The SeLeNe Project SeLeNe (Self e-Learning Networks) is an Accompanying Measure in the IST Action Line "Information and Knowledge Grids". The SeLeNe project is conducting a feasibility study into using semantic web technology for syndicating knowledge-intensive resources, such as learning objects. It is developing services for the discovery, sharing, and collaborative creation of learning resources, thus facilitating a syndicated and personalised access to such resources.

This is a report on the IST Workshop on Metadata Management in GRID and P2P Systems that took place on 16th December 2003 in London.

The report begins by introducing the aims of the Workshop, then gives a detailed account of the proceedings of each session and ends with overall conclusions from the Workshop. The appendices contain statistics on the speakers and participants, and the results of Workshop evaluation questionnaires completed by attendees.

As part of the dissemination of the SeLeNe project, and also as part of the integration role of the project as an Accompanying Measure, a Workshop was held in London on December 16th 2003. The target audience for the event were experts from the GRID, peer-to-peer and e-learning communities, as well as other application areas requiring GRID or peer-to-peer support.

The declared goal of the Workshop was "to identify recent technological achievements and open challenges regarding metadata management in novel applications requiring peerto-peer information management in a distributed or GRID setting". The discussions in all sessions focussed on models, services and architectures from the perspectives of the metadata level in general and for specific application domains.

The event attracted 43 participants (including those from within the SeLeNe consortium) from 8 different European countries, and we believe that it was an important step in coordinating research activities in this area. In order to disseminate the findings of the Workshop as widely as possible, participants' presentations have been made publicly available via the Workshop website:

http://www.ics.forth.gr/isl/ist workshop/ and also published on-line in a repository for Workshop proceedings at:

http:CEUR-WS.org

The Workshop programme was as follows. Details of the proceedings of each session can be found in Section 2. 2 Summary of Workshop Sessions

The first session of the Workshop addressed metadata management issues in Grid systems.

The session contained three presentations on related projects and issues. The presentations complement each other in that they present mutually exclusive issues; what metadata we have in real Grid systems, how we can provide Grid services using existing technology and how metadata can be used in new Grid domains. In all three presentations the need for metadata management middleware was quite evident.

The first presentation, given by Gavin McCance of the University of Glasgow, presented how DataGrid, the European flagship Grid project, addresses issues of metadata management. DataGrid is a file-based system manipulating quantum-physical, Earth-observation and biomedical data. The DataGrid project utilises two types of metadata: (i) Grid internal metadata on files and application logical names, and (ii) general application-specific metadata. Internal metadata have a relatively simple structure and their main use is for indexing purposes. The internal metadata is used to solve the "replica location problem", answering the query "given a logical file identifier, how do we find all the replicas of that file on the Grid?". The system that has been built is composed of a Replica Metadata Catalog and a Replica Location Service. It is based on bloom filters and allows quite a flexible search facility. DataGrid provides a simple Grid-enabled front end, it is up to the end-user (from one of the supported domains -physics, Earth-observation, biomedicine) to provide the specific application and to request the relevant files to manipulate. There is no generic application-metadata-management middleware that can handle generic application-type metadata.

The second talk was given by Savas Parastatidis of North East e-Science Centre and focussed on how Grid services can access an organisation's resources. He identified the need for resource metadata that can be published outside the organisation's boundaries and argues that this can be achieved via existing specifications and tools without the need to change the existing infrastructure. He presented two possible solutions, one based on Service Data Elements (SDEs), where metadata about resources is exposed through SDEs, and one based on the Grid Resource Metadata document (GRM). The claim made is that when using SDEs a number of changes need to take place: WSDL and resource descriptions are affected, new SDE-aware tools are required, Web services need added semantics, and an SDE-specific interface is necessary. However, if we utilise the concepts in GRM no changes need to take place. This is so because GRM is based on XML-Schema and no additions to WSDL or any other specifications are needed, existing tools work, it does not add semantics to a Web service (as it is just a document) and it could be published into a registry. However, it requires a particular Web service interface (a management interface or another interface) but other solutions can be built around it (e.g. peer-to-peer, registries, etc.). The main conclusion is that the GRM document provides a functional equivalent to SDEs that allows the re-use of existing technologies and work, which is not the case for SDEs.

The third talk was given by Mario Cannataro of the University "Magna Graecia" of Catanzaro. The "Knowledge Grid" was presented as a software environment that integrates data-mining techniques and Grid resources to build Grid-aware data-mining applications. The clear claim was that metadata (for describing data mining tools, data sources, Grid resources and ontologies for semantic modelling of the application domain) is a must if the Grid is to be utilised in knowledge management and data mining. The need for two types of metadata was exposed, the first type of metadata being for data-mining software and the second type for classical data-source metadata. The first type of metadata is semantically related to data mining and is split into two parts: (a) a description part, that is used to classify software and (b) a usage part, that contains the information needed by clients to access and use the software. In the system presented, metadata is utilised extensively to define execution plans and other processes. The presentation then showed how metadata and ontologies are managed to build and execute distributed data mining applications on the Knowledge Grid. It finally argues that metadata and ontologies are major players when complex applications are developed in a Grid environment.

The second session of the Workshop addressed metadata management issues in peer-to-peer systems. This session also contained three presentations on related projects and systems.

The first talk, given by Henrik Nottelmann (presenting work carried out jointly with Norbert Fuhr) of the University of Duisburg, focussed on the use of probabilistic logics for defining and using service descriptions in a peer-to-peer network with a large number of (Web) services. The goal of this work is to dynamically compute execution plans for services required to implement a given task. In this respect, the DAML-S upper ontology (a vocabulary for defining arbitrary services) is used for describing services, while probabilistic Datalog is used for match-making. A service is described in DAML+OIL by its profile, which is used for match-making, by its model, which describes how the service works, and by its grounding, which describes how to access the given service. A lower ontology for library services (e.g. search, schema mapping, or result modification services) is also introduced, containing definitions for generic search services and other query-and result-transformation services. Match-making rules were defined with probabilistic Datalog, which is a variant of predicate logic based on function-free Horn clauses and which defines probabilistic facts and rules. Match-making rules use facts derived from DAML-S directly to provide proper execution plans of services. By using the notion of chain (S1, S, S2) -i.e. a chain beginning with S1, ending with S2 and with S in between -two or more services with correct input/output can be chained to produce an execution plan. The probabilities are used as a primitive kind of cost estimation. The weight specifies the quality of the execution plan, so if its value is smaller than 0.5 then we have a quality loss, otherwise we have a quality improvement. The creation of a detailed lower ontology for library services, the introduction of composite services and the implementation of the project including grounding services are left as future work.

The second talk, given by Philippe Cudré-Mauroux (presenting work carried out jointly with Karl Aberer) of the EPFL and the Distributed Information Systems Laboratory, presented the Chatty Web approach for global semantic agreements. The problem posed is that of achieving semantic interoperability among heterogeneous data sources in a peerto-peer data management system, without relying on pre-existing global semantic models. The solution given involves the use of local translations that enable global agreements. Semantic 'gossiping' is used for query forwarding and distribution through the system. The right peers are selected and the "Per-Hop Behaviours" are query-dependent. An analysis is done between original and transformed queries based on intrinsic (syntactic distances) and extrinsic (semantic distances) measures. For the semantic similarity, query cycles can be detected and analysis of the results with content-retrieval techniques can be used. Consequently, a self-repairing semantic network can be organised with the use of evaluations based on Chatty Web simulations and the automatic correction of erroneous mappings based on gathered evidence. Finally, results concerning the system's scalability and sensitivity to TTL were presented and discussed.

The third talk, given by Wolfgang Nejdl of the University of Hannover and the Learning Lab, Lower Saxony (L3S), addressed issues related to data-centric networks and peer-topeer data management. Networks have evolved from host-centric to data-centric ones. In this context, the focus of research has shifted from the evaluation and optimisation of communication between network hosts to achieving worldwide physical and network data independence. Professor Nejdl presented the main results and challenges of several ongoing and past projects. REWERSE (Reasoning on the Web with Rules and Semantics) is an EU project addressing the issues of retrieving, protecting and integrating data. EDUTELLA has specified and implemented an RDF-based metadata infrastructure for peer-to-peer data networks. PROLEARN is a project that works towards innovative and interoperable e-learning resources and sustainable e-learning infrastructures and processes. The notion of schema and the use of RDF/S for describing distributed resources were identified in these projects as useful parts of a peer-to-peer data management system. Furthermore, RDF-QEL was presented as a Datalog-based Query Exchange Language that can be used to wrap other RDF and XML query languages. The HyperCup peer-to-peer topology and its broadcast algorithm were shown as an answer to the efficient routing problem in a peer-topeer system. Super-peer networks and their routing indices offer a new network architecture for distributing workload in a more efficient manner. Access control and automated trust negotiation were also identified as an important problem in order to protect resources from unauthorised access using credentials and access control policies. All the above issues represent essential building blocks for forthcoming schema-based peer-to-peer networks and peer-to-peer-based data management infrastructures.

The third session of the Workshop consisted of four presentations, each concerning metadata and its use.

The first talk, from Tamás Hauer, concerned the role of metadata in querying Gridresident medical images in the MammoGrid project, a EU-funded project aiming at a Grid solution for mammography and involving three hospitals among its partners. A federated system solution is proposed, whereby, with the help of shared metadata, a clinician can address a query to the system, which is then translated to remote local sub-queries whose results are then returned to the clinician. The specific characteristics of the medical domain (e.g. a user community which is heterogeneous and process-oriented rather than information-oriented, regional differences and a constantly changing knowledge base) necessitate flexibility and extensibility as well as management of domain information such as annotations of images. These and other considerations have led to a service-oriented architecture using Grid-middleware. In this context, metadata is used to describe service definitions. Each participating node is responsible for managing its own metadata and can change its service description on the fly, new sites can join seamlessly, and domain and service ontologies are defined independently.

The second presentation, from Bob Bentley, concerned metadata management in virtual solar observatories and reported experiences from the EGSO project, a Grid test-bed designed to improve access to solar data for the solar physics and other communities. The EGSO project addresses the generic problem of a distributed heterogeneous data set and a scattered user community. In such a setting, the availability of good quality metadata is important for searching. The architecture is defined in terms of roles: consumer, broker and provider. Resources are described by entries in a resource registry and managed by a broker. Brokers and registries are replicated to provide system resilience and permit load sharing. In order to provide an enhanced search capability, EGSO will improve the quality and availability of metadata, having a search registry describe all metadata available for search.

The third presentation, from Theo Dimitrakos, concerned metadata management issues underpinning emerging solutions for distributed trust and contract management and enforcement in enterprise Grid and peer-to-peer systems. Metadata management issues underlie a number of activities across e-science and information technology. In support of this claim, the speaker presented in some detail four projects where metadata play an important role:

• GRASP, Grid-based application service provision;

• CORAS, a CASE tool and method support for security-risk analysis;

• SWAD-Europe, Semantic Web technology development;

• PELLUCID, an agent-based platform supporting organisational mobility.

The final presentation, from Stavros Christodoulakis, also concerned metadata management, this time for audiovisual content to support intelligent video-content retrieval and e-learning services in digital TV (t-learning). These objectives are pursued in the context of TV-Anytime framework, MPEG-7 and SCORM. An ontology-driven framework for semantic metadata management was presented and the following points were stressed:

• TV-Anytime keywords are the only means to describe program segments;

• MPEG-7 semantic model is used to build domain specific ontologies;

• a coupling of OWL and MPEG-7 has been implemented;

• ontologies are used for filtering and retrieval of MPEG-7 multimedia content;

• semantic annotations are transformed into TV-Anytime segment keywords. With respect to t-learning, two issues were discussed:

(1) providing interoperability for educational applications in different e-learning and digital TV environments;

(2) creation of metadata for digital TV for educational purposes in order to offer educational experiences exploiting usual TV programs.

The fourth session of the Workshop consisted of four presentations, all on e-learning applications that rely on some form of schema-based repository of Learning Object (LO) metadata.

The first talk, from Klaus Jantke, gave an outline of the DaMiT e-learning system developed to provide a tutorial on data mining, and went on to discuss particular problems posed by the development of the system and aspects that may distinguish it from other projects.

The system allows users to specify several of their preferences as a "profile", including the type of material preferred (example-oriented, theory-oriented, etc.), style (formal, informal) and language (although the system is mostly German-only at present). The e-learning content delivered by the system is stored as a collection of fragments of learning material, each described by associated metadata. DaMiT generates complete pages of learning material from these fragments "on the fly", adapting the page to the user's needs based on their profile. DaMiT also includes a fully-functional e-payment system that uses metadata about the user to keep track of their entitlements to access to the learning material (students at DFKI can access all material for free, but others must pay to use the resource).

Professor Jantke believes that future e-learning systems will include intelligent assistants that learn about, understand and adapt to the user, to provide them with a personalised learning experience. Metadata of several kinds will be needed to support this adaptive functionality, including the kinds used in DaMiT:

(1) metadata about each fragment of content;

(2) "storyboards" that aid in the structuring of content fragments;

(3) metadata on security and payment options for users.

The second presentation in the session, from Kevin Keenoy (of work carried out with the other SeLeNe consortium members), gave an outline of the SeLeNe project, focussing on the metadata used and aspects of the project related to personalisation. SeLeNe aims to provide a tool that enables the discovery, sharing and collaborative creation of LOs within a learning community. LOs to be shared within the community are described using an extended form of the IEEE-LOM metadata schema, encoded in RDF, and these descriptions together form a distributed repository of metadata that can be queried by users searching for learning material. The metadata entries make use of taxonomies of subject and topic domains, learning objectives and learning styles.

As well as metadata describing the LOs (LO descriptions), the system manages metadata describing the users of the system (user profiles), to enable adaptive personalisation services to be provided by SeLeNe. The user profiles are also encoded using RDF, and make use of the same taxonomies as the LO descriptions. Both the LO descriptions and user profiles adapt and evolve over time, either automatically (e.g. the user profile records the history of LO accesses by the user) or as users add and amend the data manually (e.g. a LO author may add additional information to one of her LO descriptions).

Three main areas of personalisation in SeLeNe were identified: (1) Personalised views of the RDF LO metadata, defined using RDF View Language (RVL), are supported;

(2) Personalised query results are generated, based on the user profile;

(3) Advanced notification services are provided, allowing individuals to be informed of changes to the metadata repository of interest to them. These are specified using an ECA language for RDF.

Several unresolved issues were identified as areas for future research, including peer-topeer query processing over RDF, the evaluation of algorithms for personalised ranking of query results, the combination of ECA rules with transaction and consistency maintenance in RDF repositories, and the design of user interfaces enabling easy access to such advanced personalisation services.

The third presentation, from David Massart (of work carried out jointly with Frans Van Assche), described work done by the Celebrate (Context e-Learning with Broadband Technologies) project in developing a system providing metadata search and exchange of LOs, creating a European Learning Network (ELN) that is used by 500 schools across Europe.

The ELN is built around a brokerage system that manages exchanges between its members, enabling interoperability between e-learning systems by searching and exchanging LOs in their repositories. The IEEE-LOM schema is used for LO descriptions. Various communication protocols may be used between different ELN members, but the individual systems are shielded from this complexity by the "ELN client" that provides a simple communication API.

Unlike other virtual learning environment networks based either on a client-server or on a peer-to-peer architecture, Celebrate is based on a mixed approach where, although the brokerage system hosts a central metadata repository, each ELN member is also authorised to manage its own local metadata repository. Search requests are both handled centrally and propagated to local repositories.

The final presentation of the session, from Zoltán Miklós (of work carried out jointly with Bernd Simon), described work done in the ELENA project, "creating a smart space for learning".

ELENA, like both SeLeNe and Celebrate, is concerned with the sharing of learning resources that are contained in heterogeneous, distributed repositories. In Smart Spaces for Learning metadata repositories are connected in a peer-to-peer fashion, and Semantic Web techniques are used to achieve interoperability. Personal Learning Assistants support learners in selecting appropriate learning services from a large number of available sources.

The metadata for learning services is based on a common formal ontology, but learners do not need to be familiar with the domain ontology in detail as they can define views of the ontology using the TRIPLE language. These views allow queries to be formulated in terms of a user-specific ontology. A Simple Query Interface translates queries into an appropriate form, establishing interoperability among heterogeneous learning repositories.

The presentations in this session brought to light several core issues that arise for systems providing personalised e-learning. In systems that employ a user profile two main issues are content and representation: what information about the user is needed, and what format should it be stored in? There is also the issue of where the information for the profile comes from -is it user-supplied, teacher-supplied, automatically derived somehow, or some combination of all these three?

Once answers to questions about the generation and maintenance of the profile have been decided, the remaining issue (and perhaps the issue that has so-far been least explored) is the resolution of the profile with LO metadata -how can the profile be used to provide the most useful personalised LOs or personalised ranking? It seems that there is scope for much more work in empirically testing different matching algorithms in different learning situations -present methods generally seem to be "sensible" ad-hoc choices rather than empirically validated solutions.

Another area requiring future investigation is learning style taxonomies -which ones are most useful in the context of e-learning, and what the relationships are between different learning style taxonomies.

From the above summaries of the four Workshop sessions, we see three sets of research challenges emerging in the area of metadata management for Grid and peer-to-peer systems:

1. Challenges stemming from the distribution, autonomy and heterogeneity of information and services, leading to the need for:

• metadata describing information and services available at the nodes of the Grid/P2P system;

• searching and matching techniques utilising this metadata for discovery of information and services;

• metadata and techniques for controlling access to, and privacy of, information and services;

• techniques for automatic service composition and orchestration;

• metadata and techniques for translation of queries and query results in the absence of a controlled global schema;

• metadata supporting replication and consistency of information;

• automatic/semi-automatic extraction of metadata from large volumes of heterogeneous data.

2. Challenges arising from the dynamicity of Grid/P2P environments, e.g. changes in the network topology, information content at nodes, and service availability at nodes:

• this leads to the need for the techniques being developed under (1) to be scalable, adaptive, extensible and fault tolerant.

3. Challenges arising from the heterogeneity of users accessing Grid/P2P systems:

• this leads to the need for techniques that personalise content and presentation to different users' needs and preferences.

During the discussion at the conclusion of the Workshop it was generally agreed that the Workshop had been very beneficial and timely in bringing together these various common strands of research from the Grid, P2P and applications communities, and that a similar event in a year's time should be aimed for. We very much enjoyed organising and hosting this Workshop, and we plan to pursue the possibility of a similar, possibly larger, event next year.

We expected the number of participants external to the SeLeNe consortium to be around 75% of the total attendance, and so set a minimum of 30 external attendees out of 50 as a target. As can be seen from Table 1, the target of 30 external attendees was met, despite there being slightly fewer participants overall than anticipated.

We set a target that a minimum of 50% of attendees should be from the IST community. In calculating the figures we have excluded members of the SeLeNe consortium, who also are obviously "IST participants". As can be seen from Table 2, we met our target for IST participation in the Workshop.

The targets we set for European representation were that participants would be from at least 6 different countries, and that about 60% of the attendees would be from outside the UK.

In fact, 8 different European countries were represented at the Workshop: Austria, Belgium, Cyprus, France, Germany, Greece, Italy and the UK, meaning that our target for Non-SeLeNe All participants participants United Kingdom 18 (56%) 25 (58%) Non-UK 14 (44%) 18 (42%) Total 32 43

Table 3: UK vs. non-UK participation in the Workshop the number of countries represented was exceeded. However, as can be seen from Table 3, the overall percentage of non-UK participants fell short of our target. This is likely to be the result of higher expenses for participants from abroad (who had to meet their own travel costs).

As this was an IST event, we set a target for a minimum of 50% of the speakers to be from, and to present results of, relevant IST projects. As can be seen from Table 4, the proportion of talks presenting IST project results was above our target.

Participants at the Workshop were asked to complete two evaluation questionnaires to assess the quality of the event -one at the end of the morning and one at the end of the afternoon sessions. The afternoon questionnaire contained additional questions about the quality of the Workshop as a whole. Both questionnaire forms can be found in Appendix C.

The responses (only from participants who are not part of the SeLeNe consortium) are summarised in the sections below. The results are compared with the target that we set ourselves:

We expect to obtain an overall average score of at least 4 for each of the morning and afternoon sessions and for the Workshop as a whole.