This is a simpler case than https://github.com/joepareti54/CustomerSegments and yet it uses a similar pipeline of ML tools to illustrate the concept.
The k-means model predicts which cluster a given customer belongs to. We have 6 customers, therefore the prediction is a [1,6] array. Next we calculate the proportion of customers that belong to a given cluster, by summing up the number of customers belonging to a given cluster, divided by the total number of customers (=len(purchases.index)). The k-mean model was obtained on the ORIGINAL data (from PCA with 3 principal components), and using 4 centers: when using the method model.cluster_centers, the output is a 4x3 array, where the rows are cluster centers, and the columns are principal components. To be abel to draw a comparison between original and new datasets, I need to transform the principal components back to the original features space (apples, oranges, ...)
Even though the k-means label is NOT invariant across runs, the transformed features should be consistent. Let's inspect 2 independent simulations: RUN 1 k-means predictions --- original data [3 1 0 2 1 0] k-means predictions --- new data [1 2 0 1 1 0] k-means centers coordinates in original features space apples oranges figs pears berries tomatoes onions 0 3.815409 11.140532 10.647411 13.750075 12.931495 1.829856 8.183790 1 6.856632 12.917453 21.462751 2.476402 4.464275 21.608906 3.263839 2 20.845435 1.405542 5.974547 4.056566 -0.426516 0.890788 2.932460 3 11.810483 -0.521511 -8.194870 9.490480 4.634975 5.231688 12.172282 RUN 2 k-means predictions --- original data [1 2 0 3 2 0] k-means predictions --- new data [2 3 0 2 2 0] k-means centers coordinates in original features space apples oranges figs pears berries tomatoes onions 0 3.815409 11.140532 10.647411 13.750075 12.931495 1.829856 8.183790 1 11.810483 -0.521511 -8.194870 9.490480 4.634975 5.231688 12.172282 2 6.856632 12.917453 21.462751 2.476402 4.464275 21.608906 3.263839 3 20.845435 1.405542 5.974547 4.056566 -0.426516 0.890788 2.932460
The explanation is: the clusters are the same, just called differently once you retrained the model ("0" remained "0", "1" became "2", "2" became "3", "3" became "1"). https://stackoverflow.com/questions/65749167/different-k-means-results-for-repeated-runs-of-this-program