mlpack nearest neighbor with cosine distance?

问题

I'd like to use the NeighborSearch class in mlpack to perform KNN classification on some vectors representing documents.

I'd like to use Cosine Distance, but I'm having trouble. I think the way to do this is to use the inner-product metric "IPMetric" and specify the CosineDistance kernel... This is what I have:

NeighborSearch<NearestNeighborSort, IPMetric<CosineDistance>> nn(X_train);

But I get the following compile errors:

/usr/include/mlpack/core/tree/hrectbound_impl.hpp:211:15: error: ‘Power’ is not a member of ‘mlpack::metric::IPMetric<mlpack::kernel::CosineDistance>’
 sum += pow((lower + fabs(lower)) + (higher + fabs(higher)),
           ^
/usr/include/mlpack/core/tree/hrectbound_impl.hpp:220:3: error: ‘TakeRoot’ is not a member of ‘mlpack::metric::IPMetric<mlpack::kernel::CosineDistance>’
if (MetricType::TakeRoot)
^

I suspect that the problem may be that the default tree type, KDTree, does not support this distance metric? If that's the issue, is there a tree type that does work for CosineDistance?

Finally, is it possible to use a brute-force search? I can't seem to find a way to use no tree at all...

Thanks!

回答1:

Unfortunately, like you suspected, arbitrary metric types don't work with the KDTree---this is because the kd-tree requires a distance that can be decomposed into different dimensions. But that is not possible with IPMetric. Instead, why not try using the cover tree? The build time of the tree may be somewhat longer but it should give comparable performance:

NeighborSearch<NearestNeighborSort, IPMetric<CosineDistance>, arma::mat,
    tree::StandardCoverTree> nn(X_train);

If you want to do brute-force search, specify the search mode in the constructor:

NeighborSearch<NearestNeighborSort, IPMetric<CosineDistance>, arma::mat,
    tree::StandardCoverTree> nn(X_train, NAIVE_MODE);

I hope this is helpful; let me know if I can clarify anything.

来源：https://stackoverflow.com/questions/42097957/mlpack-nearest-neighbor-with-cosine-distance

标签

c++

metrics

knn

mlpack