resolving specific clusters from data (spectral clustering)

Question

[benformatics]

I'm trying to cluster this genetic data but even with multiple different methods k-means/medoids spectral, etc... I can't seem to resolve the two big clusters. Any suggestions? I would really like to be able to identify that central cluster around y=1. The goal is to run this on multiple datasets.

[ATpoint]

What is x and y? What kind of data is that? Did you try a graph-based clustering based on these two dimensions? So building a KNN/SNN graph first and then cluster that with igraph (e.g. louvain)?

[benformatics]

I like your suggestions, I will try that. It's gene coverage at specific coordinates.

Answer 1

[Mensur Dlakic]

I think Gaussian mixture models will work well on this type of scatter, though you will likely end up with more than two clusters. If you provide [X, Y] coordinates for data points, I could tell you for sure.

• https://scikit-learn.org/stable/modules/generated/sklearn.mixture.GaussianMixture.html
• https://scikit-learn.org/stable/modules/generated/sklearn.mixture.BayesianGaussianMixture.html

You could literally plug in your data into a script below instead of random points they generate:

https://scikit-learn.org/stable/auto_examples/mixture/plot_gmm.html