Training set and accuracy calculation about gae

[tkipf]

Yes, this is indeed the case. The training setup is based on a real-world use case where one might want to predict missing links in a corrupted or incomplete graph. At training time, nothing is known about later test edges, i.e. we treat all missing edges at training time as negative examples (which will unavoidably contain a small number of false negatives, but as long as the false/positive ratio is small, this is not a big issue) -- this is related to the open world assumption in statistical relational learning.

…

On Wed, Mar 13, 2019 at 7:59 PM Hongwei Jin ***@***.***> wrote: Hi @tkipf <https://github.com/tkipf> When I review the code, the training set you generated is a subgraph of the original graph, while in model when calculating the loss, the function weighted_cross_entropy_with_logits compares the pred_scores and the subgraph adjacency matrix. The ones in the subgraph adjacency matrix represent the train_edge, while the zeros have both *val_edges_false* and *test_edge_false* as the trainings. Is that true? If yes, what I think the loss should be calculated by sampling the adj_matrix with train_edges and train_edges_false, which represent both 1s and 0s. Thus, what I think is to split the edges into E_G = E_train + E_val + E_test and split the non-existed edges into \bar{E}_G = \bar{E}_train + \bar{E}_val + \bar{E}_test Also, since the reconstructed matrix is not a matrix with values from [0, 1]. I think there is a problem when applying the sigmoid function to calculate the accuracy. Most of entries in the reconstruct matrix are positive, and it could leads to a bias calculation. Is that correct? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#26>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHAcYKOVtZ-7H0eF3VN9BZqxJp7Zx2Slks5vWUqsgaJpZM4bv6u-> .

[tkipf]

The sigmoid function is not an issue, but accuracy is not a good metric in this case as the classes are heavily unbalanced. It is better to report F1 score, average precision or area under the ROC curve.

…

On Fri, Mar 15, 2019 at 11:36 AM Thomas Kipf ***@***.***> wrote: Yes, this is indeed the case. The training setup is based on a real-world use case where one might want to predict missing links in a corrupted or incomplete graph. At training time, nothing is known about later test edges, i.e. we treat all missing edges at training time as negative examples (which will unavoidably contain a small number of false negatives, but as long as the false/positive ratio is small, this is not a big issue) -- this is related to the open world assumption in statistical relational learning. On Wed, Mar 13, 2019 at 7:59 PM Hongwei Jin ***@***.***> wrote: > Hi @tkipf <https://github.com/tkipf> > When I review the code, the training set you generated is a subgraph of > the original graph, while in model when calculating the loss, the function > weighted_cross_entropy_with_logits compares the pred_scores and the > subgraph adjacency matrix. > > The ones in the subgraph adjacency matrix represent the train_edge, while > the zeros have both *val_edges_false* and *test_edge_false* as the > trainings. Is that true? > > If yes, what I think the loss should be calculated by sampling the > adj_matrix with train_edges and train_edges_false, which represent both > 1s and 0s. > > Thus, what I think is to split the edges into > E_G = E_train + E_val + E_test > and split the non-existed edges into > \bar{E}_G = \bar{E}_train + \bar{E}_val + \bar{E}_test > > Also, since the reconstructed matrix is not a matrix with values from [0, > 1]. > I think there is a problem when applying the sigmoid function to > calculate the accuracy. Most of entries in the reconstruct matrix are > positive, and it could leads to a bias calculation. Is that correct? > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub > <#26>, or mute the thread > <https://github.com/notifications/unsubscribe-auth/AHAcYKOVtZ-7H0eF3VN9BZqxJp7Zx2Slks5vWUqsgaJpZM4bv6u-> > . >

[cshjin]

Will it be possible to predict the link based on the node embedding Z_i, Z_j using the metric learning?
Where in the case that we have learned the node embedding from the subgraph, and learn a metric to predict the existence of a link, say f(Z_i, Z_j) = Z_i' W Z_j

[tkipf]

Yes, this the main idea behind these two papers: https://arxiv.org/abs/1703.06103, https://arxiv.org/abs/1706.02263

…

On Fri, Mar 15, 2019 at 5:20 PM Hongwei Jin ***@***.***> wrote: Will it be possible to predict the link based on the node embedding Z_i, Z_j using the metric learning? Where in the case that we have learned the node embedding from the subgraph, and learn a metric to predict the existence of a link, say f(Z_i, Z_j) = Z_i' W Z_j — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#26 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHAcYEPewol6EWMgD-LYRdIabCSkR1JFks5vW8g9gaJpZM4bv6u-> .

[cshjin]

Got it, thanks for your information.

[jlevy44]

Why can't we just try to generate the graph from scratch via simulation.

I would like to use these methods to create an ensemble of different graphs that are close to the original graph rather than complete missing edges.

Mainly to compare the statistical properties and solve degeneracy problems with ERGMs.