Web-based Learning Solutions for Communities of Practice

Artificial Intelligence. More precisely, numerical
techniques emphasize on the determination of
homogeneous clusters according to some similarity
measures, but provide low-level descriptions
of clusters (Anderberg, 1973). Recently, there
are works on clustering that focus on numerical
data whose inherent geometric properties can be
exploited to naturally define distance functions
between data points, such as DBSCAN (Ester,
Kriegel, Sander & Xu, 1996), BIRTH (Zhang,
Ramakrishnan & Livny, 1996), C2P (Nanopoulos,
Theodoridis & Manolopoulos, 2001), CURE
(Guha, Rastogi & Shim, 1998), CHAMELEON
(Karypis, Han & Kumar, 1999), WaveCluster
(Sheikholeslami, Chatterjee & Zhang, 1998).
However, data mining applications frequently
involve many datasets that also consist of categorical
attributes on which distance functions are not
naturally defined.
Clustering
In our case, we have numerical data that characterize
users. In order to cluster them, we have to
form a clustering algorithm in these input data. As
we have different users and different communities
of users, it is desirable to find clusters of users by
referring to a specific community each time, so
as to result to some valuable conclusions. Referring
to a specific community, we regard a cluster
as the collection of users that have something in
common by working on the same workspace and
thus. For example, let SP1, SP2,..., SPk be the k
spaces (from now on the term workspace and space
should be treated as equal) used by a community
A. We build an array X with size n x n (n is the
number of users), where the cell Xij denotes the
correlation between user i and user j.
First of all, we have to consider both cases for
symmetric undirected and directed arrays of data
depending on the analysis we want to make. In
addition, we can use the affinity metric introduced
in section 3.1, in order to build the array X.
So we have an array for each space in the
88
Mining Unnoticed Knowledge in Collaboration Support Systems
community A. After the construction of these arrays,
we build a unified array for the whole community
by using the arrays of each space. More
precisely, we may use the average of each cell,
which is the most common way. This final array
will have the relationships between users of the
same community. Having this array as an input
and applying an appropriate clustering algorithm
into them, we can find hidden relationships and
correlations between users of a community by
observing the resulted clusters of users.
An interesting approach could be the use of all
data arrays for all spaces and not the unified data
array of a community. According to this idea, we
can have more detailed views of users. This can
be derived by applying a clustering procedure for
each certain space. By clustering user data for a
specific space, we can provide micro-clusters of
users and give lower level clusters. After that,
we can perform a macro-clustering procedure.
This procedure can exploit user properties from
micro-clusters and find higher level clusters in a
history or time horizon. Based on this method, we
can have - at the same time - both detailed and
general views of users.
Depending on the kind of the analysis we want
to make about the users of the system, we have to
follow two different approaches about the nature of
the arrays that we use, one concerning symmetric
undirected arrays, and one concerning directed
arrays. Both of them are described below.
Symmetric Undirected Arrays
In this case, we can use any known hierarchical
algorithm for clustering where we use relevant
distances. In hierarchical clustering, there is a
partitioning procedure of objects into optimally
homogeneous groups. It is based on empirical
measures of similarity among the objects that
have received increasing attention in several
different fields (Johnson, 1967). There are two
different categories of hierarchical algorithms:
those that repeatedly merge two smaller clusters