Open jmy48 opened 5 years ago
Great question! I actually do have concrete plans to integrate (spatio-)temporal graph operators and datasets. For example, we already provide the BitconOTC
dataset from the EvolveGCN paper. IMO, more temporal graph datasets can be easily created by interpreting each timestamp as a single data object and using dataloaders without shuffling to evolve through time.
Concerning operators, I'm currently working on a Know-Evolve implementation, but I am very interested in adding other operators too. Feel free to submit a PR if you find the time to implement any of them.
My use case is perfectly suited for know-evolve! Do you have an estimated completion time
It's done when it's done :P My guess is approximately 2 weeks. I am still learning about temporal point processes and their implementation and I have other stuff to do :(
For my case, the new dataset with an extra "time" dimension is preferable. Keras layer of 'Timedistributed' could be applied for easy extension. In other word, same edge information is shared with different node values.
I have integrated the RENet model. See here for documentation and an example.
ICEWS18-raw:
Epoch: 20, MRR: 0.2839, Hits@1: 0.1870, Hits@3: 0.3244, Hits@10: 0.4716
GDELT-raw:
Epoch: 20, MRR: 0.2082, Hits@1: 0.1321, Hits@3: 0.2248, Hits@10: 0.3555
I have integrated the RENet model. See here for documentation and an example.
The documentation moved?
❓ Questions & Help
Hi, I'm interested in learning over graphs that unroll through time via LSTM or other sequential learning mechanisms, like https://arxiv.org/pdf/1902.09130.pdf and https://arxiv.org/pdf/1803.07294.pdf. Is this possible/does this already exist?