16.5.5 Applying Unsupervised EdgeWise for Partitioned Graphs

You can apply unsupervised edgewise on partitioned graphs, where you have different providers and different features.

• JShell
• Java
• Python

opg4j> var model = analyst.unsupervisedEdgeWiseModelBuilder().
         setVertexInputPropertyNames("vertex_provider_features").
         setEdgeInputPropertyNames("edge_provider1_features", "edge_provider2_features").
         build()

UnsupervisedEdgeWiseModel model = analyst.unsupervisedEdgeWiseModelBuilder()
    .setVertexInputPropertyNames("vertex_provider_features")
    .setEdgeInputPropertyNames("edge_provider1_features", "edge_provider2_features")
    .build();

params = dict(vertex_input_property_names=["vertex_provider_features"],
              edge_input_property_names=["edge_provider1_features", "edge_provider2_features"])

model = analyst.unsupervised_edgewise_builder(**params)

You can select which providers you want to train or infer on:

• JShell
• Java
• Python

opg4j> var model = analyst.unsupervisedEdgeWiseModelBuilder().
         setVertexInputPropertyNames("vertex_provider_features").
         setEdgeInputPropertyNames("edge_provider1_features", "edge_provider2_features").
         setTargetEdgeLabels("provider1").
         build()

UnsupervisedEdgeWiseModel model = analyst.unsupervisedEdgeWiseModelBuilder()
    .setVertexInputPropertyNames("vertex_provider_features")
    .setEdgeInputPropertyNames("edge_provider1_features", "edge_provider2_features")
    .setTargetEdgeLabels("provider1")
    .build();

params = dict(vertex_input_property_names=["vertex_provider_features"],
              edge_input_property_names=["edge_provider1_features", "edge_provider2_features"],
              target_edge_labels=["provider1"])

model = analyst.unsupervised_edgewise_builder(**params)

If you wish to control the flow of the embeddings at each graph convolutional layer of the underlying Graphwise model, then you can enable or disable the connections of interest. By default, all the connections are enabled.

• JShell
• Java
• Python

opg4j> var convLayerConfig = analyst.graphWiseConvLayerConfigBuilder().
         setNumSampledNeighbors(25).
         useVertexToVertexConnection(true).
         useEdgeToVertexConnection(true).
         useEdgeToEdgeConnection(false).
         useVertexToEdgeConnection(false).
         build()
opg4j> var model = analyst.unsupervisedEdgeWiseModelBuilder().
         setVertexInputPropertyNames("vertex_provider1_features", "vertex_provider2_features").
         setEdgeInputPropertyNames("edge_provider_features").
         setTargetEdgeLabels("provider1").
         build()

GraphWiseConvLayerConfig convLayerConfig = analyst.graphWiseConvLayerConfigBuilder()
    .setNumSampledNeighbors(10)
    .useVertexToVertexConnection(true)
    .useEdgeToVertexConnection(true)
    .useEdgeToEdgeConnection(false)
    .useVertexToEdgeConnection(false)
    .build();

UnsupervisedEdgeWiseModel model = analyst.unsupervisedEdgeWiseModelBuilder()
    .setVertexInputPropertyNames("vertex_provider1_features", "vertex_provider2_features")
    .setEdgeInputPropertyNames("edge_provider_features")
    .setTargetEdgeLabels("provider1")
    .setConvLayerConfigs(convLayerConfig)
    .build();

conv_layer_config = dict(num_sampled_neighbors=25,
                         activation_fn='tanh',
                         weight_init_scheme='xavier',
                         neighbor_weight_property_name=weightProperty,
                         vertex_to_vertex_connection=True,
                         edge_to_vertex_connection=True,
                         vertex_to_edge_connection=False,
                         edge_to_edge_connection=False)

conv_layer = analyst.graphwise_conv_layer_config(**conv_layer_config)

params = dict(vertex_input_property_names=["vertex_provider1_features", "vertex_provider2_features"],
              edge_input_property_names=["edge_provider_features"],
              target_edge_labels=["provider1"],
              conv_layer_config=[conv_layer])

model = analyst.unsupervised_edgewise_builder(**params)