How does pinsage random walk work?

Questions
#1

Hi. I’m doing a project on the PinSage repo and I’m struggling to understand how the sample.random_walk works here:
def iter(self):
while True:

        heads = torch.randint(
            0, self.g.num_nodes(self.item_type), (self.batch_size,)
        )

        tails = dgl.sampling.random_walk(
            self.g,
            heads,
            metapath=[self.item_to_user_etype, self.user_to_item_etype],
        )[0][:, 2]

        neg_tails = torch.randint(
            0, self.g.num_nodes(self.item_type), (self.batch_size,)
        )
        mask = (tails != -1)
        yield heads[mask], tails[mask], neg_tails[mask]

If the graph is a bipartite graph, then how would the sampling work? The edges would be from the item node to user node so wouldn’t the tails list contain nodes of both types? But I assume it is somehow only the item nodes? How?

#2

It is a two-hop random walk, with an item-to-user edge and then a user-to-item edge, as specified in the metapath argument:

#3

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.