/link-prediction-examples

Primary LanguagePython

Link Prediction Examples

Trying out different network architectures on Cora Dataset using DGL & Pytorch.

Here are the models that are included in this repo:

Model Name Description
GraphSAGE Learns node representations by aggregating node neighborhood features
StochasticGraphSAGE Stochastic version of GraphSAGE. Uses DGL's DataLoader and NeighborSampler APIs. batch-size, n-neighbors and negative-samples parameters are available for this model.
VGAE Variational Graph Auto-Encoder. Only difference from standard VAE is that it uses GraphConv layer instead of a Linear layer.
ResidualMLPVGAEReduceSum Jointly learns both graph-wise (with VGAE) and attribute-wise features and aggregates them with summation
ResidualMLPVGAEReduceFF Jointly learns both graph-wise (with VGAE) and attribute-wise features and aggregates them with a feed-forward layer
ResidualMLPGraphSAGEReduceSum Jointly learns both graph-wise (with GraphSAGE) and attribute-wise features and aggregates them with summation
ResidualMLPGraphSAGEReduceFF Jointly learns both graph-wise (with GraphSAGE) and attribute-wise features and aggregates them with a feed-forward layer

Code

Most of the preprocessing stuff is taken from DGL documentation.

You can run it with the following example command:

    python train.py \
        --model ResidualMLPVGAEReduceFF \
        --epochs 80 \ 
        --graph-h-feats 16 16 \ 
        --mlp-h-feats 16 16 \
        --reduce-ff-size 16 \
        --lr 0.01
    /