[Distributed] Partition MovieLens dataset
#8815
Conversation
kgajdamo
force-pushed
the
partition-movie-lens
branch
2 times, most recently
from
8652f1c to
b0f1d7e
Compare
Codecov ReportAttention:
Additional details and impacted files@@ Coverage Diff @@
## master #8815 +/- ##
==========================================
- Coverage 89.87% 89.39% -0.49%
==========================================
Files 479 479
Lines 31136 31145 +9
==========================================
- Hits 27984 27842 -142
- Misses 3152 3303 +151 ☔ View full report in Codecov by Sentry. |
kgajdamo
force-pushed
the
partition-movie-lens
branch
from
2cf2cd5 to
6a5cacc
Compare
rusty1s
changed the title
Distributed] Partition MovieLens dataset
| train_data, val_data, test_data = T.RandomLinkSplit( | ||
| num_val=0.1, | ||
| num_test=0.1, | ||
| neg_sampling_ratio=0.0, | ||
| edge_types=[edge_type], | ||
| rev_edge_types=[('movie', 'rev_rates', 'user')], | ||
| )(dataset[0]) |
There was a problem hiding this comment.
Looks like we drop the adjustment of message passing here computed in train_data, val_data and test_data. This is not necessarily a blocker (since I don't necessarily see a good way to fix this for now), but something you should be aware of. Currently, you would like information.
There was a problem hiding this comment.
Sorry, I didn't get that. What do you mean by "we drop the adjustment of message passing here computed in train_data, val_data and test_data"?
There was a problem hiding this comment.
The edge_index is different across different splits for link prediction tasks (in order to not leak information).
There was a problem hiding this comment.
I see. What about the solution used in the temporal_link_pred example?: temporal_link_pred.py #L27 Maybe we can use that one instead?
There was a problem hiding this comment.
Yeah, using temporal sampling is usually a good option to resolve this.
Changes made: