search

uhh-lt / path2vec

Learning to represent shortest paths and other graph-based measures of node similarities with graph embeddings
Apache License 2.0
33 stars 14 forks source link

Add results of node2vec 600 to the paper to figure Figure 3 #15

Closed alexanderpanchenko closed 5 years ago

snkohail commented 6 years ago

Where to ?

akutuzov commented 6 years ago

I'll do this.

snkohail commented 6 years ago

Thanks @akutuzov

akutuzov commented 6 years ago

@snkohail is this the node2vec 600 model you trained? http://ltdata1.informatik.uni-hamburg.de/shortest_path/models/converted_wordnet.600.10.70.5.F.emb.gz

akutuzov commented 6 years ago

@snkohail have you trained node2vec embeddings with the dimensionality 600 and 25 random walks? I see only 5-walk models results in the googledoc.

akutuzov commented 6 years ago

I updated the tables with the nodevec600 results and created new plots. They are already in the Overleaf paper. Node2vec seems to radically improve when moving from 300 to 600 dimensions. However, it can be explained by the fact that the hyperparameters of 300D and 600D models are different: 300D were trained with 25 random walks for 10 epochs and 600d were trained with 5 random walks for 5 epochs. Anyway, these results do not significantly change the overall picture.

alexanderpanchenko commented 6 years ago

Thanks!

@snkohail we are preparing a new submission for NAACL. Could you compute node2vec with the same hyperparameters as for 600 dimensions for the dimensions 300, 200, 100, and 50?

snkohail commented 6 years ago

@alexanderpanchenko Yes, on it.

alexanderpanchenko commented 6 years ago

Great, thanks!

On Fri, Sep 14, 2018, 6:31 PM Sarah notifications@github.com wrote:

@alexanderpanchenko https://github.com/alexanderpanchenko Yes, on it.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/uhh-lt/path2vec/issues/15#issuecomment-421413595, or mute the thread https://github.com/notifications/unsubscribe-auth/ABY6vsA7HNITkyzy1T1eFhy-30GGJ8GPks5ua9nmgaJpZM4Wlgd_ .

snkohail commented 6 years ago

I am getting this error when I run node2vec on the server (gpu or cpu1). cannot execute binary file: Exec format error I had no problem running the script b4 ?! Any help?

alexanderpanchenko commented 6 years ago

Can you provide the command you launch and the full error msg?

Did you try to recompile?

On Sat, Sep 15, 2018, 3:46 AM Sarah notifications@github.com wrote:

I am getting this error when I run node2vec on the server (gpu or cpu1). cannot execute binary file: Exec format error I had no problem running the script b4 ?! Any help?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/uhh-lt/path2vec/issues/15#issuecomment-421521978, or mute the thread https://github.com/notifications/unsubscribe-auth/ABY6vvvrtSErrft589jGZfbbW4JUZCigks5ubFwIgaJpZM4Wlgd_ .

snkohail commented 6 years ago

it is the node2vec script: ./node2vec -i:graph/karate.edgelist -o:emb/karate.emb -l:3 -d:24 -p:0.3 -dr -v

the same script runs normally in my computer, I am running it now.

I tried both node2vec compiled on my computer and moved to the server, and another one compiled on the server. both gives the same error. Even the one which I used before in my experiments didn't work. I googled the error. seems something with the bash cofig or directory. But I don't have sudo permission to do the solutions.

alexanderpanchenko commented 6 years ago

Ok. I do not know. We can look together in the office.

On Sat, Sep 15, 2018, 2:18 PM Sarah notifications@github.com wrote:

it is the node2vec script: ./node2vec -i:graph/karate.edgelist -o:emb/karate.emb -l:3 -d:24 -p:0.3 -dr -v

the same script runs normally in my computer, I am running it now.

I tried both node2vec compiled on my computer and moved to the server, and another one compiled on the server. both gives the same error. Even the one which I used before in my experiments didn't work. I googled the error. seems something with the bash cofig or directory. But I don't have sudo permission to do the solutions.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/uhh-lt/path2vec/issues/15#issuecomment-421561115, or mute the thread https://github.com/notifications/unsubscribe-auth/ABY6vhDPBpV-RECPOZs-MPMbnuE2urMIks5ubO_9gaJpZM4Wlgd_ .

snkohail commented 6 years ago

Sorry, but I had to run the experiments on my computer coz I don't have permission to run shell scripts. I also couldn't upload the results to the server. Please find all the embeddings of size 600 here:

https://drive.google.com/open?id=1SbegQaXxH_t3BqKeGsdkyTC5UUYcprvj

alexanderpanchenko commented 6 years ago

thanks! sounds stange you do not have the rights - ask Steffen to fix that on ltcpu1

On 25 Sep 2018, at 17:43, Sarah notifications@github.com wrote:

s on my computer coz I don't have permission to run shell scripts. I also couldn't upload the resu

alexanderpanchenko commented 6 years ago

@Andrei /@Sarah could you compute the performance on this new model please?

On 25 Sep 2018, at 17:43, Sarah notifications@github.com wrote:

Sorry, but I had to run the experiments on my computer coz I don't have permission to run shell scripts. I also couldn't upload the results to the server. Please find all the embeddings of size 600 here:

https://drive.google.com/open?id=1SbegQaXxH_t3BqKeGsdkyTC5UUYcprvj https://drive.google.com/open?id=1SbegQaXxH_t3BqKeGsdkyTC5UUYcprvj — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/uhh-lt/path2vec/issues/15#issuecomment-424392834, or mute the thread https://github.com/notifications/unsubscribe-auth/ABY6vgp7dz9U0BExE3mbGkAANzy3Rn4Zks5uek89gaJpZM4Wlgd_.

akutuzov commented 6 years ago

@snkohail thanks! @alexanderpanchenko but should we use the 600D models with the hyperparameters used for models of lower dimensionalities, or should we instead use the scores of those lower-dimensional models with 5 random walks and 5 epochs, as in the original 600-D model by @snkohail ? What do you think?

akutuzov commented 6 years ago

OK, I think will be less time-consuming if we simply use the hyperparameters which we already used before (25 random walks, context size 10, 10 epochs). I updated the node2vec results in our tables with the scores for node2vec.600.25.10.10 model and recreated the plots in the paper using these scores. As I've said, it does not change the overall picture anyway.

alexanderpanchenko commented 6 years ago

Thanks! Sounds good to me: I have no special opinion about it.

On Sep 26, 2018, at 10:42 PM, Andrey Kutuzov notifications@github.com wrote:

OK, I think will be less time-consuming if we simply use the hyperparameters which we already used before (25 random walks, context size 10, 10 epochs). I updated the node2vec results in our tables with the scores for node2vec.600.25.10.10 model and recreated the plots in the paper using these scores. As I've said, it does not change the overall picture anyway.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/uhh-lt/path2vec/issues/15#issuecomment-424862950, or mute the thread https://github.com/notifications/unsubscribe-auth/ABY6vmmhmQOag6-Lrub8DHWSZyz72dVbks5ue-bOgaJpZM4Wlgd_.

akutuzov commented 5 years ago

@snkohail Where can I find the node2vec models you trained before? http://ltdata1.informatik.uni-hamburg.de/shortest_path/models/ certainly does not have all of them. For example, what about the 300_25_10_10 model?