دانلود کتاب بیش از 200 نکته پایتون و علم داده

The performance of KMeans is entirely dependent on the centroid initialization step. Thus, obtaining inaccurate clusters is highly likely. While KMeans++ offers smarter centroid initialization, it does not always guarantee accurate convergence (read how KMeans++ works in my previous post). This is especially true when the number of clusters is high. Here, repeating the algorithm may help. But it introduces an unnecessary overhead in run-time. Instead, Breathing KMeans is a better alternative here. Here’s how it works: • Step 1: Initialise k centroids and run KMeans without repeating. In other words, don’t re-run it with different initializations. Just run it once. • Step 2 — Breathe in step: Add m new centroids and run KMeans with (k+m) centroids without repeating. • Step 3 — Breathe out step: Remove m centroids from existing (k+m) centroids. Run KMeans with the remaining k centroids without repeating. • Step 4: Decrease m by 1. • Step 5: Repeat Steps 2 to 4 until m=0.