GraphSAGENodeGenerator in GraphSAGE

[kakaroa]

The parameter weighted in GraphSAGENodeGenerator, if weighted=True, is the sampling process sampling the first n nodes with the highest weight? Is it possible to interpret the weighted undirected graph as the original input data?

[cloudmadeofcandy]

After following a string of many different functions, particularly, this string:

GraphSageNodeGenerators -> _samplers = SeededPerWalk(SampledBreadthFirstWalk) -> _sample_neighbours_untyped() -> naive_weighted_choices()

I reached the definition of "naive_weighted_choices":

def naive_weighted_choices(rs, weights, size=None):
    """
    Select indices at random, weighted by the iterator `weights` of
    arbitrary (non-negative) floats. That is, `x` will be returned
    with probability `weights[x]/sum(weights)`.

    For doing a single sample with arbitrary weights, this is much (5x
    or more) faster than numpy.random.choice, because the latter
    requires a lot of preprocessing (normalized probabilties), and
    does a lot of conversions/checks/preprocessing internally.
    """
    probs = np.cumsum(weights)
    total = probs[-1]
    if total == 0:
        # all weights were zero (probably), so we shouldn't choose anything
        return None

    thresholds = rs.random() if size is None else rs.random(size)
    idx = np.searchsorted(probs, thresholds * total, side="left")

    return idx

So, the weights are normalized by the sum of weights, then are converted into probabilities.

Hope that this helps!!!

Quân Trương