11.3 Quick Start: Using Graph Machine Learning on PGQL Property Graphs

This tutorial helps you get started on applying the DeepWalk machine learning algorithm on a PGQL property graph.

The instructions assume that the PGQL property graph is already existing in your current database.

Run the following steps to build and work with a Deep Walk model.

1. Load the PGQL property graph into the graph server (PGX).
   • JShell
   • Java
   • Python

   opg4j> var instance = GraphServer.getInstance("https://localhost:7007", "<username>", "<password>".toCharArray())
   instance ==> ServerInstance[embedded=false,baseUrl=https://localhost:7007]
   opg4j> var session=instance.createSession("mySession")
   session ==> PgxSession[ID=5af9c362-10a3-4a7c-953c-602553d4606b,source=mySession]
   opg4j> var graph = session.readGraphByName("BANK_GRAPH",GraphSource.PG_VIEW)
   graph ==> PgxGraph[name=BANK_GRAPH,N=1000,E=4997,created=1684315831352]

   ServerInstance instance = GraphServer.getInstance("https://localhost:7007", "<username>", "<password>".toCharArray());
   PgxSession session = instance.createSession("my-session");
   PgxGraph graph = session.readGraphByName("BANK_GRAPH",GraphSource.PG_VIEW);

   >>> instance = graph_server.get_instance("https://localhost:7007","<username>","<password>")
   >>> session = instance.create_session("my_session")
   >>> graph = session.read_graph_by_name("BANK_GRAPH", "pg_view")
   >>> graph
   PgxGraph(name: BANK_GRAPH, v: 1000, e: 4997, directed: True, memory(Mb): 0)

2. Build a Deep Walk model using customized hyper-parameters as shown:
   • JShell
   • Java
   • Python

   opg4j> var model = session.createAnalyst().deepWalkModelBuilder().
   ...>                 setMinWordFrequency(1).
   ...>                 setBatchSize(512).
   ...>                 setNumEpochs(1).
   ...>                 setLayerSize(100).
   ...>                 setLearningRate(0.05).
   ...>                 setMinLearningRate(0.0001).
   ...>                 setWindowSize(3).
   ...>                 setWalksPerVertex(6).
   ...>                 setWalkLength(4).
   ...>                 setNegativeSample(2).
   ...>                 build()
   model ==> oracle.pgx.api.mllib.DeepWalkModel@6e0f259e

   import oracle.pgx.api.mllib.DeepWalkModel;
   DeepWalkModel model= session.createAnalyst().deepWalkModelBuilder()
           .setMinWordFrequency(1)
           .setBatchSize(512)
           .setNumEpochs(1)
           .setLayerSize(100)
           .setLearningRate(0.05)
           .setMinLearningRate(0.0001)
           .setWindowSize(3)
           .setWalksPerVertex(6)
           .setWalkLength(4)
           .setNegativeSample(2)
           .build();

   >>> model = session.create_analyst().deepwalk_builder(min_word_frequency= 1,
   ...                                 batch_size= 512,
   ...                                 num_epochs= 1,
   ...                                 layer_size= 100,
   ...                                 learning_rate= 0.05,
   ...                                 min_learning_rate= 0.0001,
   ...                                 window_size= 3,
   ...                                 walks_per_vertex= 6,
   ...                                 walk_length= 4,
   ...                                 negative_sample= 2)

3. Train the Deep Walk model as shown:
   • JShell
   • Java
   • Python

   opg4j> model.fit(graph)

   model.fit(graph);

   >>> model.fit(graph)

4. Get the loss value as shown:
   • JShell
   • Java
   • Python

   opg4j> var loss = model.getLoss()
   loss ==> -2.097562355629634E-5

   double loss = model.getLoss();

   >>> loss = model.loss
   >>> loss
   -2.0706271243398078e-05

5. Compute similar vertices as shown:
   • JShell
   • Java
   • Python

   opg4j> var similars = model.computeSimilars("ACCOUNTS(280)",10)
   batchSimilars ==> oracle.pgx.api.frames.internal.PgxFrameImpl@308e465b
   opg4j> batchSimilars.print()

   import oracle.pgx.api.frames.*;
   
   PgxFrame similars = model.computeSimilars("ACCOUNTS(280)", 10);
   similars.print();

   >>> similars = model.compute_similars("ACCOUNTS(280)",10)
   >>> similars.print()

   The example produces a similar output:

   +-------------------------------------+
   | dstVertex     | similarity          |
   +-------------------------------------+
   | ACCOUNTS(280) | 1.0                 |
   | ACCOUNTS(486) | 0.3253505229949951  |
   | ACCOUNTS(615) | 0.2806776463985443  |
   | ACCOUNTS(660) | 0.27348122000694275 |
   | ACCOUNTS(737) | 0.2734076678752899  |
   | ACCOUNTS(368) | 0.2707795202732086  |
   | ACCOUNTS(479) | 0.27019545435905457 |
   | ACCOUNTS(845) | 0.2618815004825592  |
   | ACCOUNTS(834) | 0.2543807625770569  |
   | ACCOUNTS(249) | 0.24260951578617096 |
   +-------------------------------------+

6. Get all trained vectors and store them in a database table as shown:
   • JShell
   • Java
   • Python

   opg4j> var vertexVectors = model.getTrainedVertexVectors().flattenAll()
   vertexVectors ==> oracle.pgx.api.frames.internal.PgxFrameImpl@46cb9794
   opg4j> vertexVectors.write().db().name("deepwalkframe").tablename("vertexVectors").overwrite(true).store()

   PgxFrame vertexVectors = model.getTrainedVertexVectors().flattenAll();
   vertexVectors.write()
     .db()
     .name("vertex vectors")
     .tablename("vertexVectors")
     .overwrite(true)
     .store();

   >>> vertex_vectors = model.trained_vectors.flatten_all()
   >>> vertex_vectors.write().db(). \
   ...         table_name("vertex_vectors"). \
   ...         overwrite(True). \
   ...         store()

7. Store the trained model in the database as shown:
   • JShell
   • Java
   • Python

   opg4j> model.export().db().
   ...>   modelstore("bank_model").
   ...>   modelname("model").
   ...>   description("DeepWalk Model for Bank data").
   ...>   store()

   model.export().db()
     .modelstore("bank_model")  
     .modelname("model2")                 
     .description("DeepWalk Model for Bank data") 
     .store();

   >>> model.export().db(model_store="bank_model",
   ...                   model_name="model",
   ...                   model_description="DeepWalk Model for Bank data")

8. Load a pre-trained model from the database as shown:
   • JShell
   • Java
   • Python

   opg4j> session.createAnalyst().loadDeepWalkModel().db().
   ...>   modelstore("bank_model").
   ...>   modelname("model").
   ...>   load()

   model = session.createAnalyst().loadDeepWalkModel().db()
     .modelstore("bank_model")
     .modelname("model")   
     .load();

   >>> model = session.create_analyst().get_deepwalk_model_loader().db(model_store="bank_model",
   ...                                        model_name="model")

9. Destroy the model as shown:
   • JShell
   • Java
   • Python

   opg4j> model.destroy()

   model.destroy();

   >>> model.destroy()

   See Using the Machine Learning Library (PgxML) for Graphs for more information on the supported machine learning algorithms.