Train BlogCatalog with no feature reconstruction (and normal input features)

ixxi-dante / an2vec

Bringing node2vec and word2vec together for cool stuff

GNU General Public License v3.0

22 stars 6 forks source link

Train BlogCatalog with no feature reconstruction (and normal input features) #43

Closed wehlutyk closed 6 years ago

wehlutyk commented 6 years ago

Simpler version of #40.

wehlutyk commented 6 years ago

[x] With flat adjacency loss (i.e. after #38) (running this diff on top of 868d5e49b327ba7075175eda500c6588bd0d509e)
[x] With weighed adjacency loss (i.e. before #38) (running this diff on top of 868d5e49b327ba7075175eda500c6588bd0d509e)

wehlutyk commented 6 years ago

The first run (with flat adjacency loss) is done, results in 0ae284e8d35339b1550596e461929ae4d06461ff, see the projects/scale/blogcatalog-dim_ξ=10-no_adj_cross_entropy_weighing-no_feature_reconstruction-results.ipynb notebook. Nothing seems to have changed from #38.

wehlutyk commented 6 years ago

The second run (with weighted adjacency loss) is done, results in e9ef69181b870f81db9db7f4260af77808f0bd75, see the projects/scale/blogcatalog-dim_ξ=10-no_feature_reconstruction-results.ipynb notebook. This looks basically like the results of #31, i.e. with the same bad training result on adjacency loss (fixed by #38).

The conclusion here is that feature reconstruction is not a problem: when disabled, the adjacency reconstruction performance is unchanged (and still fixed the same by the changes of #38).