Many Web Usage Mining methods rely on clustering algorithms in order to produce homogeneous classes of documents (when the content of the web site is analyzed) and/or of users (when browsing behaviors are analysed). Information extracted from web server logs are complex and noisy, but can be used to define usage based dissimilarities between users of the site or between pages of the site. There are however many possibilities to define this type of dissimilarity measure.

We have defined a benchmark site that allows to compare dissimilarity measures via the clustering results they produce. The benchmark consists in one year of log of the web site of the CIn, the laboratory of Francisco De Carvalho. This site is small (91 pages) and is very well organized. This allows to define a meaningful semantic structure and to build an expert partition of its content. Then, this expert partition can be compared to the results of a clustering algorithm applied to the dissimilarity matrix constructed with a specific measure.

The results that will be published in the EGC 2006 conference show that the Jaccard index and the ``term frequency inverse document frequency'' approach obtain quite good results, whereas the cosine measure performs badly. It seems also that better results could be obtained by taking into account the structure of the site together with usage data.

In the context of a research project with India and some others French partners, about a hundred ancient manuscripts written in Sanskrit, all arisen from the same text (the Benares Glose), should be compared in order to make a critical edition and provide some classification between the manuscripts (cf. section  ).

During his internship, S. Tandabany developed some tools that were required for clustering homogeneous XML Sanskrit Documents. First, a modified longest common substring algorithm is proposed to deal with Sanskrit characters. Then, as in Sanskrit inversions of characters are not always meaningful, a detection of possible inversions is applied. Finally, the Agglomerative 2-3 Hierarchical classification (cf. section  ) is used as the classification algorithm. To do this, we proposed a new distance between texts taking into account some Sanskrit specificities and allowing the addition of meta-data (worn state and shape of the manuscripts as objects, annotations about words forgotten, etc.). The text is splitted into paragraphs, ``sub-distances'' are computed between each corresponding paragraph, taking into account adds, deletions, transformations and inversions. Then, some constraints needed to obtain a distance (triangular inequality) are removed to get a dissimilarity instead of a distance for the 2-3AHC. The impact of these modifications on the classification was analysed. Finally, our results are highlighted with some experiments and examples .

In the context of the MobiVIP project (cf. section  ), during Rémi Busseuil's internship we studied the possibility of clustering tourists itineraries in the town center (Antibes in this case). Thus, a software for tourist itineraries generation and clustering was developed (in Visual .NET), taking into account not only the geographical characteristics of an itinerary, but also the symbolical ones: street type, buildings type, etc. This new use of semantical data, opens new directions for the road itineraries recommendations, by addressing new issues like the purpose of the itinerary or the nature of the crossed areas.

Clustering itineraries has many advantages besides the possibility of choosing the most suitable one: it is also an analysis and comparison tool. This can have multiple applications: route or destination prediction, traffic anticipation, etc. As clustering algorithm, we used the Agglomerative 2-3 Hierarchical Classification (2-3 AHC) algorithm. The 2-3 AHC has the advantage of being easily visualized compared to a classical clustering method.

In order to compare different itineraries, we basically divided each itinerary into fragments and then we computed a distance/dissimilarity value using the Longest Common Subsequence algorithm (LCS) and a spread function developed in  . Different ways of defining the dissimilarity and of comparing the fragment were tested.

An example of clustering 40 itineraries issued from 4 different profession types is presented in Figure bellow.

The main goal of the Semantic Web is to ease a computer-based data mining and discovery, formalizing data that is mostly textual. Our approach is different as we are concerned in the way Web sites are constructed, taking into account their development and their semantics. In this respect we are closer to what is called content management.

Our formal approach is based on the analogy between Web sites and programs when there are represented as terms, although differences between Web sites and programs can be pointed out :

We are developing a specification language to express global constraints in Web sites or in a collection of XML documents in an operational way.

An initial version of this language as been described in , together with its application to a real sized collection of documents: the Inria scientific activity report for the years 2001 and 2001.

The language and its implementation has been developped and improved in 2005, in particular for efficiency. At the same time our XML core parser has been extended to allow parsing of HXTML documents.

The same language has been used to extract information in XML documents. This has been the case to choose and extract words from different part of XML documents. These words was first passed to a tagger, then used to cluster the different documents. A first experiment has been done in 2004 and has been presented to the EG2005 conference , (cf. section ).

In a more long term experiment, we have initiated a regular monitoring of the Inria activity reports to see how the number of bad URLs in these reports evoluates. This monitoring, started in december 2004 and then performed every two weeks, takes into account the activity report for 2002, 2003 and 2004.

This work was conducted in the context of the PhD of A. Baldé.

A huge volume of data is produced by many applications. Data mining techniques are part of knowledge discovery methods whose aim is to discover knowledge in large databases without predetermined information about the application field which is well-known as KDD. But data mining is a complex process for an end-user and the main difficulties consist in the interpretation of the results. Metadata can help the interpretation process by providing additional information. Our objective is to facilitate the interpretation process and to point out that metadata can play a major role for this purpose. In spite of the visual representation of the results, the user should acquire a significant experience to be able to interpret the clusters. Data mining tools generally offer visualization modules which are not adapted to analysis. The original contributions of our work made in collaboration with Marie-Aude Aufaure (Supelec) concern new approaches to representing clustering's metadata and interpreting clustering's results by using metadata.

First, we propose a metadata model that could be automatically exploited . We also propose a tool in order to help the end-user to interpret the clusters obtained. This tool is based upon the architecture described in Figure  .

This architecture is composed by three layers: metadata model, metadata manager which manages metadata extraction and storage and manipulations performed on these metadata and user query layer using XQuery. In order to implement these queries, we use the Saxon processor. Saxon is a set of tools dedicated to XML documents processing: it has established a reputation for fast performance, the highest level of conformance to the W3C specifications. This method can be applied to a wide variety of data mining methods.

This work was performed in the context of the PhD of H. Behja (France-Morocco Cooperation - Software Engineering Network).

Our goal is to make more explicit the notion of "viewpoint" from analysts during their activity and to propose a new approach to integrate the viewpoint notion in a multi-view Knowledge Discovery from Databases (KDD) analysis. We define a viewpoint in KDD as an analyst's perception of a KDD process, perception referring to its own knowledge  . Our purpose is to facilitate both reusability and adaptability of a KDD process, and to reduce his complexity whilst maintaining the trace of the past analysis viewpoints. The KDD process will be considered as a view generation and transformation process annotated by metadata to store the semantics of a KDD process.

In 2004 we started with an analysis of the state of the art and identified three directions: 1) the use of the viewpoint notion in the Knowledge Engineering Community including object languages for knowledge representation, 2) modelling KDD process adopting a Semantic Web based approach and 3) the use of annotations of KDD processes. Then we designed and implemented an object oriented platform for KDD processes including the viewpoint notion (via design patterns and UML using Rational Rose). The current platform is based on the Weka library.

In 2005 we proposed and implemented the knowledge conceptual model integrating the viewpoint concept (cf. Figure  ). It is composed of four models structured in two types of knowledge:

First, for the domain knowledge, the domain model that describes the analyzed domain knowledge in terms of objects, attributes, data, etc. and the analyst domain knowledge that will relate to the tasks carried out by the analyst; choice of methods, variables, etc. We propose a formal representation of the domain model as a datawarehouse that allows the business information to be viewed from many viewpoints. For our example of the HTTP log in Web Usage Mining (WUM), the used database design is the star schema.

Second, for the strategic knowledge, we find:

This work is accepted for publication in january 2006 in a special issue on ``Méthodes Avancées de Développement des SI'' of the french journal ISI (D. Rieu and G. Giraudin editors).

A critical edition is the edition of a well known text taking into account all possible versions of this text. Critical editions are particularly needed for texts issued from manuscripts where the variations can be very significant from one manuscript to the other.

Production of critical edition of Sanskrit Text.This is particularly important for the Indian subcontinent where at least one third of the manuscripts existing through the whole world are supposed to be found, the main part of them being written in Sanskrit. It was not an Indian tradition to deal with critical edition, so very few of them exist at present time. The idea is to provide a computer assisted construction of critical edition. Such tools exist for occidental languages, but, due to some Sanskrit specificities they could not suit our purpose:

In order to avoid the complexity problem induced by the thousands of characters sentences, and to be able to provide to the philologist the exact words where a difference occurs, we need either a lexicon, or a text where all the words appear separately. We will use the second solution and we call such a text a Padapatha according to a certain form of recitation used by the Sanskritist. Due to the sandhi, such a text is not readily comparable with the manuscript text. We must provide a pre-processing based on LEX to construct all the sandhi related to the padapatha in order to be able to provide a suitable comparison. For the comparison we use the Longest Common Subsequence (LCS) algorithm based on dynamic programming. The algorithm output will be used as input of the following project steps :

Electronic display of critical edition of Sanskrit Text.Traditionally critical editions are represented as particularly boring (for unfamiliar) books, where the text itself is very small, and where the notes are numerous and enormous. It's a philologist dream to see only the points they care about. An electronic form of critical edition could be a proper answer.

But there are a lot of problems related to the Sanskrit we have to care about. Thanks to Unicode it is now possible to get some standard about the display of sandhi characters. But because of the ligature formation, two Sanskrit characters separated by a blank do not look similar as if they were put one after the other.

Marina Dufresne during her internship (cf. section  ) developed a software tool what allows an interactive display of critical edition of Sankrit text starting from an XML text. This tool is not perfect but it has been greatly appreciated by the Sanskrit community. This work has bee done in collaboration with François Patte.

The standard Self Organizing Map (SOM) is restricted to vector data from Rⁿ. In our previous work , , we proposed an adapted version of the SOM, called the DSOM (for Dissimilarity SOM) that can be applied to any data for which a dissimilarity can be defined.

In 2005, we improved DSOM by defining a new algorithm associated to an improved implementation. This implementation significantly reduces the execution time of the method without changing the results . The new algorithm is based on a factorization technique applied to the computation of the criterion optimized by the method (the sum of weighted dissimilarities between a prototype candidate and the data to cluster). It is associated to an early stopping scheme and to some memorization techniques that leverage the iterative nature of the method.

We have also applied the DSOM to usage-based web content clustering and visualization , see section for details.

Functional Data Analysis is an extension of traditional data analysis to functional data. In this framework, each individual is described by one or several functions, rather than by a vector of Rⁿ. This approach allows to take into account the regularity of the observed functions.

In 2005, we have extended our approach based on neural methods for functional data to the case of Support Vector Machines (SVMs) applied to function classification:

In 2005, our earlier work on functional multi-layer perceptrons has been published in international journals , , .

The main contribution is the proposal of a new partitional dynamic clustering method for interval data based on the use of an adaptive Hausdorff distance at each iteration. The idea of dynamical clustering with adaptive distances is to associate a distance to each cluster, which is defined according to its intra-class structure. The advantage of this approach is that the clustering algorithm recognizes different shapes and sizes of clusters. Here the adaptive distance is a weighted sum of Hausdorff distances. Explicit formulas for the optimum class prototype, as well as for the weights of the adaptive distances, are found. When used for dynamic clustering of interval data, these prototypes and weights ensure that the clustering criterion decreases at each iteration.

Let be a set of nobjects indexed by iand described by pinterval variables indexed by j. An interval variable X is a correspondence defined from in such that for each i, X( i) = [ a, b] , where is the set of closed intervals defined from , i.e., = {[ a, b]: a, b, ab}. Each object iis represented as a vector of intervals , where x_i^j= [ a_i^j, b_i^j] .

An interval data table { x _i ^j} _{n× p}which is used by our clustering method is made up of nrows that represent nobjects to be clustered and pcolumns that represent pinterval variables. Each cell of this table contains an interval x_i^j= [ a_i^j, b_i^j] . In our approach a prototype y_kof cluster C_kPis also represented as a vector of intervals , where y_k^j= [ _k^j, _k^j] .

It is now a matter of choosing an adaptive distance between vectors of intervals and properly defining the representation step of the dynamic algorithm with adaptive distances given in the previous section. In other words, we will give an explicit formula for the prototype y_kand for the vector of weights _kthat minimizes both the adequacy criterion .

This work was conducted in the context of the PhD of S. Chelcea.

We have continued this year our study of the Agglomerative 2-3 Hierarchical Clustering , as a part of Chelcea Sergiu's PhD thesis. A study of different aggregation indexes and cluster indexing measures combined with the 2-3 AHC algorithm execution has revealed a particular case of clusters merging, which can influence the resulting induced dissimilarity matrix. This case that we denoted blind merging, is present when two clusters are merged whilst one of them is not maximal. Based on our previous theoretical study , the next (intermediate) merging, will merge together two clusters possibly at a high indexing degree. This can be avoided by minimizing the final cluster's indexing degree when choosing the two merging clusters.

A slightly modified version of a 2-3 AHC algorithm was proposed and implemented in order to avoid such situations. The interest of this new 2-3 algorithm variant is its resulting induced dissimilarity matrix which is ``better'' or equal to the classic ultrametric.

We experimentally validated this new 2-3 AHC algorithm variant on different artificial datasets and we also integrated it into our Hierarchies Visualization Toolbox http://axis.inria.fr/.

This new 2-3 AHC algorithm variant was also applied and validated on other types of datasets: on Web Usage Data , on Sanskrit XML documents (see also Section ) and on tourists itineraries .

This work was conducted in the context of the master of A. Marascu.

In recent years, emerging applications introduced new constraints for data mining methods. These constraints are particularly linked to new kinds of data that can be considered as complex data. One typical kind of such data is known as data streams. In a data stream processing, memory usage is restricted, new elements are generated continuously and have to be considered as fast as possible, no blocking operator can be performed and the data can be examined only once.

At this time and to the best of our knowledge, no method has been proposed for mining sequential patterns in data streams. We argue that the main reason is the combinatory phenomenon related to sequential pattern mining. Actually, if itemset mining relies on a finite set of possible results (the set of combinations between items recorded in the data) this is not the case for sequential patterns where the set of results is infinite. In fact, due to the temporal aspect of sequential patterns, an item can be repeated without limitation leading to an infinite number of potential frequent sequences.

The SMDS (Sequence Mining in Data Streams) method, proposed in , , is designed for extracting sequential patterns from a data stream. More precisely, our goal is to extract significant patterns that will be representative of Web usage streaming data. To this end, SMDS performs as follows:

All those steps have to be performed as fast as possible in order to meet the constraints of a data stream environment. Approximation has been recognized as a key feature for this kind of applications, explaining our choice for an alignment method for extracting the summaries of clusters. The SMDS method is illustrated in figure . SMDS has been tested over both real and synthetic datasets. Experiments could show the efficiency of our approach and the relevance of the extracted patterns on the Web site of Inria Sophia Antipolis.

The analysis of the content of a web site based on usage data is an important task as it allows to obtain insight on the organization of the site and of its adequacy to user needs. The (dis)agreement between the prior structure of the site (in terms of hyperlinks) and the actual trajectories of the users is of particular interest. In many situations, users have to follow some complex paths in the site in order to reach the pages they are looking for, mainly because they are interested in topics that appeared unrelated to the creators of the site and thus remained unlinked. On the contrary, some hyperlinks are not used frequently, for instance because they link documents that are accessed by different user groups.

In 2005 we have studied two general tools for visualizing the content of a site based on usage data:

In the context of the ECML/PKDD 2005 Discovery Challenge, we improved  , our preprocessing methodology for intersites Web Usage Mining  . A clickstream dataset was proposed in the Discovery Challenge this year for the first time. The dataset consisted in requests for page views on seven different e-commerce Web sites from the Czech Republic. A request contained a PHP SessionID automatically generated for each new user visit on each server (unique IDs).

Based on Tanasa's preprocessing methodology , we defined a new methodology to preprocess the provided datasets and to store it in a data warehouse. Since a user changing shops can have (during a single visit) multiple SessionIDs, one on each shop, we regrouped these PHP SessionIDs into intersite users visits. More precisely we regrouped SessionIDs belonging to a single user (same IP) into a Group of SessionIDs, corresponding to the user's actual (intersite) visit. This was done by comparing the Referrer with the previously accessed URLs (in a reasonable time window), each time the user moves to another shop. We thus reduced by 23.88% the number of user visits.

To analyze the traffic load in the seven shop sites, we grouped the requests in terms of Time Periods(slices of date and hour). We cross-clustered these time periods against the visited products using a generalized dynamic algorithm .

The result consisted in the confusion table containing classes of periods and products (see Figure  ).

Such analyses allow us to identify best hours for marketing strategies, like fast promotions, online advices and publish banners, etc. Others analyses could be planned in the future, exploiting for example the link between the consumer activities and the time periods by shop or focusing on multi-shop user visits, etc.

In fact we apply a previous work published in 2003: indeed we proposed in a crossed clustering algorithm in order to partition a set of objects in a predefined number of classes and to determine, in the same time, a structure (taxonomy) on the categories of the object descriptors. This procedure is a simultaneous clustering algorithm on contingency tables. The convergence of the algorithm is guaranteed at the best partitions of the objects in rclasses and of the categories of the descriptors in cgroups, respectively. This algorithm extended the dynamical algorithms hereafter proposed in the context of the Web Usage Mining. In particular, we had already performed it on the Web Logs Data, coming from the HTTP log files of INRIA web server .

This work has been done in collaboration with the LGI2P and the LIRMM (see ) and has been published in . Existing Web Usage Mining techniques are currently based on an arbitrary division of the data ( e.g.``one log per month'') or guided by presumed results ( e.g``what is the customers behaviour for the period of Christmas purchases?''). Those approaches have two main drawbacks. First, they depend on this arbitrary organization of the data. Second, they cannot automatically extract ``seasons peaks'' among the stored data.

The work presented in this section performs a specific data mining process (and particularly to extract frequent behaviours) in order to automatically discover the densest periods. Our method extracts, among the whole set of possible combinations, the frequent sequential patterns related to the extracted periods. A period will be considered as dense if it contains at least one frequent sequential pattern for the set of users connected to the Web site in that period.

Our method is based on:

The third step is based on Perio, the heuristic we have developed for that purpose and it is widely inspired from genetic algorithms.

XML documents are becoming ubiquitous because of their rich and flexible format that can be used for a variety of applications. Standard methods have been used to classify XML documents, reducing them to their textual parts. These approaches do not take advantage of the structure of XML documents that also carries important information.

Last year we studied the impact of selecting different parts (sub-structures) of XML documents for specific clustering tasks. Our approach integrated techniques for extracting representative words from documents elements with unsupervised classification of documents. We illustrated and evaluated this approach with the collection of XML activity reports written by Inria research teams for year 2003. The objective was to cluster projects into larger groups (Themes), based on the keywords or different chapters of these activity reports. We then compared the results of clustering using different feature selections, with the official theme structure used by Inria between 1985 and 2003, and with the new one proposed officially in 2004. The results (published this year) show that the quality of clustering strongly depends on the selected document features , .

This year we developed a new representation model for clustering XML documents. The standard vector model for classification or clustering of documents represents documents by weighted vectors of words contained in the documents. This model takes into account only the textual content of documents. With XML documents, we want a representation that takes into account either the structure of the documents or both the structure and the content. Since XML documents can be seen as trees, we represent documents by the set of their (node) paths of length L, n < =L < = m, n and m being two given values. Paths can be contrained to be root-beginning paths, or leaf-ending paths. For dealing with both the structure and the content, we define text pathsthat extend the node paths with the word contained in the subtree of their final node. Then by regarding paths as words, we can cluster documents by applying standards clustering methods based on the vector model. There is one difficulty, though, since the vector model is based on the independance between the dimensions of the vectors. In our case, when two paths are embedded in each other they are obviously not independant. To deal with this problem of dependency, we partition the paths by their length and treat each set of paths as a different modality in the clustering algorithm.

We evaluate our approach using four standard metrics, namely the F-measure, the Corrected-Rand, the entropy and the purity. That for a given clustering task, we compare the resulting clusters with a priori known classes. We made several experiments using the INEX IEEE collections and INRIA activity reports . The results that will be published in the EGC 2006 conference show that our approach works both for structure-based clustering and Structure-and-content clustering. However, using leaf-ending paths may result in damaging the clustering time, as the number of paths increases dramatically. We need to find good ways to reduce the number of paths, especially for text paths.

We also started to apply this approach to the collections proposed by the INEX XML Document Mining tracks .

The goal of this work is to provide a classification (``classification supervisée'' in french) over a collection of XML documents. For this purpose we consider that we are provided with a set of clusters coming from a previous clustering on an past collection. More formally: let us consider S₁a first collection of XML documents and C= { c₁, c₂, ... c_n}the set of clusters defined for the documents of S₁. Let us now consider S₂a new collection of XML documents. Our goal is to provide a classification on S₂by taking into account the distribution of documents in C.

To this end, our method will perform as illustrated in figure . It is based on the following three steps:

The matching techniques developed in this work are described in .

When searching information from structured documents collections such as XML collections, it is expected that the use of the structure will help in two ways:

In the context of INEX (an International Initiative for the Evaluation of XML search), we are interested in evaluating what granularity of elements the users find relevant. We first analysed the relevance assessments to identify the types of highly relevant elements and identified three retrieval scenarios: Original, generaland specificand compare the performance of three different systems - a native XML database, a full text retrieval engine and a hybrid system -, for those different scenarios . We then developed a novel retrieval heuristics that dynamically determines the preferable units of retrieval called Coherent retrieval Element .

In , we analyse and compare the assessors'judgement on the relevance of returned document components with the users' behaviour when interacting with components of XML documents. By analysing the level of agreement between the assessor and the users, we show that the highest level of agreement is on highly relevant and on non-relevant document components, suggesting that only the end points of the INEX 10-point relevance scale are perceived in the same way by both the assessor and the users.