I am in Athens [+1 to location-driven-research] to attend ECML PKDD 2011 and I am at the second MultiClust workshop today. This workshop is held in conjunction with ECML PKDD 2011.
The first talk of the morning was a invited talk by Michael Houle on "Combinatorial approaches to clustering". Michael Houle began the talk with how in high dimensions, all out intuitions about space go for a toss. One example that he gave illustrated how most of the points are concentrated along the boundaries in high dimensions. The talk revolved around how similarity measures between partitions cannot be fully trusted though there have been a range of combinatorial and spatial measures that have been introduced, since each measure suffers from some kind of drawback. He also talked about constructing secondary similarity measures [“Secondary similarity between two points v and w is defined in terms of data objects in the common intersection of neighborhoods based at v and w, where the neighborhoods themselves are determined according to a supplied primary similarity measure”] and how they can get around the curse of dimensionality.
One interesting question that was thrown out was "Can we do a fully automated clustering?", i.e., can we convert similarity functions into ranking on the neighbors, assume no knowledge of distribution of the data, and automatically pick 'k'. The talk moved on to approaches towards computing quality of a partition by computing a significance score for each point which depends on the correlation between a member of the set to the members in neighboring sets and other members of the same set, along with the size of cluster and the feature set. Once we know an object's significance, we can use this to change the shape of a cluster, by adding points and removing them.
Bart Goethals gave another invited talk titled “Cartification: from similarities to item set frequencies”. He talked about doing a k-nearest neighbors for all data points to find the right centers of clusters. The idea is that true 'centers' occurs in many knn baskets.
There were many interesting ideas out of the other talks at the workshop. I will highlight a few of them:
- One of the first talks about “Subjectively interesting alternate clusters”, uses an information theoretical framework to find interesting alternate clusters.
- Jilles Vreeken gave a very good talk about the current approaches in pattern mining and how we can use that knowledge in data mining. On this talk titled, “where pattern met subspace cluster”, he highlighted the similarities between subspace clustering and pattern mining where the goal is to find informative local structures.
- The talk on “Factorial clustering” was about using latent variables to generate multiple clusterings. One question that they wanted an answer for was a way to compute the right distance between partitions and I said, "Hey ! Use ours!”.
- Another talk on “Browsing clustering alternatives" outlined constructing a Merge-Split tree structure on the space of partitions and enable a user driven browsing of partitions.
The day ended with a rooftop reception with a magnificent view of the Acropolis and the temple of Zeus! I had very useful dinner conversations with Michael Houle and Thomas Seidl about general problems in the area of meta-clustering and on how to compute distances metrics on partitions. We chatted about combinatorial distances and spatial distances, and secondary level distances, where one constructs a ranking on top of a existing distance between many members.