| SparseMap |
- LLR embedding is still impractical
|
- For large databases,
 ,  are very large
|
- Dimensionality of embedded space is large
|
- For large sized
 , time to determine the corresponding dimension by computing  is large
|
- SparseMap proposes two practical heuristics
|
- Compute
 by considering some (and not all) 
|
| |
- This is a way to avoid computing all
 distances
|
| |
- is an upper bound of
|
- Reduce dimensionality by retaining some (and not all)
|
| |
- Greedy resampling: Delete 's that contributes largest to stress
|
| |
- Sample only some object pairs when computing stress
|
|
