To help users better understand and improve a Machine Learning (ML) process, we propose a visual method to analyze a large number of supervised learning results. Each classification result is a “representation -classifier- parameters” triplet which behavior can be described with standard indices (accuracy, etc) and also with the probabilities of predicted classes for each data. This information can be used to compute a distance/similarity between classification results. Then, to build the visualization, each triplet is positioned in 2D with a dimension reduction algorithm (MDS). This method can be applied to any dataset and we present results with a real-world application (detection of autistic disorders from eye tracking data). With different settings of this visualization, we show that the user can visually observe and evaluate the effectiveness of the representations” classifiers and parameters.

This paper presents a survey on automatic or semi-automatic recommendation systems which help users to create dashboards. It starts by showing the important role that dashboards play in data science, and give an informal definition of dashboards, i.e., a set of visualizations possibly with linkage, a screen layout and user feedback. We are mainly interested in systems that use a fully or partially automatic mechanism to recommend dashboards to users. This automation includes the suggestion of data and visualizations, the optimization of the layout and the use of user feedback. We position our work with respect to existing surveys. Starting from a set of over 1000 papers, we have selected and analyzed 19 papers/systems along several dimensions. The main dimensions were the set of considered visualizations, the suggestion method, the utility/objective functions, the layout, and the user interface. We conclude by highlighting the main achievements in this domain and by proposing perspectives.

In this paper, we present a method for the automatic generation of dashboards (DBo) using a genetic algorithm (GA). A DBo is a set of visualizations, with possible linkage, intended to help users explore and analyse a dataset. Our Automatic Dashboard Generation System (ADGS) considers several input models for data, user and visualizations. We propose to represent a solution (i.e. a DBo) as a variable size matrix in which rows are visualizations and columns are data attributes. This representation can be evolved with a GA. For this purpose, we define genetic operators for DBo such as random generation, crossover, and mutation. We propose a fitness function to evaluate the quality of a DBo, as well as selection and replacement schemes. Finally, we present results with a benchmark dataset and a given user scenario. We show that the GA can find DBo that maximizes the evaluation function and that can be interesting for novice users. In perspectives, we will improve further the GA and we will study how to automatically propose a layout of the optimized DBo.

We present in this paper a visual method to explore the properties of an image dataset and its internal representation learned by a convolutional neural network. We consider the inner characteristics extracted by the network just before the classification layers. We build a neighborhood graph from this vector space by connecting data together according to specific topological properties. We define typical examples of topological anomalies to be detected (isolated points, erroneous points, class boundaries). Then we propose a visualization of this graph highlighting this information and offering an overview of the graph (groups of data) as well as local details (fine topological properties). This visualization includes a representation of the images in order to let the user understand what can cause an error (errors during image acquisition, pre-processing or labeling, or errors due to the choice of the network or the learning parameters, etc.). We perform several tests with the VGG16 network on samples of standard datasets.

We consider in this paper a radial visualization called POIViz as a starting point to be improved with an optimization procedure. In our previous work, we studied POIViz and showed that it was able to represent multidimensional data in 2D within a few seconds, even for datasets with millions of records. We provided a simple heuristic to select the Points Of Interest (POIs), i.e., the 2D anchors that determine the layout of the data. In this paper, we extend POIViz to Gen-POIViz by proposing a genetic algorithm (GA) that can greatly optimize the quality of the visualization. The GA searches for a set of POIs that minimizes a cost function that is based on Kruskal's stress. Furthermore, Gen-POIViz can find relevant POIs with a small sample of the data only, and thus it can compute a projection of the complete data in a very short time. We provide comparative results with standard methods in data projection. Gen-POIViz obtains results with a quality that is between force-directed Multidimensional Scaling (MDS) and Principal Components Analysis (PCA). On larger datasets, we show the advantage of our method when it works on a data sample. It can be much faster than MDS, and it can be run with even larger datasets for which other methods fail.

Metagenomics is an important field in biology where an environmental sample is sequenced to study the genomic content of species present in it. The data obtained from sequencing is a mixture of DNA fragments obtained from several species present in the sample. So an important step in this data analysis is to group together the DNA fragments originating from same specie or genera. In this paper we present an approach named ProxiClust, where we exhibit how Community Detection methods can be used to handle this task. The large size of the dataset poses challenge in using the traditional data mining technique given their computation and memory complexity. We aim to achieve scalability through a deterministic approach by converting the data from cloud points in to a graph and leverage community detection on it for identifying groups. Firstly the relevant pairwise relationships between DNA fragments are extracted by building proximity graphs on the data so storing complete distance matrix in memory can be avoided. The groups on graph are identified by leveraging community detection methods. We perform exploratory study to examine properties of several approaches for this framework and exhibit specific instances of this approach that perform comparably with state of the art binning methods.

Thanks to the capturing devices cost reduction and the advent of social networks, the size of image collections is becoming extremely huge. Many works in the literature have addressed the indexing of large image collections for search purposes. However, there is a lack of support for exploratory data mining. One may want to wander around the images and experience serendipity in the exploration process. Thus, effective paradigms not only for organising, but also visualising these image collections become necessary. In this article, we present a study to jointly index and visualise large image collections. The work focuses on satisfying three constraints. First, large image collections, up to million of images, shall be handled. Second, dynamic collections, such as ever-growing collections, shall be processed in an incremental way, without reprocessing the whole collection at each modification. Finally, an intuitive and interactive exploration system shall be provided to the user to allow him to easily mine image collections. To this end, a data partitioning algorithm has been modified and proximity graphs have been used to fit the visualisation purpose. A custom web platform has been implemented to visualise the hierarchical and graph-based hybrid structure. The results of a user evaluation we have conducted show that the exploration of the collections is intuitive and smooth thanks to the proposed structure. Furthermore, the scalability of the proposed indexing method is proved using large public image collections.

In this paper, a visual analysis system to qualitatively assess the features and distance functions that are used for calculating dissimilarity between two word images is presented. Computation of dissimilarity between two images is the prerequisite for image matching, indexing and retrieval problems. First, the features are extracted from the word images and a distance between each image to others is computed and represented in a matrix form. Then, based on this distance matrix, a proximity graph is built to structure the set of word images and highlight their topology. The proposed visual analysis system is a web based platform that allows visualisation and interactions on the obtained graph. This interactive visualisation tool inherently helps users to quickly analyse and understand the relevance and robustness of selected features and corresponding distance function in a unsupervised way, i.e. without any ground truth. Experiments are performed on a handwritten dataset of segmented words. Three types of features and four distance functions are considered to describe and compare the word images. Theses material are leveraged to evaluate the relevance of the built graph, and the usefulness of the platform.

The amount of captured images has increased exponentially these last years. Online available image collections are becoming common thanks to social networks or institutes digitization programs. The context of our work falls into the need to explore such image collections. The literature paradigms are leveraged to meet three constraints: (i) handling medium to large image collections, (ii) handling dynamic image collections and (iii) providing interactive visualisations. In this paper, we describe how our work has been used to build a recommendation system and an interactive exploration platform for dynamic image collection. To illustrate the relevance of such tools, we present APoD eXplorer, an online available platform that enhances the exploration of the NASA Astronomy Picture of the Day image collection. The platform is available at http://frederic.rayar.free.fr/apod/.

Nowadays, more and more information is flowing in and is provided on the Web. Large datasets are made available covering many fields and sectors. Open Data (OD) plays an important role in this field. Thanks to the volumes and the variety of the released datasets, OD brings high societal and business potential. In order to realize this potential, the reuse of the datasets (e.g. in internal business processes) becomes primordial. However, if the aim is to reuse OD, it is also necessary to be able of assessing its quality. This paper demonstrates how Information Visualization may help on this task and presents Stacktab chart - a new chart to analyse and assess CSV files in order to understand their structure, identify the location of relevant information and detect possible problems in the datasets.

We study in this paper the visualization of large multidimensional datasets with a focus on Open Data. Starting from our early work in which we defined a visualization based on points of interest, we improve this method in several ways with the aim of dealing with larger datasets and especially Open datasets. We propose the parallelization, using CPU and GPU, of the most costly steps of our method, like the computation of the data layout. We improve the visualization with a density rendering so as to keep the display informative for large datasets and for Open Data. We propose a layered visualization with interactions that can support several users tasks such as data filtering and labeling. We show that, even with common hardware, the performances of our approach are such that any user graphical queries can be processed in a few seconds. We detail how we were able to visualize and explore a collection of 300,000 Open datasets from the French Open Data web site. With the resulting visualization, we were able to improve our previous results.

International audience

In this paper, we propose an original solution to the problem of point cloud clustering. The proposed technique is based on a d-dimensional formulated Delaunay Triangulation (DT) construction algorithm and adapts it to the problem of cluster detection. The introduced algorithm allows this detection as along with the DT construction. Precisely, a criterion that detects occurrences of gaps in the simplex perimeter distribution is added during the incremental DT construction. This detection allows to label simplices as being inter - or intra cluster. Experimental results on 2D shape datasets are presented and discussed in terms of cluster detection and topological relationship preservation.

In this paper we study a 3D tubular visualization of software activity log data, with the aim of supporting multithreaded software development and debugging. We consider an existing visualization called Datatube2 that has already been used to help various domain experts in the analysis of large amounts of time-dependent data. Since software logs are also time series, DataTube2 has been enhanced to support the specific data and tasks that are currently found in debugging, yielding to a specific visualization that we called DataTube4log. In this visualization, each line in the tube is devoted to the activity of a thread. Dependencies between threads are materialized with arrows. Synchronization objects are also represented. Using a real dataset, we show how a domain expert has solved a debugging problem. In this experiment, we found that DataTube4log can be easily learned and adopted, and that the ability of the tube to efficiently render an overview of large time series was beneficial to the software engineer.

This paper presents an original feature vector extraction process based on the Delaunay triangulation (DT) and a zoning technique. The presented work provides an illustration of the equivalency between a zoning and the Delaunay triangulation in the context of handwritten character recognition. A novel technique that relies on the approximation of a DT and an automatic pruning calculation is introduced. We call this technique the alpha-approximation. To discuss our contribution, experiments are conducted on the MNIST database of handwritten digits using a support vector machine classifier for the classification task.

We study in this paper the use of a 3D interface for OLAP. We analyze the state of the art in OLAP 3D visualizations. In a first step, we propose a new interface, called VR4OLAP, that combines the relative advantages of the studied methods. VR4OLAP can visualize 3 dimensions of an OLAP data cube and up to two measures. It represents the OLAP operators in the visualization with 3D widgets. It can use a 3D stereoscopic screen with a 3D mouse. In a second step, we perform a user study to evaluate this interface and to compare it with a standard cross-table. We evaluate the technical performances of the 3D interface. We conclude that, for the studied data and methods, 1) users are quite enthusiastic with a 3D representation, 2) the performances in 2D are equal or better than those obtained in 3D, 3) no advantage was found with the use of the 3D immersive setup.

We deal in this paper with the problem of creating an interactive and visual map for a large collection of Open datasets. We first describe how to define a representation space for such data, using text mining techniques to create features. Then, with a similarity measure between Open datasets, we use the k-nearest neighbors method for building a proximity graph between datasets. We use a force-directed layout method to visualize the graph (Tulip Software). We present the results with a collection of 293,000 datasets from the French Open data web site, in which the display of the graph is limited to 151,000 datasets. We study the discovered clusters and we show how they can be used to browse this large collection.

We deal in this paper with the problem of creating an interactive and visual map for a large collection of Open datasets. We first describe how to define a representation space for such data. We use text mining techniques to create features. Then, with a similarity measure between Open datasets, we use the Relative Neighbors method for building a proximity graph between datasets. We use a force-directed layout method to visualize the graph (Tulip Software). We present the results with a collection of 300,000 datasets from the French Open data web site, in which the display of the graph is limited to 150,000 datasets. We study the discovered clusters and we show how they can be used to browse this large collection.

In this paper, we propose a new method for the visual reorganization of online analytical processing (OLAP) cubes that aims at improving their visualization. Our method addresses dimensions with hierarchically organized members. It uses a genetic algorithm that reorganizes k-ary trees. Genetic operators perform permutations of subtrees to optimize a visual homogeneity function. We propose several ways to reorganize an OLAP cube depending on which set of members is selected for the reorganization: all of the members, only the displayed members, or the members at a given level (level by level approach). The results that are evaluated by using optimization criteria show that our algorithm has a reliable performance even when it is limited to 1 minute runs. Our algorithm was integrated in an interactive 3D interface for OLAP. A user study was conducted to evaluate our approach with users. The results highlight the usefulness of reorganization in two OLAP tasks.

We deal in this paper with the problem of automating the process of choosing an appropriate visualization and its parameters in the context of visual data mining (VDM). To solve this problem, we develop a user assistant that performs 2 steps: the system starts by suggesting to users different mappings between their data and possible visualizations. This is performed with a simple but generic heuristic that can be applied to any visualization. Then, the user selects a visualization among those proposed by our assistant, and he may further improve the parameters set that defines the mapping between the visual attributes and the data attributes. For this purpose, we use an interactive genetic algorithm (IGA), which allows users to visually evaluate and adjust the mappings. We present a user evaluation that confirms the interest of our system in two tasks.

In this paper we present a new incremental algorithm for building neighborhood graphs between data. It is inspired from the self-assembly behavior observed in real ants where ants progressively become attached to an existing support and then successively to other attached ants. Each artificial ant represents one data. The way ants move and build a graph depends on the similarity between the data. We have compared our results to those obtained by the relative neighborhood algorithm on several databases (either artificial or real), and we show that our method is competitive especially with respect to execution times.

We present an incremental algorithm for building a neighborhood graph from a set of documents. This algorithm is based on a population of artificial agents that imitate the way real ants build structures with self-assembly behaviors. We show that our method outperforms standard algorithms for building such neighborhood graphs (up to 2230 times faster on the tested databases with equal quality) and how the user may interactively explore the graph.

This paper presents a new bio-inspired algorithm (FClust) that dynamically creates and visualizes groups of data. This algorithm uses the concepts of a flock of agents that move together in a complex manner with simple local rules. Each agent represents one data. The agents move together in a 2D environment with the aim of creating homogeneous groups of data. These groups are visualized in real time, and help the domain expert to understand the underlying structure of the data set, like for example a realistic number of classes, clusters of similar data, isolated data. We also present several extensions of this algorithm, which reduce its computational cost, and make use of a 3D display. This algorithm is then tested on artificial and real-world data, and a heuristic algorithm is used to evaluate the relevance of the obtained partitioning.

The aim of web pages visualization is to present in a very informative and interactive way a set of web documents to the user in order to let him or her navigate through these documents. In the web context, this may correspond to several user's tasks: displaying the results of a search engine, or visualizing a graph of pages such as a hypertext or a surf map. In addition to web pages visualization, web pages clustering also greatly improves the amount of information presented to the user by highlighting the similarities between the documents [6]. In this paper we explore the use of a cellular automata (CA) to generate such maps of web pages.

We present a new bio-inspired algorithm that dynamically creates and visualizes groups of data. This algorithm uses the concepts of a flock of agents that move together in a complex manner with simple local rules. Each agent represents one data. The agents move with the aim of creating homogeneous groups of data that evolve together in a 2D environment. These created groups are visualized in real time and help the domain expert to understand the underlying class structure of the data set, like for example a realistic number of classes, clusters of similar data, isolated data, etc. We present several extensions of this algorithm and present results from artificial and real-world data.

We present how we use an interactive genetic algorithm to find the best parameters to build an artificial art work according to user's aesthetic taste. Ants are used to spread colors on a numerical painting and behave with very simple rules to follow and deposit colors. These rules and colors are considered as parameters for the evolutionary process. This work can be considered as a contribution to naturally inspired artificial art and evolutionary techniques are used to help artists in their creative process.

We present a new clustering algorithm for unsupervised learning. It is inspired from the self-assembling behavior observed in real ants where ants progressively become attached to an existing support and then successively to other attached ants. The artificial ants that we have defined similarly builds a tree. Each ant represents one data. The way ants move and build this tree depends on the similarity between the data. We have compared our results to those obtained by the k-means algorithm and by AntClass on numerical databases (either artificial, real, or from the CE.R.I.E.S.). We show that AntTree significantly improves the clustering process.

Supervised classification has already been the subject of numerous studies in the fields of Statistics, Pattern Recognition and Artificial Intelligence under various appellations which include discriminant analysis, discrimination and concept learning. Many practical applications relating to this field have been developed. New methods have appeared in recent years, due to developments concerning Neural Networks and Machine Learning. These "hybrid" approaches share one common factor in that they combine symbolic and numerical aspects. The former are characterized by the representation of knowledge, the latter by the introduction of frequencies and probabilistic criteria. In the present study, we shall present a certain number of hybrid methods, conceived (or improved) by members of the SYMENU research group. These methods issue mainly from Machine Learning and from research on Classification Trees done in Statistics, and they may also be qualified as "rule-based". They shall be compared with other more classical approaches. This comparison will be based on a detailed description of each of the twelve methods envisaged, and on the results obtained concerning the "Waveform Recognition Problem" proposed by Breiman et al.,4 which is difficult for rule based approaches.