Yes. This problem is of particular importance in drug response prediction, where small molecules are represented as graphs where the atoms are nodes and the bonds are edges. These are difficult to work with because of the variability in the number of nodes.
> Also I'd like to thank Ferenc Huszar for highlighting some drawbacks of these kinds of models.
As someone who understands ML concepts but hasn't used ML in practice, can someone elaborate on these drawbacks compared to using a traditional neural network?
In a traditional CNN you use the same convolution kernel across an entire image. Here you do the same but you convolve the features of a person (pixel) with the feature of its neighbors, but those change depending on the network. Drawbacks are: I) you add one new person to the network, the whole thing has to be retrained; ii) convolutions are more complex and expensive
I didn't get a chance to look at the paper, but I wanted to share my experience working at two social network companies and publishing an academic paper on the topic of machine learning on a social network graph. Many people think it's necessary to use graph based machine learning (or graph databases).
In practice I found it almost always better to just use a "flat" representation and do regular machine learning (or database schemas). I would love to hear if others disagree.
I think it's highly task-specific. In network biology for example, taking advantage of the network structure, either through node embeddings or directly, is almost always helpful (even necessary). I'm not as familiar with applications to social networks, or graph DBs.
16 comments
[ 2.8 ms ] story [ 47.5 ms ] thread[1] https://arxiv.org/abs/1611.07308
It was written by the authors of GraphSAGE, who also implemented it at Pinterest: https://arxiv.org/abs/1706.02216 and https://arxiv.org/abs/1806.01973
https://www.manning.com/books/graph-powered-machine-learning
Any recommendations on an up to date book on Social Network Analysis and Graph Clustering/Community Detection?
[1] https://github.com/FluxML/Flux.jl/issues/625
https://danielegrattarola.github.io/spektral/
[1] https://arxiv.org/abs/1509.09292
[2] https://arxiv.org/abs/1603.00856
[3] http://proceedings.mlr.press/v48/niepert16.html
[4] https://persagen.com/files/misc/xu2017seq2seq.pdf (PDF)
As someone who understands ML concepts but hasn't used ML in practice, can someone elaborate on these drawbacks compared to using a traditional neural network?
https://medium.com/octavian-ai/how-to-get-started-with-machi...
In practice I found it almost always better to just use a "flat" representation and do regular machine learning (or database schemas). I would love to hear if others disagree.