I’ve got several thousand observations in 350-dimensional space, in a relatively sparse matrix (median observation has 11 non-zero dimensions). I'm using a density-based clustering algorithm, DBSCAN, to identify clusters and noise points (points that do not fall into clusters).

DBSCAN requires setting two parameters. The first is the minimum threshold number of points that defines a cluster. The guidance in the original paper (Ester et al. 1996) is to use the number of dimensions in the space. In a prior post (where I asked way too many questions) a helpful forum user pointed out that because the value of most dimensions is zero for most observations, the intrinsic dimensionality of my dataset is much lower than 350, and is more like the median or mean number of nonzero dimensions.

I've been reading about intrinsic dimensionality, but it is well outside my expertise. I am looking for a paper or two that I can cite (and learn more from) that documents why the intrinsic dimensionality is closer to the median number of nonzero dimensions. Any help is much appreciated!


First of all, you can compute the actual intrinsic dimensionality for some points, and just give the data.

Secondly, it should be obvious that adding a constant dimension does not affect the intrinsic dimensionality. Thus, if for most points most dimensions are 0, the intrinsic dimensionality must be lower.

You don't need a citation for that.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.