System information

• TensorFlow version (you are using): 2.1
• Are you willing to contribute it (Yes/No): Yes (partially, what can be done on the python side, and maybe CPU kernels)

Will this change the current api? How?
Add an operation to tf.sparse.broadcast_indices (feel free to suggest a beter name), doing the following:

tf.sparse.broadcast_indices(a, b, default_value): Given two sparse tensors of same dense shape, inserts `default_value` at positions such that a.indices == b.indices.

This would provide a basic building block for arbitrary pointwise-binary operations on SparseTensors. Therefore I would also suggest adding a function like

tf.sparse.map_binary_op(a, b, op, def_value=0):
   a, b = broadcast_indices(a, b, def_value)
   return SparseTensor(a.indices, op(a.values, b.values), a.shape) 

This would then allow for more efficient implementations e.g. of binary metrics when both labels and predictions are sparse, e.g. when calculating P@k statistics.

I think I can provide an initial, sub-optimal implementation that builds this on top of tf.sparse.add for default values of zero. Would you be interested in a PR like that (maybe putting it into experimental namespace at first)?