compute pairwise_distance with custom metric function for non-numeric data

This is a feature request: I’m working with strings and have a custom metric function to compute similarities between the strings. I would love to use the pairwise_distance function together with this custom metric to compute the whole distance/similarity matrix for my data. However, I’m getting a ValueError: could not convert string to float: 'some string' when the X and Y arrays are checked. It would be great if these checks could be made optional for custom metric functions.

[cod3licious]

I wanted to use the function to compute a distance matrix that I was using elsewhere, i.e., outside of sklearn; I figured the function had some nice parallelization built in or was in some other way more efficient than a naive implementation. So yes, it’s probably of limited value in conjunction with sklearn models, but even if there the better solution would be to pass a precomputed distance matrix, this matrix needs to be computed somehow as well. And considering that it’s probably just a matter of adding one parameter check_input=True and then one if statement before the arrays are checked, I think it’s worth it, even if the benefit doesn’t extend to other sklearn models.

[vnmabus]

I have to add something to this topic. I am the main maintainer of scikit-fda, a project that implements functional data methods compatible with scikit-learn. In our case we do not even have arrays, as our data represent functions, so we have our own objects analog to a 1d array of functions (but sharing common things between them). Also we have developed functional metrics, such as the Lp metrics (which use integrals instead of sums). We even know how to compute the pairwise distance for some of these metrics in a more efficient way than the naive implementation (for example multiplying the weights of the quadrature and using einsum). Moreover, we want to apply some distance-based methods almost verbatim to our objects, such as knn and agglomerative clustering, an objective that we currently achieve wrapping the estimators and using the “precomputed” distance.

In summary, it would be nice if we had support for the following things:

• Computing the pairwise distances with our types and metrics, relying in the optimized implementation if available.
• Using the distance based classifiers directly with our types and metrics, again relying in the optimized implementation of the pairwise distances if available.