Martine Collard

Social-Based Conceptual Links: Conceptual Analysis Applied to Social Networks

In this work, we propose a novel approach for the discovery of frequent patterns in a social network on the basis of both vertex attributes and link frequency. With an analogy to the traditional task of mining frequent item sets, we show that the issue addressed can be formulated in terms of a conceptual analysis that elicits conceptual links. A social-based conceptual link is a synthetic representation of a set of links between groups of vertexes that share similar internal properties. We propose a first algorithm that optimizes the search into the concept lattice of conceptual links and extracts maximal frequent conceptual links. We study the performances of our solution and give experimental results obtained on a sample example. Finally we show that the set of conceptual links extracted provides a conceptual view of the social network.

How to Semantically Enhance a Data Mining Process

This paper presents the KEOPS data mining methodology centered on domain knowledge integration. KEOPS is a CRISP-DM compliant methodology which integrates a knowledge base and an ontology. In this paper, we focus first on the pre-processing steps of business understanding and data understanding in order to build an ontology driven information system (ODIS). Then we show how the knowledge base is used for the post-processing step of model interpretation. We detail the role of the ontology and we define a part-way interestingness measure that integrates both objective and subjective criteria in order to eval model relevance according to expert knowledge. We present experiments conducted on real data and their results.

D2SNet: Dynamics of diffusion and dynamic human behaviour in social networks

In this paper, we present an original and formal framework, the D2SNet model designed to combine both the social network evolution and the diffusion dynamics among individuals. We have conducted experiments on three social networks that show identical characteristics as real social networks. A formal definition of the model is provided and we describe its implementation in a simulation tool. We represent human behaviors and information dissemination strategies by standard and synthetic scheme. In a first step, we study the impact of network growing strategies only and we highlight important parameters such as the evolution speed and mainly the kind of strategies that favour or not the spread. Then we study a more complete evolution strategy that involves link creation and deletion. We provide a deep analysis on the impact of each parameter such as evolution speed, creation and deletion probabilities and dynamic human behaviors on the diffusion amplitude and coverage. Our study gives a novel and useful insight in the diffusion process in a dynamic context.

Diffusion in dynamic social networks: application in epidemiology

Structure and evolution of networks have been areas of growing interest in recent years, especially with the emergence of Social Network Analysis (SNA) and its application in numerous fields. Researches on diffusion are focusing on network modeling for studying spreading phenomena. While the impact of network properties on spreading is now widely studied, involvement of network dynamicity is very little known. In this paper, we address the epidemiology context and study the consequences of network evolutions on spread of diseases. Experiments are conducted by comparing incidence curves obtained by evolution strategies applied on two generated and two real networks. Results are then analyzed by investigating network properties and discussed in order to explain how network evolution influences the spread. We present the MIDEN framework, an approach to measure impact of basic changes in network structure, and DynSpread, a 2D simulation tool designed to replay infections scenarios on evolving networks.

Multi-criteria evaluation of interesting dependencies according to a data mining approach

This paper addresses the problem of the goodness of dependency rules extracted by mining data. Our approach is experimental and based on the idea that model quality such as accuracy, interestingness or domain-dependent criteria. Most works on model quality are focusing on one criterion at a time only and do not take into account multiple factors simultaneously. A few works combine different measures in weighted expressions. In order to combine multiple measures, we have first realized a comparative study which highlights the relative contribution of different factors and reveals trade-offs among some of them. This situation suggests looking in the rules which may not exist. Thus, we show that a multi-objective evolutionary approach is able to reveal interesting rules which are ignored by standard solutions.

Mobility and Information Flow: Percolation in a Multi-Agent Model

Abstract In real world situations, each person is generally in contact with only a small fraction of the entire population and exchange information through these interactions. Their number and their frequency vary from one to another individual and may be much depending on mobility of individuals. The objective of this article is to better understand how human mobility may have an impact on mobile social networking systems. This should help to answer a question as: “How might an information, a rumor, a pathogen, etc., driven by physical proximity, spread through a population?”. We present a first stage of our work in which we focus on percolation processes as information flow mechanisms. We propose a synthetic mobility model and we define an artificial world populated by heterogeneous agents who differ in their mobility. Simulations are conducted on a multi-agent programmable environment. Our experimental results clearly demonstrate positive correlations between agent mobility factors and percolation thresholds.

Towards a hybrid algorithm for extracting maximal frequent conceptual links in social networks

One of the most common tasks in the area of social network mining is the extraction of frequent patterns from social networks. Although traditional approaches have been mainly focused on subgraphs occurring frequently in a network or a set of networks, new approaches have attempted to exploit network structure and node properties in order to elicit new kinds of patterns. One of these news approaches is the extraction of conceptual links, a solution that combines both structure and node properties for providing knowledge on the groups of nodes the most connected in a social network. However, if the extraction of conceptual links offers a great potential in terms of knowledge discovery as well as network visualization, the search for these kinds of patterns remains a computationally intensive problem. Thus, the efficiency of the extraction processes highly depends on the efficiency of the underlying algorithm. In this paper, we focus on the extraction of maximal frequent conceptual links (MFCL) and we propose a hybrid algorithm that applies a filter on the nodes for reducing the search space to the most frequent groups of nodes. Although our solution potentially causes loss of some patterns, we demonstrate its efficiency by comparison with the MFCLMin algorithm. We investigate the potential loss of patterns, the gain on the runtime, and the gain on the number of comparisons.

FLMin: An Approach for Mining Frequent Links in Social Networks

This paper proposes a new knowledge discovery method called FLMin to discover frequent patterns in a social network. The algorithm works without previous knowledge on the network and exploits both the structure and the attributes of nodes to extract regularities called Frequent Links. Unlike traditional works in this area that solely exploit structural regularities of the network, the originality of FLMin is its ability to gather these two kinds of information in the search for patterns. In this paper, we detail the method proposed for extracting frequent links and discuss its complexity and its flexibility. The efficiency of our solution is evaluated by conducting qualitative and quantitative studies for understanding how behaves FLMin according to different parameters.

Wireless sensor network for habitat monitoring: A counting heuristic

More and more animal species are endangered. In order to study and protect them, several measures have been taken, which now seem to show their limit. This work focuses on the counting process, a key issue for any project that aims to protect animals. Indeed, this paper proposes an algorithm for automatic counting of singing birds in their habitat by using wireless sensors fitted with microphone. Sensors are used to record audio samples of bird songs. These samples are used to extract some kinds of song fingerprint that are thereafter used to recognize the bird species, through a classification process. Afterwards, by analyzing data which have been collected by a central base, the network structure derived from detection fields of sensors, allows our algorithm to propose an estimate of singing birds in the habitat. This paper details the counting algorithm which uses graph theory and audio inputs comparison. Finally, we demonstrate our scheme efficiency through experimentations.

MAX-FLMin: An Approach for Mining Maximal Frequent Links and Generating Semantical Structures from Social Networks

The paper proposes a new knowledge discovery method called MAX-FLMin for extracting frequent patterns in social networks. Unlike traditional approaches that mainly focus on the network topological structure, the originality of our solution is its ability to exploit information both on the network structure and the attributes of nodes in order to elicit specific regularities that we call “Frequent Links”. This kind of patterns provides relevant knowledge about the groups of nodes most connected within the network. First, we detail the method proposed to extract maximal frequent links from social networks. Second, we show how the extracted patterns are used to generate aggregated networks that represent the initial social network with more semantics. Qualitative and quantitative studies are conducted to evaluate the performances of our algorithm in various configurations.