Graph Classifier Sample Clauses

Graph Classifier. After constructing the resume undirected graph, we feed the graph into a Graph Convolutional Network (GCN) [4], which is formulated as L(i+1) = σ(D˜− 1 A˜D˜− 1 L(i)W (i)), where i denotes the layer number, σ is the non-linear activation function, e.g. a ReLU σ(x) = max(0, x) in our case, A˜ is the adjacent matrix of the graph with additional self-connections, D˜ is the degree matrix where D˜ii = Σj A˜ij , and W (i) is the learnable weight matrix of the i-th layer of GCN. The input node features L(i) are represented by the context embeddings of nodes and the graph-level embedding is derived by a mean pooling operation on the node- level embedding output. Then, an MLP is attached to produce the predicted label yˆ, that is, yˆ = MLP (Pooling(L(i+1))). In addition, the cross-entropy loss Σdi∈D −p(yˆi) log p(yi) is computed over all labeled resumes.