Introduction
Wikis provide flexible ways for supporting the quick and simple creation, sharing and
management of content. Based upon the established wiki-technology, semantic wikis
(e.g., [
1, 2
]) enhance this by providing enriched content and features. For example,
flexible inline queries and according results that are generated based on these dynamically
are such prominent features. While the queries and answers (results) can be flexibly
handled by the system, and can usually be formalized as textual content, the wiki
system also provides appropriate means for the persistent storage and management of the
generated content.
Semantic data mining systems enable the inclusion of a large set of background
knowledge, for example, in order to access knowledge services, for selecting the
applied data mining methods, or for postprocessing the obtained data mining results. Thus,
integrating wikis is a convenient option for semantic data mining systems, since the
semantic core components can support the semantic mining features, while the wiki
component provides for a convenient front-end and user-management, enables the persistent
storage of queries and mining results, and supports their extended annotation.
This paper presents a wiki-enabled approach for collaborative semantic data
mining: The semantic data mining system VIKAMINE [
3
] is combined with the semantic
JSPWiki (http://www.jspwiki.org) extension KnowWE [
1
]. We describe the
interaction and exchange of query and results data, and the integration of semantic
information and knowledge.
The rest of the paper is structured as follows: Section 2 describes the basics of
semantic data mining, provides an overview of the presented approach, and describes
its implementation. Section 3 concludes with a summary and interesting directions for
future work.
Method
This section briefly introduces the general semantic data mining approach. After that,
we first give a general overview, before we describe the architecture of the proposed
approach in detail. Finally, we discuss related work.
2.1
Semantic Data Mining
Semantic data mining can be considered as an approach utilizing formal methods and
techniques in order to explicitly integrate data semantics, background knowledge, or
reasoning in the mining process. The knowledge is typically represented in a
knowledge repository, such as an ontology, or a knowledge base. The main aspect of
semantic data mining is the explicit integration of this knowledge into the data mining and
knowledge discovery process, where the algorithms for data pre-processing, mining or
post-processing make use of the formalized knowledge to improve the overall process.
There has been growing interest in this issue, e.g., [
4–6
], in various domains, especially
in the medical domain [
4, 7, 8
].
With the advent of the semantic web and standardized knowledge representations of
semantic web techniques, e.g., the web ontology language OWL, utilizing these
knowledge representation formalisms for data mining is a promising direction for task design,
evaluation and refinement, as discussed below. In the following, we outline the different
aspects of semantic data mining, and discuss their implications.
The general data mining process can be structured along the CRISP-DM process
model (http://www.crisp-dm.org ) and consists of the following phases: (a)
Business understanding, i.e., understanding the application domain, (b) Data
Understanding, i.e., considering the (potential) objects of analysis, (c) Data Preparation, e.g.,
preprocessing and schema-matching of the data elements, (d) Modeling, e.g., given by
concrete mining sessions, (e) Evaluation, i.e., assessment of the mined models, (f)
Deployment, i.e., putting the extracted knowledge into action. The semantic data mining
approach integrates ontologies in each of the six steps [
7
]. In the following, we
provide examples for each of the phases structured along the dimensions of task design,
evaluation, and refinement.
– Task Design:
• In the Business Understanding phase ontologies help inexperienced users
getting accostumed to the domain, by structuring the relations between the
concepts, and explaining the concepts in terms of their properties.
• In the Data Understanding phase, important data elements (contained in the
ontology) need to be selected. Then, missing attributes, or redundant attributes
can be added or removed from the data set. This can be accomplished by a
datato-ontology mapping step [
5
] where the data elements are mapped to concepts
of the ontology, e.g., for integrating heterogenous data.
• The Data Preparation phase is strongly connected to the Modeling phase.
Depending on the latter, for example constraints on attributes or values can be
derived. This concerns constraints on the relations between the attributes, as
described in [
4
], for example, grouping constraints or exclusion constraints for
certain attribute groups that should not be considered. A further possible
inclusion of the ontology is given by a more abstract task composition phase, for
which the modeling phase can be hierarchically decomposed along the
generalization/specialization hierarchies modeled in the ontology. Then, more concise
results can potentially be obtained on lower levels, but for efficiency reasons
higher levels can be considered first and be used for filtering interesting
hypotheses in an earlier stage, cf., [
5
].
– Task Evaluation:
• During the Evaluation phase (of CRISP-DM), the discovered patterns can
be interpreted and explained in a structured way using the concepts and/or
contained patterns. Various post-processing options are available at this point,
cf., [
5
]. Specifically, due to the data-to-ontology mapping, the discovered
patterns can be matched to semantic relations or more complex relations between
these. Additionally, such knowledge provides a potential (explaining) context
for the discovered patterns. Furthermore, prior knowledge can be compared to
the patterns, e.g., for confirming known relations, identifying new knowledge,
and/or detecting exceptions and conflicts with formalized expectations.
Concerning possible explanations, causal relations can often help in this respect,
for validating and confirming discovered patterns, or for their analysis.
• The Deployment phase concerns the integration of the discovered models into
the business setting. It is easy to see, that for distributed processing and storage
(e.g., on the semantic web) a shared ontology is inevitable. This is especially
relevant for deploying results as semantic analytic reports (an extension of
analytic reports [
5
]), described below. In a late evaluation step, the
models/patterns can be tested during their practical application. In that case, the persistent
sessions stored in the wiki provide direct access in a collaborative manner.
– Task Refinement: The task refinement step is activated after the evaluation step
has been performed. It is accomplished either manually using the wiki system – by
modifying the textual task description, or by applying formalized knowledge with
respect to the applied data mining method. Then, parameters and/or the method
itself can be adapted. Refinement is performed according to the results of the
evaluation phase, so both steps are tightly coupled. Due to the application of the wiki,
different persons can collaborate in separate sessions, such that previous results can
be included in the refinement of other (related) sessions. Furthermore, previous
experiences can be documented using the wiki, for example, explanations/comments
by previous users. Furthermore, special refinement and/or evaluation knowledge
can be formalized for further improving the respective steps.
2.2
General Overview
As discussed in the last section semantic data mining is concerned with the utilization of
ontological knowledge and semantic annotations to be used throughout the data mining
and knowledge discovery process, similar to ontology-enhanced [
5
] data mining.
However, further semantic features are enabled by including a semantic core component,
e.g., a RDF-Store: Using that, results can be incrementally formalized and provided to
the store, while subsequent mining and semantic queries can make use of the collected
knowledge. The data mining query, results, and additional knowledge can then be
transparently integrated into a semantic analytic report: The idea of such reports is based on
analytical reports [
5
] that are simple text documents containing the mining results with
additional text (which is created by humans). In the semantic setting, we can
automatically transform the mining results into a format suitable for the report. Additionally,
the content can be enriched using semantic annotations and links between the reports
(and background information). The wiki also provides for flexible versioning which is
especially useful in a collaborative setting.
The sketched scenario is especially suitable for inexperienced users that are mainly
interested in reporting features of a data mining system. Such reports provide high-level
access to pre-specified queries that can be evaluated routinely. However, using the wiki
query mechanism, such queries can also be formalized in an ad-hoc fashion. Further
more detailed reports, analyses and mining sessions can then be implemented using
more advanced data mining tools, e.g., by applying the VIKAMINE [
3
] system.
On the application side, specialized sessions with domain experts, e.g., medical
doctors, and data mining engineers can be easily implemented using the collaborative
tool. In this context, the proposed approach provides, for example, flexible query
formalization, versioning, a history of queries and results, and the potential for knowledge
and experience management since the obtained semantic analytical reports can be
commented on, and can be linked to other (similar) documents. Further sessions can thus
easily build on results of previous sessions, with the same or new participants. For
experience management, the wiki can also be combined with a tagging system, e.g., [
9
].
2.3
Architectural Overview
Due to the limited space, we only provide a brief summary of the architecture of the
proposed approach. A more comprehensive discussion and overview is given in [
10
].
The architecture consists of two core components: The basic wiki system (provided by
JSPWiki (http:/www.jspwiki.org) is extended by the semantic wiki extension
KnowWE [
1
]. The wiki component provides basic features like editing, versioning,
user management, access management and attachment management. Additionally, it
directly supports the collaborative aspects of the sketched semantic mining approach.
KnowWE itself is designed as a highly extensible minimal core providing basic
semantic wiki features like formalization and reasoning. Therefore, for communication with
the mining component we designed the connector plugin KnowWE-RIP (REST [
11
]
Interface Plugin) that facilitates the connection to the mining web-service. The
semantic core component for storage and reasoning is given by a combination of the
Sesame (http://www.openrdf.org) framework and OWLIM. Sesame is a
javabased framework with support for storing and analyzing RDF data. OWLIM is a
semantic repository with reasoning capabilities that is packaged as a storage and inference
layer for sesame. As such, KnowWE integrates a semantic component and contains a
connector to the Sesame/OWLIM components for providing the semantic functionality.
We utilize the VIKAMINE [
3
] system (http://www.vikamine.org) for data
mining. VIKAMINE features a web-service that can be queried using XML based on
a specialized query language. The result (i.e., the answer) is also formulated as XML
and can thus transparently be integrated with the wiki.
The semantic mining process is initiated by the user, that is, by formulating a query
to the wiki system. Similar to other wiki-systems, the query is provided in the form
of an inline-query (e.g., [
2
]): The query is directly entered in textual form. Whenever
the wiki page is stored and/or reloaded with a new or modified query the result is
requested. In addition, we provide ’extended’ inline queries, such both the query and the
result (i.e., the ’answer’) can also be shown as required. Technically, the query is first
transformed to an XML-representation (VPDL, the VIKAMINE Pattern Description
Language, and then forwarded to the mining engine that produces an result in
XML/VPDL format. Finally, this result is re-transformed into human-readable textual form
to be displayed by the wiki. However, internally the ’raw’ result can be retained by the
versioning system of the wiki, such that always the latest result is available and can be
cached for efficiency. Therefore, changes, for example, due to an updated dataset, can
be easily extracted. The general architecture is shown in Figure 1. The seamless
integration of the result presentation enables (inexperienced) users to quickly evaluate the
obtained results by themselves and according to the formalized ontological knowledge.
2.4
Related Work and Discussion
Using ontologies for
enhancing data mining has
been discussed, e.g., by
Svatek et al. [
5
] and by
Antunes [
6
] in the context of
mining association rules.
Furthermore, Cespivova et
al. [
7
] and Kuo et al. [
8
]
describe applications in the
medical domain. While the
application of ontologies is
also a focus of the
presented approach, the
proposed method aims at a Fig. 1. Semantic Data Mining Architecture
more comprehensive
integration of semantic
information and knowledge. In contrast to the existing approaches, the proposed approach
considers a comprehensive two-way integration of semantic and data mining methods
for semantic data mining, with feedback in both directions. In this way, prior knowledge
can be transparently integrated. Using the wiki-support of the presented approach
collaborative sessions can be implemented. Furthermore, semantic annotations using the
wiki, linking unstructured, semi-structured and structured information is another novel
issue with respect to the presented approach. Semantic analytical reports can include
semantic annotations at the document level, global tagging, and associated query – data
mining results that are stored in the semantic store and thus provide powerful options
for knowledge-rich applications.