scikit learn - Distance matrix for AgglomerativeClustering with Ward's linkage

Question

I have a distance matrix from a RandomForest and I want to use Hierarchical clustering with Ward's linkage to look for clusters via Sci-Kit learn AgglomerativeClustering . I know that Ward linkage only works with Euclidean distances and the RandomForest distance matrix consist of (squared) Euclidean distances. I also know that I can use affinity = 'precomputed' and input my distance matrix, but then I cannot use Ward's linkage (according to Sci-Kit learn's documentation). Should I then just input my square distance matrix (n x n) and use affinity = 'euclidean' and linkage = 'ward' instead, since I am not violating any mathematical assumptions in my humble opinion?

I also read that I could perform a PCA on the distance matrix (after double-centering) and then use this in Kmeans, since Kmeans doesn't handle (implicilty) distance matrices. Normally, it takes sample x feature matrix (Data matrix) as input. Would this post-PCA matrix be a better input for AgglomerativeClustering with Ward than the distance matrix?

Answer 1

等待大神答复