I spent some time today reflecting on how to cluster the messages of the STAMPS system. So I started back from the main goal, which is to support navigation and information retrieval. Therefore the goal is to find messages that have some similarities, that either describe the same resource in a different place or either that characterize the same place with content that is related.

In the system we can expect different degrees and kinds of similarities. We will have direct similarity when a search keyword will match directly to one of the keyword of the message. Else, a bit indirect will be to have a search keyword that matches one of the stemming alternative of the messages keywords.

Then a variety of degrees of relevance can be visualized on the messages that are connected by threads to the original messages.

A different case is a search string that contains multiple keywords, in which case we can imagine a plain match in a single message but also a match on different message. In this last case also, this search set the ground for a new connection between the keywords. The same thing can goes for messages that have different keywords but that are in physical proximity. In this case is the spatial proxemics that sets the relevance between the messages.

I still have in mind that an external referrer can be used to add semantics to the playground.

Another concept keep bouncing in my mind: is it the case to have a kind of permanence of this keywords concentration in a certain spot? Can we use this information to define landmarks that characterize a certain cluster? But then the point is that a cluster needs dimensions and I cannot figure out a proper way to find these parameters along the way.

Tags: spatial clustering