Unsupervised Graph Learning with GraphSage

GraphScope provides the capability to process learning tasks. In this tutorial, we demostrate how GraphScope trains a model with GraphSage.

The task is link prediction, which estimates the probability of links between nodes in a graph.

In this task, we use our implementation of GraphSAGE algorithm to build a model that predicts protein-protein links in the PPI dataset. In which every node represents a protein. The task can be treated as a unsupervised link prediction on a homogeneous link network.

In this task, GraphSage algorithm would compress both structural and attribute information in the graph into low-dimensional embedding vectors on each node. These embeddings can be further used to predict links between nodes.

This tutorial has following steps:

  • Creating session and loading graph
  • Launching the learning engine and attaching to loaded graph.
  • Defining train process with builtin GraphSage model and hyperparameters
  • Training and evaluating

First, let's create a session and load the dataset as a graph.

In [ ]:
import os
import graphscope

k8s_volumes = {
    "data": {
        "type": "hostPath",
        "field": {
          "path": "/testingdata",
          "type": "Directory"
        },
        "mounts": {
          "mountPath": "/home/jovyan/datasets",
            "readOnly": True
        }
    }
}

# create session
graphscope.set_option(show_log=True)
sess = graphscope.session(k8s_volumes=k8s_volumes)

# loading ppi graph
graph = sess.load_from(
    edges={
        "link": [
            (
                "/home/jovyan/datasets/ppi/edge.csv",
                [],
                ("src_id", "protein"),
                ("dst_id", "protein"),
            ),
        ]
    },
    vertices={
        "protein": (
            "/home/jovyan/datasets/ppi/node.csv",
        ),
    },
)

Launch learning engine

Then, we need to define a feature list for training. The training feature list should be seleted from the vertex properties. In this case, we choose all the properties prefix with "feat-" as the training features.

With the featrue list, next we launch a learning engine with the learning method of session. (You may find the detail of the method on Session.)

In this case, we specify the GCN training over protein nodes and link edges.

With gen_labels, we take protein nodes as training set.

In [ ]:
# define the features for learning
paper_features = []
for i in range(50):
    paper_features.append("feat-" + str(i))

# launch a learning engine.
lg = sess.learning(graph, nodes=[("protein", paper_features)],
                  edges=[("protein", "link", "protein")],
                  gen_labels=[
                      ("train", "protein", 100, (0, 100)),
                  ])

We use the builtin GraphSage model to define the training process. You can find more detail about all the builtin learning models on Graph Learning Model

In the example, we use tensorflow as NN backend trainer.

In [ ]:
import numpy as np
from graphscope.learning.examples import GraphSage
from graphscope.learning.graphlearn.python.model.tf.trainer import LocalTFTrainer
from graphscope.learning.graphlearn.python.model.tf.optimizer import get_tf_optimizer

# unsupervised GraphSage.

def train(config, graph):
    def model_fn():
        return GraphSage(
            graph,
            config["class_num"],
            config["features_num"],
            config["batch_size"],
            categorical_attrs_desc=config["categorical_attrs_desc"],
            hidden_dim=config["hidden_dim"],
            in_drop_rate=config["in_drop_rate"],
            neighs_num=config["neighs_num"],
            hops_num=config["hops_num"],
            node_type=config["node_type"],
            edge_type=config["edge_type"],
            full_graph_mode=config["full_graph_mode"],
            unsupervised=config['unsupervised'],
        )
    trainer = LocalTFTrainer(
        model_fn,
        epoch=config["epoch"],
        optimizer=get_tf_optimizer(
            config["learning_algo"], config["learning_rate"], config["weight_decay"]
        ),
    )
    trainer.train()
    embs = trainer.get_node_embedding()
    np.save(config['emb_save_dir'], embs)

# define hyperparameters
config = {
    "class_num": 128,  # output dimension
    "features_num": 50,
    "batch_size": 512,
    "categorical_attrs_desc": "",
    "hidden_dim": 128,
    "in_drop_rate": 0.5,
    "hops_num": 2,
    "neighs_num": [5, 5],
    "full_graph_mode": False,
    "agg_type": "gcn",  # mean, sum
    "learning_algo": "adam",
    "learning_rate": 0.01,
    "weight_decay": 0.0005,
    'unsupervised': True,
    "epoch": 1,
    'emb_save_dir': './id_emb',
    "node_type": "protein",
    "edge_type": "link",
}

Run training process

After define training process and hyperparameters,

Now we can start the traning process with learning engine lg and the hyperparameters configurations.

In [ ]:
train(config, lg)

Finally, don't forget to close the session.

In [ ]:
sess.close()
In [ ]: