0

We have multiple Gaussian variables, which could be the locations of 2-d points. Suppose the 2-d points are measured independently.

If we connect the adjacent points, then we will get a structure (graph). Suppose we have a model to compute whether the structure is stable or not.

How can we use the information about stable structure to correct the given measurements? We can anticipate that some positions of the 2-d points may cause an unstable structure. Hence, we can prune these positions and get a better estimate.

merv
  • 67,214
  • 13
  • 180
  • 245
firefighter
  • 25
  • 3
  • This isn't really a programming question - more like statistical learning methods. I'm voting to migrate the question to [CrossValidated](https://stats.stackexchange.com), where the question would be more appropriate. However, this process often takes a few days, so you may be better off just reposting there manual. If you do that please make sure to delete your question here. Cross-posting is considered bad form. – merv Feb 05 '19 at 15:37
  • Nevertheless, I will say this reminds me of a (non-Bayesian) method used in [the single-cell RNA-seq imputation method called MAGIC](https://www.biorxiv.org/content/10.1101/111591v1). Basically, it computes kNNs, then runs a Gaussian filter to compute affinity between neighbors, which reduces noise in the neighborhood graph. Then diffusion along the graph is used to impute values. I.e., if you prune, then diffuse, the final values will reflect the unpruned neighbors more closely. And if I remember correctly, they assume the transformed variables (log(counts)) are Gaussian. – merv Feb 05 '19 at 15:43

0 Answers0