Summary: - Theoretically, usually running time is between O(n~2) and O(n~3). - Empirically, it often behaves like O(n~2). - Simple and as such parallelizes well even on cloud platforms. Highly effective on datasets with as many as one million training examples. Limitations: - O(n~2) memory usage - blows up memory usage quickly. - Many parameters, large prediction time.
展开▼