what does the training parameter off APP

[ghost]

does it like the following:

../cli/app -train net.txt -save rep_dw.txt -undirected 1 -dimensions 64 -walk_times 1 -walk_steps 40 -window_size 5 -negative_samples 5 -alpha 0.025 -threads $Threads

because I can not run any good result under this method

[chihming]

Increasing the -walk_times (maybe > 10 or 20) and -sample_times usually helps.

BTW, if you test it on a public dataset, please tell me which dataset you used. I am willing and glad to have a check.

[ghost]

I just test on the amazon dataset refer to the original paper Scalable Graph Embedding for Asymmetric Proximity, but I can not get the result in the Node Recommendation experiment, and the result far less than the result in the paper

my running parameters is:

../cli/app -train net.txt -save rep_app.txt -undirected 1 -dimensions 128 -walk_times 20 -walk_steps 40 -window_size 5 -negative_samples 20 -alpha 0.025 -threads $Threads

Could you give me a comment if you have time to check and get any result? thank you very much

[khosla]

I believe that the implementation of APP is a bit incorrect. You should also save the context embeddings as a node has two representations in the source and target role.

[chihming]

Thanks for the feedbacks, I'll have a check.

[khosla]

in addition , do you know the best parameters for Cora dataset? I am unable to reproduce the paper's results for Cora even after using the context embeddings.

[chihming]

do you have the link of the dataset? I can test it later

[khosla]

here you go : http://konect.uni-koblenz.de/networks/subelj_cora