4 ms·

Thanks for the feedback/ideas :) You're right, it won't scale as well as could be for large datasets -- or work on large dimensionalities. It's meant to be a l

by mck- 13y ago

Thanks for the feedback/ideas :) You're right, it won't scale as well as could be for large datasets -- or work on large dimensionalities. It's meant to be a light-weight solution for the more common use-cases.. For larger cases, it would be good indeed to resort to better methods.. I wanted it to be stateless and functional, whereas with space-partitioning trees, don't you need to maintain the tree (and not generate it on the fly for it to be scalable)? As for your second point, could you elaborate on why the movie example is non-metric?

yid 13y ago
> I wanted it to be stateless and functional, whereas with space-partitioning trees, don't you need to maintain the tree (and not generate it on the fly for it to be scalable)? Yes, but it's essential to maintain some sort of summary or index data structure to make the method scalable. A linear scan may be stateless, but that's hardly an issue when the method won't scale beyond a few hundred examples. > As for your second point, could you elaborate on why the movie example is non-metric? No triangle inequality.
- mck- 13y ago
  The simple algorithm as is uses sorting, hence nlogn -- scales reasonably well into thousands of examples.. a tree would be klogn -- minor improvement unless n >> k?