20:30 Not sure if the author (or anyone who knows the answer) will ever see this message but this part is a bit vague. Since the embedding module uses a GNN encoder and thus samples the neighborhood (they sample the most recent edges i.e. nodes connected via those edges) wouldn't it make sense that the "Updated memory" part also contains all of those neighbors included as well? So not only s1, s2, and s3 but also their temporal neighbors which we'll use to compute the final embeddings z1, z2, z3. Also, the most interesting part of the paper aside from memory was not explained - time embeddings and how they integrate that information to get the final embeddings (like z1, z2, z3 in the example given). Finally related to the memory staleness problem. Say I'm not on Twitter for a month. But if somebody retweeted my tweet or liked my tweet (or more generally interacted with me in any way) that means we had an edge interaction event and I (my node) was involved and so "my" memory should get updated, right? Both mi (source node message) and mj (target node message) get computed once we have eij according to the paper.
20:30 Not sure if the author (or anyone who knows the answer) will ever see this message but this part is a bit vague. Since the embedding module uses a GNN encoder and thus samples the neighborhood (they sample the most recent edges i.e. nodes connected via those edges) wouldn't it make sense that the "Updated memory" part also contains all of those neighbors included as well? So not only s1, s2, and s3 but also their temporal neighbors which we'll use to compute the final embeddings z1, z2, z3.
Also, the most interesting part of the paper aside from memory was not explained - time embeddings and how they integrate that information to get the final embeddings (like z1, z2, z3 in the example given).
Finally related to the memory staleness problem. Say I'm not on Twitter for a month. But if somebody retweeted my tweet or liked my tweet (or more generally interacted with me in any way) that means we had an edge interaction event and I (my node) was involved and so "my" memory should get updated, right? Both mi (source node message) and mj (target node message) get computed once we have eij according to the paper.
Very interesting presentation, but I get advertisements every 5 minutes. That sucks.
Great talk, and an insightful discussion!
Starts at 00:59
Incredibly annoying to have advertisement every 2.5 minute. Besides that, very nice!
Too many ads