sanidhyaanand / graphretrievaltoolkit

Graph Matching

Models Used

1. SimGNN:

   • Encoder:
     • Inputs: Initial one-hot encoded node embedding matrix $U \in R^{NXD}$
     • Outputs: Aggregated node Embedding Matrix $U \in R^{NXD}$
     • Uses: Neighbour Aggregation with Conv Nets (SAGE, GCN, GAT)
   • Attention Mechanism:
     • Inputs: Node Embedding Matrix $U \in R^{NXD}$
     • Outputs: Attention Weighted Graph Embedding Vector $h \in R^{D}$
     • Uses: Non linear weighted transform ($\tanh$) for context, sigmoid layers for att. weights, $\sum$ aggregate for h
   • Graph Interaction Extraction:
     • Inputs: Graph Embedding Vectors $h_{q}, h_{c} \in R^{D}$
     • Outputs: Interaction Score Vector $g \in R^{K}$, K being the depth of the NTN
     • Uses: Neural Tensor Network
   • Score Predictor:
     • Inputs: Graph Similarity Score Vector $g \in R^{K}$
     • Outputs: Graph Similarity Score s
     • Uses: Fully Connected Network

2. GMN Embed:

   • Encoder:
     • Inputs:
       1. Initial Node Representation Matrix $U \in R^{NXD}$
       2. Initial Edge Representation Matrix $X \in R^{NXN}$
     • Outputs: Encoded Node and Edge Embedding Vectors $H^{0} \in R^{NXD}$ and $E \in R^{NXN}$
     • Uses: Multi Layer Perceptron Networks
   • Propagation:
     • Inputs: Encoded embeddings $H^{0} \in R^{NXD}$ and $E \in R^{NXN}$