Execute this notebook: Download locally

GraphSAGE with Neo4j

This example shows the application of GraphSAGE on the Cora dataset stored in Neo4j.

Subgraphs are sampled directly from Neo4j, which eliminate the need to store the whole graph structure in NetworkX. In this demo, node features are cached in memory for faster access since the dataset is relatively small.

import pandas as pd
import os

import stellargraph as sg
from stellargraph.connector.neo4j import Neo4jGraphSAGENodeGenerator, Neo4jStellarGraph
from stellargraph.layer import GraphSAGE

import numpy as np
from tensorflow.keras import layers, optimizers, losses, metrics, Model
from sklearn import preprocessing, feature_extraction, model_selection

import time
%matplotlib inline

Connect to Neo4j

It is assumed that the Cora dataset has already been loaded into Neo4j. This notebook demonstrates how to load Cora dataset into Neo4j.

import py2neo

default_host = os.environ.get("STELLARGRAPH_NEO4J_HOST")

# Create the Neo4j Graph database object; the arguments can be edited to specify location and authentication
neo4j_graphdb = py2neo.Graph(host=default_host, port=None, user=None, password=None)

Create a Neo4jStellarGraph object

Using the py2neo.Graph instance, we can create a Neo4jStellarGraph object that will be useful for the rest of the stellargraph workflow.

For this demo, we introduce two additional details that help us execute the notebook faster:

• Our Cora dataset loaded into Neo4j has an index on the ID property of paper entities. By specifying the node_label parameter when constructing Neo4jStellarGraph, this lets downstream queries to use this index and therefore run much faster, since many of the queries will require matching nodes by their ID.

• We also cache the node features in memory. Since we know the graph is relatively small, we can safely cache the node features in memory for faster access. Note that this should be avoided for larger graphs where the node features are large.

neo4j_sg = Neo4jStellarGraph(neo4j_graphdb, node_label="paper")
neo4j_sg.cache_all_nodes_in_memory()

Data labels

Each node has a “subject” property in the database, which we want to use as labels for training. We can load the labels into memory using a Cypher query and split them into training/test sets.

labels_query = """
    MATCH (node:paper)
    WITH {ID: node.ID, label: node.subject} as node_data
    RETURN node_data
    """
rows = neo4j_graphdb.run(labels_query).data()
labels = pd.Series(
    [r["node_data"]["label"] for r in rows], index=[r["node_data"]["ID"] for r in rows]
)

# split the node labels into train/test

train_labels, test_labels = model_selection.train_test_split(
    labels, train_size=0.1, test_size=None, stratify=labels
)

target_encoding = preprocessing.LabelBinarizer()

train_targets = target_encoding.fit_transform(train_labels)
test_targets = target_encoding.transform(test_labels)

Creating the GraphSAGE model in Keras

To feed data from the graph to the Keras model we need a data generator that feeds data from the graph to the model. The generators are specialized to the model and the learning task so we choose the Neo4jGraphSAGENodeGenerator as we are predicting node attributes with a GraphSAGE model, sampling directly from Neo4j database.

We need two other parameters, the batch_size to use for training and the number of nodes to sample at each level of the model. Here we choose a two-level model with 10 nodes sampled in the first layer, and 5 in the second.

batch_size = 50
num_samples = [10, 5]
epochs = 20

A Neo4jGraphSAGENodeGenerator object is required to send the node features in sampled subgraphs to Keras.

generator = Neo4jGraphSAGENodeGenerator(neo4j_sg, batch_size, num_samples)

Using the generator.flow() method, we can create iterators over nodes that should be used to train, validate, or evaluate the model. For training we use only the training nodes returned from our splitter and the target values. The shuffle=True argument is given to the flow method to improve training.

train_gen = generator.flow(train_labels.index, train_targets, shuffle=True)

Now we can specify our machine learning model, we need a few more parameters for this:

• the layer_sizes is a list of hidden feature sizes of each layer in the model. In this example we use 32-dimensional hidden node features at each layer.

• The bias and dropout are internal parameters of the model.

graphsage_model = GraphSAGE(
    layer_sizes=[32, 32], generator=generator, bias=True, dropout=0.5,
)

Now we create a model to predict the 7 categories using Keras softmax layers.

x_inp, x_out = graphsage_model.in_out_tensors()
prediction = layers.Dense(units=train_targets.shape[1], activation="softmax")(x_out)

Training the model

Now let’s create the actual Keras model with the graph inputs x_inp provided by the graph_model and outputs being the predictions from the softmax layer

model = Model(inputs=x_inp, outputs=prediction)
model.compile(
    optimizer=optimizers.Adam(lr=0.005),
    loss=losses.categorical_crossentropy,
    metrics=["acc"],
)

Train the model, keeping track of its loss and accuracy on the training set, and its generalisation performance on the test set (we need to create another generator over the test data for this)

test_gen = generator.flow(test_labels.index, test_targets)

history = model.fit(
    train_gen, epochs=epochs, validation_data=test_gen, verbose=2, shuffle=False
)

Epoch 1/20
6/6 - 6s - loss: 1.8420 - acc: 0.3000 - val_loss: 1.6646 - val_acc: 0.3888
Epoch 2/20
6/6 - 6s - loss: 1.5992 - acc: 0.4704 - val_loss: 1.4935 - val_acc: 0.5447
Epoch 3/20
6/6 - 6s - loss: 1.4202 - acc: 0.6741 - val_loss: 1.3581 - val_acc: 0.6653
Epoch 4/20
6/6 - 6s - loss: 1.2718 - acc: 0.7889 - val_loss: 1.2407 - val_acc: 0.7215
Epoch 5/20
6/6 - 5s - loss: 1.1286 - acc: 0.8593 - val_loss: 1.1424 - val_acc: 0.7543
Epoch 6/20
6/6 - 5s - loss: 1.0310 - acc: 0.8852 - val_loss: 1.0652 - val_acc: 0.7732
Epoch 7/20
6/6 - 5s - loss: 0.9338 - acc: 0.8926 - val_loss: 1.0227 - val_acc: 0.7580
Epoch 8/20
6/6 - 5s - loss: 0.8403 - acc: 0.9296 - val_loss: 0.9651 - val_acc: 0.7633
Epoch 9/20
6/6 - 5s - loss: 0.7482 - acc: 0.9481 - val_loss: 0.9277 - val_acc: 0.7908
Epoch 10/20
6/6 - 6s - loss: 0.6885 - acc: 0.9778 - val_loss: 0.8926 - val_acc: 0.7945
Epoch 11/20
6/6 - 6s - loss: 0.6433 - acc: 0.9667 - val_loss: 0.8423 - val_acc: 0.8068
Epoch 12/20
6/6 - 5s - loss: 0.5693 - acc: 0.9593 - val_loss: 0.8221 - val_acc: 0.8015
Epoch 13/20
6/6 - 5s - loss: 0.5259 - acc: 0.9778 - val_loss: 0.8009 - val_acc: 0.7986
Epoch 14/20
6/6 - 5s - loss: 0.4732 - acc: 0.9926 - val_loss: 0.7852 - val_acc: 0.7974
Epoch 15/20
6/6 - 5s - loss: 0.4317 - acc: 0.9889 - val_loss: 0.7527 - val_acc: 0.8093
Epoch 16/20
6/6 - 5s - loss: 0.4011 - acc: 0.9926 - val_loss: 0.7533 - val_acc: 0.8019
Epoch 17/20
6/6 - 5s - loss: 0.3802 - acc: 0.9889 - val_loss: 0.7341 - val_acc: 0.7986
Epoch 18/20
6/6 - 5s - loss: 0.3410 - acc: 0.9963 - val_loss: 0.7174 - val_acc: 0.8002
Epoch 19/20
6/6 - 5s - loss: 0.3153 - acc: 0.9815 - val_loss: 0.7141 - val_acc: 0.8023
Epoch 20/20
6/6 - 5s - loss: 0.2734 - acc: 1.0000 - val_loss: 0.7045 - val_acc: 0.8121

sg.utils.plot_history(history)

Now we have trained the model we can evaluate on the test set.

test_metrics = model.evaluate(test_gen)
print("\nTest Set Metrics:")
for name, val in zip(model.metrics_names, test_metrics):
    print("\t{}: {:0.4f}".format(name, val))

49/49 [==============================] - 5s 92ms/step - loss: 0.7004 - acc: 0.8105

Test Set Metrics:
        loss: 0.7004
        acc: 0.8105

Making predictions with the model

Now let’s get the predictions themselves for all nodes using another node iterator:

all_nodes = labels.index
all_mapper = generator.flow(all_nodes)
all_predictions = model.predict(all_mapper)

These predictions will be the output of the softmax layer, so to get final categories we’ll use the inverse_transform method of our target attribute specification to turn these values back to the original categories

node_predictions = target_encoding.inverse_transform(all_predictions)

Let’s have a look at a few:

df = pd.DataFrame({"Predicted": node_predictions, "True": labels})
df.head(10)

	Predicted	True
31336	Neural_Networks	Neural_Networks
1061127	Rule_Learning	Rule_Learning
1106406	Reinforcement_Learning	Reinforcement_Learning
13195	Reinforcement_Learning	Reinforcement_Learning
37879	Probabilistic_Methods	Probabilistic_Methods
1126012	Probabilistic_Methods	Probabilistic_Methods
1107140	Reinforcement_Learning	Theory
1102850	Neural_Networks	Neural_Networks
31349	Neural_Networks	Neural_Networks
1106418	Theory	Theory

Please refer to the non-Neo4j GraphSAGE node classification demo for node embedding visualization.

Execute this notebook: Download locally