机器学习管道：节点分类

%pip install graphdatascience

import os

from graphdatascience import GraphDataScience

# Get Neo4j DB URI, credentials and name from environment if applicable
NEO4J_URI = os.environ.get("NEO4J_URI", "bolt://:7687")
NEO4J_AUTH = None
NEO4J_DB = os.environ.get("NEO4J_DB", "neo4j")
if os.environ.get("NEO4J_USER") and os.environ.get("NEO4J_PASSWORD"):
    NEO4J_AUTH = (
        os.environ.get("NEO4J_USER"),
        os.environ.get("NEO4J_PASSWORD"),
    )
gds = GraphDataScience(NEO4J_URI, auth=NEO4J_AUTH, database=NEO4J_DB)

# On AuraDS:
#
# gds = GraphDataScience(NEO4J_URI, auth=NEO4J_AUTH, database=NEO4J_DB, aura_ds=True)

from graphdatascience import ServerVersion

assert gds.server_version() >= ServerVersion(2, 5, 0)

import json

import numpy as np
import pandas as pd

CORA_CONTENT = "https://data.neo4j.com/cora/cora.content"
CORA_CITES = "https://data.neo4j.com/cora/cora.cites"

SUBJECT_TO_ID = {
    "Neural_Networks": 100,
    "Rule_Learning": 1,
    "Reinforcement_Learning": 2,
    "Probabilistic_Methods": 3,
    "Theory": 4,
    "Genetic_Algorithms": 5,
    "Case_Based": 6,
}

HOLDOUT_NODES = 10

# Define a string representation of the SUBJECT_TO_ID map using backticks
subject_map = json.dumps(SUBJECT_TO_ID).replace('"', "`")

# Cypher command to load the nodes using `LOAD CSV`, taking care of
# converting the string `subject` field into an integer and
# replacing the node label for the holdout nodes
load_nodes = f"""
    LOAD CSV FROM "{CORA_CONTENT}" AS row
    WITH
      {subject_map} AS subject_to_id,
      toInteger(row[0]) AS extId,
      row[1] AS subject,
      toIntegerList(row[2..]) AS features
    MERGE (p:Paper {{extId: extId, subject: subject_to_id[subject], features: features}})
    WITH p LIMIT {HOLDOUT_NODES}
    REMOVE p:Paper
    SET p:UnclassifiedPaper
"""

# Cypher command to load the relationships using `LOAD CSV`
load_relationships = f"""
    LOAD CSV FROM "{CORA_CITES}" AS row
    MATCH (n), (m)
    WHERE n.extId = toInteger(row[0]) AND m.extId = toInteger(row[1])
    MERGE (n)-[:CITES]->(m)
"""

# Load nodes and relationships on Neo4j
gds.run_cypher(load_nodes)
gds.run_cypher(load_relationships)

# Create the projected graph containing both classified and unclassified nodes
G, _ = gds.graph.project(
    "cora-graph",
    {"Paper": {"properties": ["features", "subject"]}, "UnclassifiedPaper": {"properties": ["features"]}},
    {"CITES": {"orientation": "UNDIRECTED", "aggregation": "SINGLE"}},
)

assert G.node_count() == 2708
assert G.relationship_count() == 10556

# Create the pipeline
node_pipeline, _ = gds.beta.pipeline.nodeClassification.create("cora-pipeline")

# List all pipelines
gds.pipeline.list()

# Alternatively, get the details of a specific pipeline object
gds.pipeline.list(node_pipeline)

# "Mark" some node properties that will be used as features
node_pipeline.selectFeatures(["features"])

# If needed, change the train/test split ratio and the number of folds
# for k-fold cross-validation
node_pipeline.configureSplit(testFraction=0.2, validationFolds=5)

# Add a model candidate to train
node_pipeline.addLogisticRegression(maxEpochs=200, penalty=(0.0, 0.5))

# Explicit set the number of trials for autotuning (default = 10)
node_pipeline.configureAutoTuning(maxTrials=5)

# Estimate the resources needed for training the model
node_pipeline.train_estimate(
    G,
    targetNodeLabels=["Paper"],
    modelName="cora-pipeline-model",
    targetProperty="subject",
    metrics=["F1_WEIGHTED"],
    randomSeed=42,
    concurrency=4,
)

# Perform the actual training
model, stats = node_pipeline.train(
    G,
    targetNodeLabels=["Paper"],
    modelName="cora-pipeline-model",
    targetProperty="subject",
    metrics=["F1_WEIGHTED"],
    randomSeed=42,
    concurrency=4,
)

# Uncomment to print all stats
# print(stats.to_json(indent=2))

# Print F1_WEIGHTED metric
stats["modelInfo"]["metrics"]["F1_WEIGHTED"]["test"]

predicted = model.predict_stream(
    G, modelName="cora-pipeline-model", includePredictedProbabilities=True, targetNodeLabels=["UnclassifiedPaper"]
)

predicted

# List of class labels
classes = stats["modelInfo"]["classes"]
print("Class labels:", classes)

# Calculate the confidence percentage for the predicted class
predicted["confidence"] = predicted.apply(
    lambda row: np.floor(row["predictedProbabilities"][classes.index(row["predictedClass"])] * 100), axis=1
)

predicted

node_pipeline_fastrp, _ = gds.beta.pipeline.nodeClassification.create("cora-pipeline-fastrp")

# Add a step in the pipeline that mutates the graph
node_pipeline_fastrp.addNodeProperty(
    "fastRP",
    mutateProperty="embedding",
    embeddingDimension=512,
    propertyRatio=1.0,
    randomSeed=42,
    featureProperties=["features"],
    contextNodeLabels=["Paper", "UnclassifiedPaper"],
)

# With the node embeddings available as features, we no longer use the original raw `features`.
node_pipeline_fastrp.selectFeatures(["embedding"])

# Configure the pipeline as before
node_pipeline_fastrp.configureSplit(testFraction=0.2, validationFolds=5)
node_pipeline_fastrp.addLogisticRegression(maxEpochs=200, penalty=(0.0, 0.5))
node_pipeline.configureAutoTuning(maxTrials=5)

# Perform the actual training
model_fastrp, stats_fastrp = node_pipeline_fastrp.train(
    G,
    targetNodeLabels=["Paper"],
    modelName="cora-pipeline-model-fastrp",
    targetProperty="subject",
    metrics=["F1_WEIGHTED"],
    randomSeed=42,
    concurrency=4,
)

print(stats_fastrp["modelInfo"]["metrics"]["F1_WEIGHTED"]["test"])

predicted_fastrp = model_fastrp.predict_stream(
    G,
    modelName="cora-pipeline-model-fastrp",
    includePredictedProbabilities=True,
    targetNodeLabels=["UnclassifiedPaper"],
)

print(len(predicted_fastrp))

model_fastrp.predict_mutate(
    G,
    mutateProperty="predictedClass",
    modelName="cora-pipeline-model-fastrp",
    predictedProbabilityProperty="predictedProbabilities",
    targetNodeLabels=["UnclassifiedPaper"],
)

predicted_fastrp = gds.graph.nodeProperty.stream(G, "predictedClass", ["UnclassifiedPaper"])

predicted_fastrp

# Retrieve node information from Neo4j using the node IDs from the prediction result
nodes = gds.util.asNodes(predicted_fastrp.nodeId.to_list())

# Create a new DataFrame containing node IDs along with node properties
nodes_df = pd.DataFrame([(node.id, node["subject"]) for node in nodes], columns=["nodeId", "subject"])

# Merge with the prediction result on node IDs, to check the predicted value
# against the original subject
#
# NOTE: This could also be replaced by just appending `node["subject"]` as a
# Series since the node order would not change, but a proper merge (or join)
# is clearer and less prone to errors.
predicted_fastrp.merge(nodes_df, on="nodeId")

gds.graph.nodeProperties.write(
    G,
    node_properties=["predictedClass"],
    node_labels=["UnclassifiedPaper"],
)

model.drop()
model_fastrp.drop()
node_pipeline.drop()
node_pipeline_fastrp.drop()

G.drop()

gds.run_cypher("MATCH (n) WHERE n:Paper OR n:UnclassifiedPaper DETACH DELETE n")

gds.close()