Upload
walter-obrien
View
217
Download
0
Tags:
Embed Size (px)
Citation preview
What is it
Clustering is the classification of objects into different groups, or more precisely, the partitioning of a data set into subsets (clusters), so that the data in each subset (ideally) share some common trait - often proximity according to some defined distance measure.
We have being doing it
We have been grouping people, cars, etc. We are just not very good when we have too many
items to keep track Experts can track five to six dimensions, we may
have data set with many times of that We can only see the obvious groups, most likely It is difficult for us to see the hidden ones, or the
combined ones
An Example
You can group your customers (for a bike store) into several groups based on • Gender • Income• Age• Etc
There may be other things, such as do they play game?
Principles of Clustering
Guessing and lying (MS)• Setting clusters
Training with data Calibrating your clusters Training again Repeating until converged or going nowhere
The clustering mythology is very sensitive to the starting points and can converge at local solutions that many not be optimal global solution
Scalable clustering
Ideally, the data point that will not change its cluster do not need to be considered
In MS’ implementation, it will read the first 50,000. If that don’t converge, we process the next 50K, rather than read in and process all 100K.
Few interesting parameters Clustering_Method
• What method to use 1~4 Clustering_Count
• The number of clusters to find• 0 makes the algorithms to guess a good number
Minimum_Support• What case count can be considered as empty
Stopping_tolerance• The number of cases switch clusters
Sample_size• For scalable clustering
Cluster_Seed• Where to put the clusters
Maximum_Input_attributes• A number before attributed considered before automatic feature selection kicks in. Automatic feature selection,
selects the most popular attributes Maximum_states
• Possible values
Understanding The Results
Comprehending the results can be difficult because you have to look for many directions• High-level overview• Look into a cluster• Determine how a cluster is different from a near
by one
High-level overview Cluster Profiles view -- too much info
• Getting some sense regarding who/what are in each cluster
Look into a cluster
The Cluster characteristic view• See the attributes that are going together • Note that an attribute ranks high may be
because it is ranked high on all the cluster. In that case, it is not that interesting.