Feature representation for feature set clustering

Question

I'm studying customer requirements clustering. Each customer's requirements are collected as a set of application features. I'd like to cluster those set of features, so that I can know what are the typical customer requirement clusters/patterns.

My first naive attempt is to represent each feature's presence by 1 when the feature presences in a feature set, or 0 otherwise. It results in high dimensional binary vectors for those feature sets. I then tried hierarchical cluster analysis. I'm afraid that the results are hard to interpret, as there is not much geometry intuition to explore.

My next attempt would be to categorize the features and represents only features belong in to the same category as a presence binary vector interpreting as coefficient to the polynomial with 2, and use the categories as the coordinates, this will resulting vectors of numbers. I hope that the resulted representation would have lower dimensions, and more geometry to explore, thus resulting more effective clustering.

Please give me pointers of existing good practice, or research on effective representation of features for feature set clustering. Thanks a lot!

Answer 1

I do not think clustering is what you are looking for. These methods assume everything is equally important and often would try to just partition the data one way or another. But do you really need partitions?

Instead, consider this market basket analysis. Every customer "bought" some requirements. Use frequent itemset mining to identify typical requirements and requirement combinations such as "customers that want A also usually want B".