|
I have data which has several properties (metadata, as key value pairs, where the keyspace is shared over the whole dataset) per object. I took a sample of objects and divided them in n groups according to an unknown algorithm. What statistical methods or algorithm exists to find the relevant properties and their weight for the division to receive a similar grouping of the data than the "unknown algorithm" did? |
||||
|
|
Since your metadata is probably discrete, I suggest using classification trees. Note that from your example it is highly likely that your unknown algorithm is random, i.e. there is no algorithm. |
|||
|
|
|
This isn't a clustering, it's (supervised) classification. There are a lot of methods to do classification, such as naïve bayes for binary features and linear discriminants for continuous, or neural networks for some opaque combination of both. For the example you mention, the feasibility of what you suggest depends a lot on the metadata - for example, if it's been tagged (=classified) properly, you could expect to be able to find some tags they like/dislike. However, the problem you stated is more general. Discovering the exact algorithm is unlikely, and in the example one doesn't even exist. |
|||
|
|
|
You can use SOM which is a kind of supervised clustering. Or as sesqu said their are tons of other algorithms such as support vector machines, regression or logistic regression. Neural Networks, etc .. |
|||
|
|
