Execute this notebook: Download locally
Directed GraphSAGE with Neo4j¶
This example shows the application of directed GraphSAGE to a directed graph, where the in-node and out-node neighbourhoods are separately sampled and have different weights.
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.
[3]:
import pandas as pd
import numpy as np
import os
import stellargraph as sg
from stellargraph.connector.neo4j import Neo4jDirectedGraphSAGENodeGenerator, Neo4jStellarDiGraph
from stellargraph.layer import DirectedGraphSAGE
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.
[4]:
import py2neo
default_host = os.environ.get("STELLARGRAPH_NEO4J_HOST")
# Create the Neo4j Graph database object; port, user, password parameters can be add to specify location and authentication
neo4j_graphdb = py2neo.Graph(host=default_host)
Create a Neo4jStellarDiGraph object¶
Using the py2neo.Graph
instance, we can create a Neo4jStellarDiGraph
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 ofpaper
entities. By specifying thenode_label
parameter when constructingNeo4jStellarGraph
, this lets downstream queries to use this index and therefore run much faster, since many of the queries will require matching nodes by theirID
.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.
[5]:
neo4j_sg = Neo4jStellarDiGraph(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.
[6]:
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]
)
[7]:
# 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 Neo4jDirectedGraphSAGENodeGenerator
as we are predicting node attributes with a DirectedGraphSAGE
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 (5 in-nodes and 5 out-nodes), and 4 in the second layer (2 in-nodes and 2 out-nodes).
[8]:
batch_size = 50
in_samples = [5, 2]
out_samples = [5, 2]
epochs = 20
A Neo4jDirectedGraphSAGENodeGenerator
object is required to send the node features in sampled subgraphs to Keras.
[9]:
generator = Neo4jDirectedGraphSAGENodeGenerator(
neo4j_sg, batch_size, in_samples, out_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.
[10]:
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, which corresponds to 12 weights for each head node, 10 for each in-node and 10 for each out-node.The
bias
anddropout
are internal parameters of the model.
[11]:
graphsage_model = DirectedGraphSAGE(
layer_sizes=[32, 32], generator=generator, bias=False, dropout=0.5,
)
Now we create a model to predict the 7 categories using Keras softmax layers.
[12]:
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
[13]:
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)
[14]:
test_gen = generator.flow(test_labels.index, test_targets)
[15]:
history = model.fit(
train_gen, epochs=epochs, validation_data=test_gen, verbose=2, shuffle=False
)
Epoch 1/20
6/6 - 10s - loss: 1.9096 - acc: 0.2370 - val_loss: 1.7668 - val_acc: 0.3409
Epoch 2/20
6/6 - 9s - loss: 1.7038 - acc: 0.4741 - val_loss: 1.6493 - val_acc: 0.4606
Epoch 3/20
6/6 - 9s - loss: 1.5779 - acc: 0.5778 - val_loss: 1.5424 - val_acc: 0.6300
Epoch 4/20
6/6 - 9s - loss: 1.4403 - acc: 0.7259 - val_loss: 1.4561 - val_acc: 0.6727
Epoch 5/20
6/6 - 9s - loss: 1.3461 - acc: 0.7926 - val_loss: 1.3726 - val_acc: 0.6858
Epoch 6/20
6/6 - 9s - loss: 1.2412 - acc: 0.8519 - val_loss: 1.2986 - val_acc: 0.7059
Epoch 7/20
6/6 - 9s - loss: 1.1368 - acc: 0.8741 - val_loss: 1.2314 - val_acc: 0.7281
Epoch 8/20
6/6 - 9s - loss: 1.0673 - acc: 0.9111 - val_loss: 1.1734 - val_acc: 0.7391
Epoch 9/20
6/6 - 9s - loss: 0.9848 - acc: 0.9074 - val_loss: 1.1162 - val_acc: 0.7424
Epoch 10/20
6/6 - 9s - loss: 0.8907 - acc: 0.9296 - val_loss: 1.0722 - val_acc: 0.7408
Epoch 11/20
6/6 - 9s - loss: 0.8161 - acc: 0.9333 - val_loss: 1.0221 - val_acc: 0.7527
Epoch 12/20
6/6 - 9s - loss: 0.7358 - acc: 0.9519 - val_loss: 0.9911 - val_acc: 0.7572
Epoch 13/20
6/6 - 9s - loss: 0.7578 - acc: 0.9370 - val_loss: 0.9490 - val_acc: 0.7666
Epoch 14/20
6/6 - 9s - loss: 0.6305 - acc: 0.9667 - val_loss: 0.9255 - val_acc: 0.7662
Epoch 15/20
6/6 - 9s - loss: 0.6118 - acc: 0.9741 - val_loss: 0.8930 - val_acc: 0.7728
Epoch 16/20
6/6 - 9s - loss: 0.5467 - acc: 0.9704 - val_loss: 0.8725 - val_acc: 0.7748
Epoch 17/20
6/6 - 9s - loss: 0.5354 - acc: 0.9704 - val_loss: 0.8470 - val_acc: 0.7801
Epoch 18/20
6/6 - 9s - loss: 0.4920 - acc: 0.9852 - val_loss: 0.8328 - val_acc: 0.7785
Epoch 19/20
6/6 - 9s - loss: 0.4557 - acc: 0.9704 - val_loss: 0.8178 - val_acc: 0.7851
Epoch 20/20
6/6 - 9s - loss: 0.4396 - acc: 0.9630 - val_loss: 0.8101 - val_acc: 0.7830
[16]:
sg.utils.plot_history(history)
Now we have trained the model we can evaluate on the test set.
[17]:
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 [==============================] - 8s 164ms/step - loss: 0.8131 - acc: 0.7859
Test Set Metrics:
loss: 0.8131
acc: 0.7859
Making predictions with the model¶
Now let’s get the predictions themselves for all nodes using another node iterator:
[18]:
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
[19]:
node_predictions = target_encoding.inverse_transform(all_predictions)
Let’s have a look at a few:
[20]:
df = pd.DataFrame({"Predicted": node_predictions, "True": labels})
df.head(10)
[20]:
Predicted | True | |
---|---|---|
31336 | Neural_Networks | Neural_Networks |
1061127 | Theory | Rule_Learning |
1106406 | Reinforcement_Learning | Reinforcement_Learning |
13195 | Reinforcement_Learning | Reinforcement_Learning |
37879 | Probabilistic_Methods | Probabilistic_Methods |
1126012 | Probabilistic_Methods | Probabilistic_Methods |
1107140 | Theory | Theory |
1102850 | Neural_Networks | Neural_Networks |
31349 | Neural_Networks | Neural_Networks |
1106418 | Theory | Theory |
Please refer to the non-Neo4j directed GraphSAGE node classification demo for node embedding visualization.
Execute this notebook: Download locally