Recent progress has been made in the field of multilabel streaming classification, where an instance can be associated with several labels simultaneously. Most recent research has focused on adapting models to the dynamic distribution of nonstationary data streams. However, continual learning is not only adaptation to concept drift: phenomena such as catastrophic forgetting, as well as forward and backward transfers, appear when new classification tasks are introduced in the data stream. This paper aims to develop a standardized and operational evaluation protocol specifically adapted to the study of these phenomena, in order to identify the most promising strategies for this new multi-label, multi-task learning problem on tabular data stream. This protocol, which is reproducible, fair, and flexible enough to allow the simulation of a large number of scenarios, includes the creation of multi-label and multi-task streams, and an evaluation protocol for measuring (a) online performance, (b) phenomena linked to continual learning and (c) resources consumed. It is tested to compare 18 continual multi-label classification strategies on 6 open literature datasets and 3 simulated datasets. This exploratory analysis has enabled us to identify the promising nature of neural networks coupled with data replay.

Declarative variables of self-description have a long-standing tradition in matchmaking media. With the advent of online dating platforms and their brand positioning, the volume and semantics of variables vary greatly across apps. However, a variable landscape across multiple platforms, providing an in-depth understanding of the dating structure offered to users, has hitherto been absent in the literature. In this study, more than 300 declarative variables from 22 Anglophone and Francophone dating apps are examined. A mixed-method research design is used, combining hierarchical classification with an interview analysis of nine founders and developers in the industry. We present a new typology of variables in nine categories and a classification of dating apps, which highlights a double mimetic-distinctive mechanism in the variable definition and reflects the dating market. From the interviews, we extract three main factors concerning the economic and sociotechnical framework of coding practices, the actors' personal experience, and the development methodologies including user traces that influence this mechanism. This work, which to our knowledge is the most extensive thus far on dating app declarative variables, provides a new perspective on the analysis of the intersection between developers and users of online dating, which is mediated through variables, among other components.

Multi-label classification has gained in importance in the last decade and it is today confronted to the current needs to process massive raw data from heterogeneous sources. Therefore, dimensionality reduction, which aims at reducing the number of features, labels, or both, knows a renewed interest to enhance the scaling properties of the classifiers and their predictive performances. In this paper we review more than fifty papers presenting dimensionality reduction approaches for multi-label classification and we propose an analysis in three steps : (i) a typology of the methods describing the main components of their strategies, the problem they tackle and the way they solve it (ii) a unified formalization of the problems to help to distinguish the similarities and differences between the approaches, and (iii) a meta-analysis of the published experimental results inspired by the consensus theory to identify the most efficient algorithms.

Training machine learning models requires the storage of an ever-growing number n of parameters, which leads to memory management issues. In this paper, we propose a novel approach to accurately estimate the largest parameters sufficient for prediction tasks while saving memory. Inspired by the heavy hitter identification problem in data stream analysis, our count-sketch based strategy consists in independently storing t ~ 10 times all parameter values on different sets of r <;<; n counters by randomly aggregating several values in the same counters. Each parameter is then approximated by the median of the values stored in its t associated counters. We conduct experiments on the popular use case of the linear regression problem for nine multi-label datasets inducing various model sizes (up to 11.7 GB of RAM). It shows that replacing the regular storage of all parameters with the proposed count-sketch storage preserves the quality of predictive performances while significantly reducing the use of memory (from 80% to 99.9%). A theoretical bound on the approximation error completes the experiments. In practice, this work gives a chance to several potentially interesting memory-hungry models to be implemented without having to resort to costly computational environments.

Extreme Multi-label Learning (XML) considers large sets of items described by a number of labels that can exceed one million. Tree-based methods, which hierarchically partition the problem into small scale sub-problems, are particularly promising in this context to reduce the learning/prediction complexity and to open the way to parallelization. However, the current best approaches do not exploit tree randomization which has shown its efficiency in random forests and they resort to complex partitioning strategies. To overcome these limits, we here introduce a new random forest based algorithm with a very fast partitioning approach called CRAFTML. Experimental comparisons on nine datasets from the XML literature show that it outperforms the other tree-based approaches. Moreover with a parallelized implementation reduced to five cores, it is competitive with the best state-of-the-art methods which run on one hundred-core machines.

Time series analysis is the centerpiece of numerous research fields from stock analysis to topic mining. While the univariate case is still commonplace, there is an increasing need for tools providing features to study the relationships between multiple time series. Initially motivated by a interdisciplinary research agenda with sociologists and musicologists, we propose an extension of the famous stacked graph to display an overview of a set of multiple and item-set time series. This visualization allows the exploration of the general tendencies observed on a population and the comparison of patterns between groups. A proof-of-concept is presented on real-life data extracted from a recent study on the daily music listening behavior.

Stimulated by the rise of music recommendation services, the visualization of music collections has undergone a remarkable transformation over the last decade. In this communication, we focus on the visual comparison of personal music collections. We introduce an interactive and tridimensional system which allows their visualization using a music archipelago metaphor and relying on a facet-based engine. In the industrial context of music related services where social and sharing aspects are omnipresent, we also present a novel feature to enable the visual comparison of music collections.

International audience

The validation of ontology quality is still a challenging task and is becoming even more so whenever ontologies are automaticall built. In previous works, we proposed a quite large classification of quality problems that need to be carefully analyzed for accomplishing the validation task. In this communication, we are going to identify the links between potential quality problems and the steps required to build an ontology. These problems result in what we call the "suboptimal executions" of those steps.

This paper presents a new hands-free prototype TempoSpring for exploring complex social networks. Based on the Microsoft Kinect, our system lets the user rotate, zoom and filter the graph, and aggregate nodes to highlight structural properties. TempoSpring provides complementary visual restitutions (monoscopy, passive and active stereoscopy) to a user standing in front of the wall/screen on which the graph is displayed. To overcome screen limitations or information mixtures when details are required on graph components we propose a multimodal approach which displays subgraphs on a secondary screen (e.g mobile device) to analyze details at a lower scale.

Date du colloque : 07/2011

Defining a suitable semantic similarity between concept pairs of a subsumption hierarchy is becoming a generic problem for many applications in knowledge engineering exploiting ontologies. In this paper, we define a generic framework which can guide the proposition of new measures by making explicit the information on the ontology which has not been integrated into existing definitions yet. Moreover, this framework allows us to rewrite numerous measures, originally proposed in various contexts, which are in fact closely related to each other. From this observation, we show some metrical and ordinal properties. Experimental comparisons on Word-Net and on collections of human judgments complete the theoretical results and confirm the relevance of our propositions.

Metaheuristics are now so common that for some classical hard combinatorial problems, there exist more than ten variants. Thus, the issue of comparing optimization methods is crucial. In this paper, we focus on one aspect of this question: the impact of the choice of the test instances on the metaheuristic performances and the possible link with the fitness landscape structure. We base our experimental framework on the arc crossing minimization problem for layered digraphs. We compare a hybridized genetic algorithm and a multistart descent which are among the best approaches to this problem. We worked on two instance families with various sizes and structural complexities: small graphs which are easy to draw on a standard size support, and large graphs specifically built for our experiments. We show that, for the smallest instances, there is no significant difference between methods whereas for graphs similar to those classically used nowadays in applications the genetic algorithm is better, and for the largest graphs (with a scaling factor up to 10300), the multistart descent is the best method. These results suggest that for "structured" fitness landscapes associated with real-life instances the GA exploits its implicit learning. On the other hand for very large landscapes with probably numerous local optima, only one exploration on a larger scale can be provided by local searches from a random starting point, cheap in computing effort.

Presents a Web cartography approach to be used in the context of online site promotion. The overall objective is to provide users with handy maps offering information about candidate sites for the creation of hyperlinks that enable a large flow of targeted visitors. Two main types of data must be considered: texts and hyperlinks. We propose to exploit the latter to construct a relevant corpus on which semantic as well as graph analyses can be applied. The stress is put on the clustering of Web resources based on the link network, which makes it possible to highlight groups of strongly connected sites which are of the utmost interest for our application. To tackle the site graph partitioning problem, we turn to a promising iterative approach initially developed in the context of computer-aided design. It uses spectral decomposition of the Laplacian matrix to embed the considered graph in a geometric space where efficient methods can be applied. An algorithm that was adapted from an existing one implements the method. Experiments were conducted on a real application case concerning the promotion of a site dealing with Cognac. We present the obtained map as well as leads to exploit it.

The recent importance given to the integration of the user in a KDD (knowledge discovery in databases) process, which gives him the opportunity to direct his mining towards his own specific needs, requires the development of new highly interactive visualization tools. For graph based representation of discovered knowledge, this means that layout algorithms must dynamically take modifications into account. The authors present a genetic approach to draw a series of layered directed graphs which model relationships between association rules. We develop new problem-specific genetic operators and show that genetic algorithms are well-adapted to solve a multiobjective problem: meeting static aesthetic requirements such as minimizing arc crosses and preserving the "user's mental map" when a transformation is interactively performed on the graph. Experimental results are presented on several randomly generated series of graphs.

The paper presents new experimental results on a recently developed heuristic inspired by dead body clustering and larval sorting behavior of ants. The heuristic highlights organization of connected graphs by displaying spatially separated clusters of highly connected vertex subsets on a two-dimensional grid. After discussing some theoretical properties of graph visualization in small dimensional metric spaces, we compare our heuristic with more classical approaches coming from multidimensional scaling.