what does 'lookback' means?
Closed this issue · 5 comments
Hello,
I notice that 'lookback' parameter existed in dynAE,dynRNN,dynAERNN , but I don't konw what it means.
I set it to 2 , input 10 graphs by timestamp and output 8 embeds as a result.
Do you use two graphs to get an embed?
The above three patterns generate an embedding predictive of the graph at next time step. For this they take the previous "look back" number of graphs and learn the embedding.
@palash1992 When i input a dynamic graph with 10 timestamps and the lookback is 2, the number of the results is 8. Why not the result is 9 (predict 3-11)
@palash1992 Hi sir, I'm just working with your algorithm. I have the problem too. When I set the 'loopback' parameter at 2, will the algorithm work like that (4 time steps input and 3 output predictions)?
Hi, if the lookback is k, the model uses previous k graphs as input and the current graph as output to understand the dynamic patterns. Thus, if you have 10 graphs and look back 2, it will use graphs 1 and 2 to predict links in graph 3 and graphs 8 and 9 to predict graph 10. The learned parameters can then be used to predict links at time 11 by inputting graphs 9 and 10 but by default, this is not done.
Hi,I have some questions. What does it mean if I set 'lookback' to 1? In your paper ,you says lookback can be set to 1 ,but it seems not to make sense. How could we get dynamic information just through one graph?