My thesis topic is about the distributed knowledge management. A good data source of distributed knowledge is Wikipedia. That's why I am here and looking forward to some exciting discovery.
A more exact topic is Constructing a Knowledge Evolution Map System on Wikipedia. My thesis proposal was just on 1/14. An important reason to build such a system on Wikipedia is that the knowledge resource is rich and the quality of knowledge is good on Wiki.
I need a list of users to keep track the history which can be a good data source cause the knowledge evolution map system will take individuals as subjects. A map for a user.
The criterion to select a user to be a subject are as below:
The user must have edited on wiki for a certain period of time.
It has come to an idea: using the concepts of n-gram and hierarchical clustering (HAC). N-gram Clustering by date can find out the periods when the user has edited the similar pages, while hierarchical clustering can find out the similar periods which may not be in the sequent time.
The experiment results look not bad. It works to identify the different knowledge periods in time line.
Now we've collected the data mentioned above and clustered them by date. There comes some problems:
Every cluster hasn't been clustered by knowledge domain. This would cause the ambiguous knowledge structure in a cluster. The idea which taking the categories of Wikipedia seems not good because Wiki's categories are also defined by users, and intermingle with some categories which are not well-defined or not related to domain knowledge;
We use bottom-up hierarchical clustering to classify the data. The threshold of merging two clusters in every hierarchy is totally the same, i.e. 0.8. I wounder why this would work in hierarchical clustering. It should be less similar when the hierarchy is getting higher;
With TFxIDF value implemented, the computing time increases exponentially when the hierarchy is higher.
Now it has been the end of May. The progress of thesis is still going. We have conducted the mechanism to find out the knowledge evolution map. We argue that a user may have accessed the similar topics on Wiki, so we firstly implement n-gram algorithm to identify the periods with similar knowledge structure. After that, we use HAC to cluster these periods. In order to decide a good clustering result, we use Minmax to determine the final clustering result in HAC.
Now the problem has come to visualize the clustering result. My classmate had suggested me a java-based visualization tooltip: JFreeChart, and I have produced some charts. But it looks like not so user-friendly...
