Chabane Djeraba

Mining Multimedia and Complex Data

Today’s technology makes it possible to easily access huge amounts of complex data. As a consequence, techniques are needed for accessing the semantics of such data and supporting the user in selecting relevant information. While meta-languages such as XML have been proposed, they are not suitable for complex data such as images, video, sounds or any other non-verbal channel of communication, because those data have very subjective semantics, i.e., whose interpretation varies over time and between subjects. Yet, providing access to subjective semantics is becoming critical with the significant increase in interactive systems such as web-based systems or socially interactive robots. In this work, we attempt to identify the requirements for providing access to the subjective semantics of complex data. In particular, we focus on how to support the analysis of those dimensions that give rise to multiple subjective interpretations of the data. We propose a data warehouse as a support for the mining process involved. A unique characteristic of the data warehouse lays in its ability to store multiple hierarchical descriptions of the multimedia data.

What are the grand challenges for data mining?: KDD-2006 panel report

We discuss what makes exciting and motivating Grand Challenge problems for Data Mining, and propose criteria for a good Grand Challenge. We then consider possible GC problems from multimedia mining, link mining, large-scale modeling, text mining, and proteomics. This report is the result of a panel held at KDD-2006 conference.

MDM/KDD2002: multimedia data mining between promises and problems

This report presents a brief overview of multimedia data mining and the corresponding workshop series at ACM SIGKDD conference series on data mining and knowledge discovery. It summarizes the presentations, conclusions and directions for future work that were discussed during the 3rd edition of the International Workshop on Multimedia Data Mining, conducted in conjunction with KDD-2002 in Edmonton, Alberta, Canada.

A Markovian approach for web user profiling and clustering

The objective of this paper is to propose an approach that extracts automatically web user profiling based on user navigation paths. Web user profiling consists of the best representative behaviors, represented by Markov models (MM). To achieve this objective, our approach is articulated around three notions: (1) Applying probabilistic exploration using Markov models. (2) Avoiding the problem of Markov model high-dimensionality and sparsity by clustering web documents, based on their content, before applying the Markov analysis. (3) Clustering Markov models, and extraction of their gravity centers. On the basis of these three notions, the approach makes possible the prediction of future states to be visited in k steps and navigation sessions monitoring, based on both content and traversed paths. The original application of the approach concerns the exploitation of multimedia archives in the perspective of the Copyright Deposit that preserves Frenchs WWW documents. The approach may be the exploitation tool for any web site.

Rule Evaluations in a KDD System

In this paper, we address, some database problems that a knowledge discovery system deals with. In databases, data may be noisy (uncertain), sparse and redundant. To solve these problems, we describe two methods: The first one is the rule intensity measurement which is an index that answers the question ‘What is the probability of having a rule between two propositions or two conjunctions of propositions ?’ The intensity of rule enables us to measure the probability of an implication of the form: IF premise THEN Conclusion. This index seems to be adapted to the field of Knowledge Discovery in Databases (KDD). It resists noise, converges with the size of the sample, eliminates coarse rules, and can be used within the framework of an incremental algorithm. We will analyse it in detail, and compare it with others. The second one eliminates the redundant rules and superfluous propositions by using an algorithm for finding a minimal set of rules.

Is there a grand challenge or X-prize for data mining?

This panel will discuss possible exciting and motivating Grand Challenge problems for Data Mining, focusing on bioinformatics, multimedia mining, link mining, text mining, and web mining.

A web user profiling approach

People display regularities in almost everything they do. This paper proposes characteristics of an idealized algorithm that would allow an automatic extraction of web user profil based on user navigation paths. We describe a simple predictive approach with these characteristics and show its predictive accuracy on a large dataset from KDDCup web logs (a commercial web site), while using fewer computational and memory resources. To achieve this objective, our approach is articulated around three notions: (1) Applying probabilistic exploration using Markov models. (2) Avoiding the problem of Markov model high-dimensionality and sparsity by clustering web documents, based on their content, before applying the Markov analysis. (3) Clustering Markov models, and extraction of their gravity centers. On the basis of these three notions, the approach makes possible the prediction of future states to be visited in k steps and navigation sessions monitoring, based on both content and traversed paths.

Modelling of the interactive application in term of scenario in a multimedia database

The interactive multimedia application must handle the different kinds of events as well as user interaction, duration, and start and finish of a medium. We present a method for modelling these events with object technology. The scenario that comprises the multimedia application is described in terms of event, condition and action expressions. The conceptual model is specified as a Petri net. The Petri net supports the action expressions that are triggered by the events if the conditions are respected. Then the Petri net permits the simulation and interpretation of the scenario. This object model is compared to the existing standard framework MHEG.

Some aspects of rule discovery in data bases

Rule Discovery in Databases integrates machine learning, probabilistic techniques and database concepts to learn a range comprehensible knowledge in sparse, noisy and redundant data. The discovery enables the learning of rules from data and extract their underlying structure. In this paper, we present the probabilistic index and the notion of minimal set of discovered rules which enhance runtime performance, improve discovery accuracy, resist noise, converges with the size of the sample, and eliminates coarse and redundant rules. This index can be used within the framework of an incremental discovery system. In other words, in this paper, we describe the rule intensity measurement which is an index that answers the question ‘What is the probability of having a rule of the form ‘IF premise THEN Conclusion’; the premise and conclusion are conjunctions of propositions ?’

Intelligent Image Clustering

We highlight a partition clustering method, which proposes an experimental solution to the famous problem of automatic discovery of the number of clusters (k). The majority of partition clustering methods consider the manual valuation of k. Manual valuation of k may be interesting for specific domains of applications where the expert has an accurate idea of the number of clusters he wants, however it is unrealistic for generic applications, and needs important estimation efforts without any insurance of their efficiencies.