Nearest Neighbor Model #158
Conversation
akleeman
force-pushed
the
nearest_neighbor
branch
from
a5b8512 to
eb263c3
Compare
akleeman
force-pushed
the
nearest_neighbor
branch
from
eb263c3 to
772116d
Compare
| for (const auto &pair : indexer) { | ||
| assert(preds.at(pair.first).size() == pair.second.size()); | ||
| set_subset(preds.at(pair.first).mean, pair.second, &mean); | ||
| set_subset(preds.at(pair.first).covariance.diagonal(), pair.second, | ||
| &variance); | ||
| if (preds.at(pair.first).has_covariance()) { |
There was a problem hiding this comment.
Do we want to do this or just add 1e6 as variances where we currently have none?
There was a problem hiding this comment.
In this method we're taking a pair of MarginalDistribution and concatenating them, so if the both don't have a defined covariance then we want to preserve that in the concatenation. Somewhere on my list of want to dos is to remove the optional behavior for convariances in favor of a third distribution type, something like:
using MeanOnlyDistribution = Distribution<Empty>;
using MarginalDistribution = Distribution<DiagonalMatrixXd>;
using JointDistribution = Distribution<Eigen::MatrixXd>;
or something along those lines, but that's out of scope here.
| const JointDistribution &prediction) const { | ||
| const NearestNeighborModel<DistanceMetric> m(*this); | ||
| MarginalDistribution marginal_pred( | ||
| prediction.mean, prediction.covariance.diagonal().asDiagonal()); |
There was a problem hiding this comment.
Wouldn't prediction.covariance work here? Or are you looking to zero the non-diagonal elements?
There was a problem hiding this comment.
Yeah exactly, I need to zero the non-diagonal elements since the NearestNeighbor model can never actually predict off diagonals.
| std::size_t min_index = 0; | ||
| double min_distance = distance_metric(ref, features[0]); | ||
|
|
||
| for (std::size_t i = 1; i < features.size(); ++i) { |
There was a problem hiding this comment.
Could we turn min_distance into an optional so the loop can start at 0? The only difference in the loop would be !min_distance && going at the start of the if.
There was a problem hiding this comment.
I like that pattern better too ... but so far albatross doesn't used any optionals! So we'd have to add a third party lib for it which I've been avoiding (though perhaps the time has come).
There was a problem hiding this comment.
An alternative is to initialize the min_distance to DBL_MAX or some such, so it will always be replaced by the first distance.
Adds a model which (given a distance metric) will produce predictions for the nearest neighbor.
https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation