Interpreting the result of link_prediction explainer?

[kal-purush]

How can I interpret the result of link_prediction explainer?
I run the example code from the git repo and this the output

tensor([0.1473, 0.1484, 0.1455,  ..., 0.1462, 0.1458, 0.1503])
tensor([[0.1732, 0.1679, 0.1755,  ..., 0.1613, 0.1687, 0.1801],
        [0.1777, 0.1700, 0.1726,  ..., 0.1658, 0.1579, 0.1625],
        [0.1720, 0.1789, 0.1607,  ..., 0.1541, 0.1602, 0.1825],
        ...,
        [0.1782, 0.1908, 0.1695,  ..., 0.1556, 0.1664, 0.1493],
        [0.1819, 0.1635, 0.1719,  ..., 0.1755, 0.1692, 0.1742],
        [0.1662, 0.1824, 0.1678,  ..., 0.1532, 0.1635, 0.1740]])

But I am not sure how to interpret this? I want to understand how a certain edge was predicted i.e., which features or edges were used to predict a particular edge.

[wsad1]

Which example from pyG did you run, was it this.
If so the first tensor is the edge_mask indicating which edges are important, and the second is the feature importance mask for each node.

[rusty1s]

Discussed on Slack as well. I think there are still some issues with GNNExplainer on link-level tasks since we never clear feature importance/edge importance for edges/nodes never visited. I imagine a simple way to fix this would be to compare node_mask before and after GNNExplainer runs, and simply zero out any score that has not changed. WDYT?

[rusty1s]

I fixed this in #6787.