Background: Traditional K-means++ algorithm has a pitfall: given a bad initialization, it may not converge and will generate a bad result.
Method: We use a special column-pivoted QR factorization.
Benefits: Our algo is more robust and efficient in computing.
Testing: We also tested the robustness of our CPQR algorithms and the graph below shows the performance of four diffferent clustering methods.
The following graph shows the accuracy comparison of our CPQR and the traditions k-means method. Check Source