Leveraging prior concept hierarchy

[usptact]

In my application, I have binary vectors to cluster. Along the data to cluster, I also have a fixed hierarchy concepts. Is it possible to leverage it for the task of clustering?

For example, I have many 100-dimension binary vectors that I would like to cluster. I also have a tree that relates the concepts hierarchically over the 100 terms/elements. The 100 terms/elements are the terminal nodes while the concepts group them.

I can ignore the hierarchy of concepts and proceed with clustering only. Making use of this tree seems to be useful for clustering.

[gregversteeg]

Hmm... this is kind of like a semi-supervised problem. corex_topic is the only corex version that has semi-supervision built-in. In corex_topic, you can specify .fit(anchors=[[1,7], [2]], anchor_strength=1) That would anchor latent factor 0 to columns 1 and 7 and latent factor 1 to column 2.