct values of A i a low value of distance is obtained. We propose also a solution to the critical point of the choice of the attributes A j . We validate our approach by embedding our distance learning framework in a hierarchical clustering algorithm. We applied it on various real world and synthetic datasets, both low and high-dimensional. Experimental results show that our method is competitive with respect to the state of the art of categorical data clustering approaches. We also show that our approach is scalable and has a low impact on the overall computational time of a clustering task.

In common binary classification scenarios, the presence of both positive and negative examples in training data is needed to build an efficient classifier. Unfortunately, in many domains, this requirement is not satisfied and only one class of examples is available. To cope with this setting, classification algorithms have been introduced that learn from Positive and Unlabeled (PU) data. Originally, these approaches were exploited in the context of document classification. Only few works address the PU problem for categorical datasets. Nevertheless, the available algorithms are mainly based on Naive Bayes classifiers. In this work we present a new distance based PU learning approach for categorical data: Pulce. Our framework takes advantage of the intrinsic relationships between attribute values and exceeds the independence assumption made by Naive Bayes. Pulce, in fact, leverages on the statistical properties of the data to learn a distance metric employed during the classification task. We extensively validate our approach over real world datasets and demonstrate that our strategy obtains statistically significant improvements w.r.t. state-of-the-art competitors. HighlightsWe propose a distance learning framework for PU learning with categorical data.Differently from existing methods for categorical data, our method is not based on the independence assumption.Our distance learning approach leverages attribute context.The experiments show statistically significant improvements in terms of prediction rate.The results are stable and robust with respect to parameter variations.

The increasing availability of text information coded in many different languages poses new challenges to modern information retrieval and mining systems in order to discover and exchange knowledge at a larger world-wide scale. The 1st International Workshop on Modeling, Learning and Mining for Cross/Multilinguality (dubbed MultiLingMine 2016) provides a venue to discuss research advances in cross-/multilingual related topics, focusing on new multidisciplinary research questions that have not been deeply investigated so far (e.g., in CLEF and related events relevant to CLIR). This includes theoretical and experimental on-going works about novel representation models, learning algorithms, and knowledge-based methodologies for emerging trends and applications, such as, e.g., cross-view cross-/multilingual information retrieval and document mining, (knowledge-based) translation-independent cross-/multilingual corpora, applications in social network contexts, and more.

National audience; Detecting change in evolving data streams is a central issue for accurate adaptive learning. In real world applications, data streams have categorical features, and changes induced in the data distribution of these categorical features have not been considered extensively so far. Previous work on change detection focused on detecting changes in the accuracy of the learners, but without considering changes in the data distribution. To cope with these issues, we propose a new unsupervised change detection method, called CDCStream (Change Detection in Categorical Data Streams), well suited for categorical data streams. The proposed method is able to detect changes in a batch incremental scenario. It is based on the two following characteristics: (i) a summarization strategy is proposed to compress the actual batch by extracting a descriptive summary and (ii) a new segmentation algorithm is proposed to highlight changes and issue warnings for a data stream. To evaluate our proposal we employ it in a learning task over real world data and we compare its results with state of the art methods. We also report qualitative evaluation in order to show the behavior of CDCStream.; Pour mieux analyser et extraire de la connaissance de flots de données, des approches spécifiques ont été proposées ces dernières années. L’un des challenges auquel elles doivent faire face est la détection de changement dans les données. Alors que de plus en plus de données qualitatives sont générées, peu de travaux de recherche se sont intéressés à la détection de changement dans ce contexte et les travaux existants se sont principalement focalisés sur la qualité d’un modèle appris plutôt qu’au réel changement dans les données. Dans cet article nous proposons une nouvelle méthode de détection de changement non supervisée, appelée CDCStream (Change Detection in Categorical Data Streams), adaptée aux flux de données qualitatives.

Nowadays, a huge amount of high resolution satellite images are freely available. Such images allow researchers in environmental sciences to study the different natural habitats and farming practices in a remote way. However, satellite images content strongly depends on the season of the acquisition. Due to the periodicity of natural and agricultural dynamics throughout seasons, sequential patterns arise as a new opportunity to model the behaviour of these environments. In this paper, we describe some preliminary results obtained with a new framework for studying spatiotemporal evolutions over natural and agricultural areas using k-partite graphs and sequential patterns extracted from segmented Landsat images.

International audience; Data labeling is an expensive andt ime-consuming task. Choosing which labels to use is increasingly becoming important. In the active learning setting, a classifier is trained by asking for labels for only a small fraction of all instances. While many works exist that deal with this issue in non-streaming scenarios, few works exist in the data stream setting. In this paper we propose a new active learning approach for evolving data streams based on a pre-clustering step, for selecting the most informative instances for labeling. We consider a batch incremental setting: when a new batch arrives, first we cluster the examples, and then, we select the best instances to train the learner. The clustering approach allows to cover the whole data space avoiding to oversample examples from only few areas. We compare our method w.r.t. state of the art active learning strategies over real datasets. The results highlight the improvement in performance of our proposal. Experiments on parameter sensitivity are also reported.

[Departement_IRSTEA]Territoires [TR1_IRSTEA]SYNERGIE; International audience; The availability of data represented with multiple features coming from heterogeneous domains is getting more and more common in real world applications. Such data represent objects of a certain type, connected to other types of data, the features, so that the overall data schema forms a star structure of inter-relationships. Co-clustering these data involves the specification of many parameters, such as the number of clusters for the object dimension and for all the features domains. In this paper we present a novel co-clustering algorithm for heterogeneous star-structured data that is parameter-less. This means that it does not require either the number of row clusters or the number of column clusters for the given feature spaces. Our approach optimizes the Goodman–Kruskal’s τ, a measure for cross-association in contingency tables that evaluates the strength of the relationship between two categorical variables. We extend τ to evaluate co-clustering solutions and in particular we apply it in a higher dimensional setting. We propose the algorithm CoStar which optimizes τ by a local search approach. We assess the performance of CoStar on publicly available datasets from the textual and image domains using objective external criteria. The results show that our approach outperforms state-of-the-art methods for the co-clustering of heterogeneous data, while it remains computationally efficient.

International audience; Tweets exchanged over the Internet are an important source of information even if their characteristics make them difficult to analyze (e.g., a maximum of 140 characters; noisy data). In this paper, we investigate two different problems. The first one is related to the extraction of representative terms from a set of tweets. More precisely we address the following question: are traditional information retrieval measures appropriate when dealing with tweets?. The second problem is related to the evolution of tweets over time for a set of users. With the development of data mining approaches, lots of very efficient methods have been defined to extract patterns hidden in the huge amount of data available. More recently new spatio-temporal data mining approaches have specifically been defined for dealing with the huge amount of moving object data that can be obtained from the improvement in positioning technology. Due to particularity of tweets, the second question we investigate is the following: are spatio-temporal mining algorithms appropriate for better understanding the behavior of communities over time? These two problems are illustrated through real applications concerning both health and political tweets. © Springer-Verlag Berlin Heidelberg 2013.

International Workshop on Spatial and Spatiotemporal Data Mining (SSTDM-12), In Cooperation with IEEE ICDM 2012, 10 December 2012, Brussels, Belgium; International audience; This paper faces the problem of harvesting geographic information from Web documents, specifically, extracting facts on spatial relations among geographic places. The motivation is twofold. First, researchers on Spatial Data Mining often assume that spatial data are already available, thanks to current GIS and positioning technologies. Nevertheless, this is not applicable to the case of spatial information embedded in data without an explicit spatial modeling, such as documents. Second, despite the huge amount of Web documents conveying useful geographic information, there is not much work on how to harvest spatial data from these documents. The problem is particularly challenging because of the lack of annotated documents, which prevents the application of supervised learning techniques. In this paper, we propose to harvest facts on geographic places through an unsupervised approach which recognizes spatial relations among geographic places without supposing the availability of annotated documents. The proposed approach is based on the combined use of a spatial ontology and a prototype-based classifier. A case study on topological and directional relations is reported and commented.

[Departement_IRSTEA]Territoires [TR1_IRSTEA]SYNERGIE; International audience; Current work on assembling a set of local patterns such as rules and class association rules into a global model for the prediction of a target usually focuses on the identification of the minimal set of patterns that cover the training data. In this paper we present a different point of view: the model of a class has been built with the purpose to emphasise the typical features of the examples of the class. Typical features are modelled by frequent itemsets extracted from the examples and constitute a new representation space of the examples of the class. Prediction of the target class of test examples occurs by computation of the distance between the vector representing the example in the space of the itemsets of each class and the vectors representing the classes. It is interesting to observe that in the distance computation the critical contribution to the discrimination between classes is given not only by the itemsets of the class model that match the example but also by itemsets that do not match the example. These absent features constitute some pieces of information on the examples that can be considered for the prediction and should not be disregarded. Second, absent features are more abundant in the wrong classes than in the correct ones and their number increases the distance between the example vector and the negative class vectors. Furthermore, since absent features are frequent features in their respective classes, they make the prediction more robust against over-fitting and noise. The usage of features absent in the test example is a novel issue in classification: existing learners usually tend to select the best local pattern that matches the example - and do not consider the abundance of other patterns that do not match it. We demonstrate the validity of our observations and the effectiveness of LODE, our learner, by means of extensive empirical experiments in which we compare the prediction accuracy ofLODE with a consistent set of classifiers of the state of the art. In this paper we also report the methodology that we adopted in order to determine automatically the setting of the learner and of its parameters.

International audience; Tweets exchanged over the Internet are an important source of information even if their characteristics make them dif- ficult to analyze (e.g., a maximum of 140 characters; noisy data). In this paper, we address the problem of extracting relevant topics through tweets coming from different commu- nities. More precisely we are interested to address the fol- lowing question: which are the most relevant terms given a community. To answer this question we define and evaluate new variants of the traditional TF-IDF. Furthermore we also show that our measures are well suited to recommend a community affiliation to a new user. Experiments have been conducted on tweets collected during French Presiden- tial and Legislative elections in 2012. The results underline the quality and the usefulness of our proposal.

International audience; Recent improvements in positioning technology has led to a much wider availability of massive moving object data. A crucial task is to find the moving objects that travel together. Usually, these object sets are called spatio-temporal patterns. Analyzing such data has been applied in many real world applications, e.g., in ecological study, vehi- cle control, mobile communication management, etc. However, few data mining tools are available for flexible and scalable analysis of massive scale moving object data. The main reason is that there is no a uni- fying approach to manage the patterns while many different kinds of spatio-temporal patterns have been proposed in recent years. Each ap- proach only focuses on mining a specific kind of pattern. Our system, GeT Move, is designed to extract and manage different spatio-temporal patterns concurrently. A user-friendly interface is provided to facilitate interactive exploration of mining results. Since GeT Move is tested on many kinds of real data sets, it will benefit users to carry out versatile analysis on these kinds of data by exhibiting different kinds of patterns at the same time.

International audience; In many domains (e.g., data mining, data management, data warehouse), a hierarchical organization of attribute values can help the data analysis process. Nevertheless, such hi- erarchical knowledge does not always available or even may be inadequate or useless when exists. Starting from this consideration, in this paper we tackle the problem of the automatic definition of data-driven taxonomies.To do this we combine techniques coming from information theory and clustering to obtain a structured representation of the at- tribute values: the Contextual Attribute-Value Taxonomy (CAVT). The two main advantages of our method are to be fully unsupervised (i.e., without any knowledge provided by an expert) and parameter-free. We experiments the bene- fit of use CAVTs in the two following tasks: (i) the mul- tilevel multidimensional sequential pattern mining problem in which hierarchies are involved to exploit abstraction over the data, (ii) the table summarization problem, in which the hierarchies are used to aggregate the data to supply a sketch of the original information to the user. To validate our approach we use real world datasets in which we obtain appreciable results regarding both quantitative and qualitative evaluation.

In this paper, we face the problem of extracting spatial relationships from geographical entities mentioned in textual documents. This is part of a research project which aims at geo-referencing document contents, hence making the realization of a Geographical Information Retrieval system possible. The driving factor of this research is the huge amount of Web documents which mention geographic places and relate them spatially. Several approaches have been proposed for the extraction of spatial relationships. However, they all assume the availability of either a large set of manually annotated documents or complex hand-crafted rules. In both cases, a rather tedious and time-consuming activity is required by domain experts. We propose an alternative approach based on the combined use of both a spatial ontology, which defines the topological relationships (classes) to be identified within text, and a nearest-prototype classifier, which helps to recognize instances of the topological relationships. This approach is unsupervised, so it does not need annotated data. Moreover, it is based on an ontology, which prevents the hand-crafting of ad hoc rules. Experimental results on real datasets show the viability of this approach.

An important step to understand the main functions of a specific family of proteins is the detection of protein features that could reveal how protein chains are constituted. To achieve this aim we treated amino acid sequences of proteins as a formal language, building a Context-Free Grammar annotated using an n-gram Bayesian classifier. This formalism is able to analyze the connection between protein chains and protein functions. In order to design new protein chains with the properties of the considered family we performed a rule clustering of the grammar to build an Annotated Stochastic Context Free Grammar. Our methodology was applied to a class of Antimicrobial Peptides (AmPs): the Frog antimicrobial peptides family. Through this case study, our approach pointed out some important aspects regarding the relationship between sequences and functional domains of proteins and how protein domain motifs are preserved by natural evolution in to the amino acid sequences. Moreover our results suggest that the synthesis of new proteins with a given domain architecture can be one of the fields where application of Annotated Stochastic Context Free Grammars can be useful.

The beginning of post-genomic era is characterized by a rising numbers of public collected genomes. The evolutionary relationship among these genomes may be caught by means of the comparative analysis of sequences, in order to identify both homologous and non-coding functional elements. In this paper we report on the on-going BIOBITS project. It is focused on studies concerning the bacterial endosymbionts, since they offer an excellent model to investigate important biological events, such as organelle evolution, genome reduction, and transfer of genetic information among host lineages. The BIOBITS goal is two-side: on the one hand, it pursues a logical data representation of genomic and proteomic components. On the other hand, it aims at the development of software modules allowing the user to retrieve and analyze data in a flexible way.

International audience; Microblogging is a modern communication paradigm in which users post bits of information, or "memes" as we call them, that are brief text updates or micromedia such as photos, video or audio clips. Once a user post a meme, it become visible to the user community. When a user finds a meme of another user interesting, she can eventually repost it, thus allowing memes to propagate virally trough the social network. In this paper we introduce the meme ranking problem, as the problem of selecting which k memes (among the ones posted by their contacts) to show to users when they log into the system. The objective is to maximize the overall activity of the network, that is, the total number of reposts that occur. We deeply characterize the problem showing that not only exact solutions are unfeasible, but also approximated solutions are prohibitive to be adopted in an on-line setting. Therefore we devise a set of heuristics and we compare them trough an extensive simulation based on the real-world Yahoo! Meme social graph, using parameters learnt from real logs of meme propagations. Our experimentation demonstrates the effectiveness and feasibility of these methods.

Association rules are an intuitive descriptive paradigm that has been used extensively in later years and in different application domains with the purpose to identify the regularities and correlation in a set of observed objects. However, recently, association rules’ statistical measures (support and confidence) have been criticized because in some cases have shown to fail their primary goal that is to select the most relevant and significant association rules. In this paper we propose a new model that replaces the support measure. The new model, like support, is a tool for the identification of the reliable rules and is used also to reduce the traversal of the itemsets search space. The proposed model adopts new criteria in order to establish the reliability of the information extracted from the database. These criteria are based on Bayes’ Theorem and on an estimate of the probability density function of each itemset. According to our criteria, the information that we have obtained from the database on an itemset is reliable if and only if the confidence interval of the estimated probability is low compared with the most likely value of it. We will see how this method can be computed in an approximated way, but satisfactory, with computational time comparable to the test on support

Clustering data described by categorical attributes is a challenging task in data mining applications. Unlike numerical attributes, it is difficult to define a distance between pairs of values of the same categorical attribute, since they are not ordered. In this paper, we propose a method to learn a context-based distance for categorical attributes. The key intuition of this work is that the distance between two values of a categorical attribute A i can be determined by the way in which the values of the other attributes A j are distributed in the dataset objects: if they are similarly distributed in the groups of objects in correspondence of the distinct values of A i a low value of distance is obtained. We propose also a solution to the critical point of the choice of the attributes A j . We validate our approach on various real world and synthetic datasets, by embedding our distance learning method in both a partitional and a hierarchical clustering algorithm. Experimental results show that our method is competitive w.r.t. categorical data clustering approaches in the state of the art.

EMNLP, Conference on Empirical Methods in Natural Language Processing , Doha, QAT, 25-/10/2014 - 29/10/2014; International audience; A major challenge in document clustering research arises from the growing amount of text data written in different languages. Previous approaches depend on language-specific solutions (e.g., bilingual dictionaries, sequential machine translation) to evaluate document similarities, and the required transformations may alter the original document semantics. To cope with this issue we propose a new document clustering approach for multilingual corpora that (i) exploits a large-scale multilingual knowledge base, (ii) takes advantage of the multi-topic nature of the text documents, and (iii) employs a tensor-based model to deal with high dimensionality and sparseness. Results have shown the significance of our approach and its better performance w.r.t. classic document clustering approaches, in both a balanced and an unbalanced corpus evaluation.

International audience; Multilingual document classification is often addressed by approaches that rely on language-specific resources (e.g., bilingual dictionaries and machine translation tools) to evaluate cross-lingual document similarities. However, the required transformations may alter the original document semantics, raising additional issues to the known difficulty of obtaining high-quality labeled datasets. To overcome such issues we propose a new framework for multilingual document classification under a transductive learning setting. We exploit a large-scale multilingual knowledge base, BabelNet, to support the modeling of different language-written documents into a common conceptual space, without requiring any language translation process. We resort to a state-of-the-art transductive learner to produce the document classification. Results on two real-world multilingual corpora have highlighted the effectiveness of the proposed document model w.r.t. document representations usually involved in multilingual and cross-lingual analysis, and the robustness of the transductive setting for multilingual document classification.