Lernen Link Prediction and Negative Sampling

In knowledge graphs, link prediction is the task of inferring missing connections between entities. You try to predict whether a relationship (or "link") exists between two entities, given the existing structure of the graph. For example, if your knowledge graph contains the triples ("Paris", "isCapitalOf", "France") and ("Berlin", "isCapitalOf", "Germany"), you might predict whether ("Rome", "isCapitalOf", "Italy") should also exist.

To train models for link prediction, you need both positive samples (triples that exist in the graph) and negative samples (triples that do not exist). Negative samples are crucial because most knowledge graphs only store true facts; without negatives, a model cannot learn what a false or implausible triple looks like. Negative sampling involves generating triples that are not present in the knowledge graph, usually by corrupting existing triples—such as replacing the head or tail entity with a random, unrelated entity. This process is essential for both training models to distinguish true from false facts and for evaluating their performance accurately.


              1234567891011121314151617181920212223242526272829303132333435363738
            
import random

# Define a toy knowledge graph as a set of triples (head, relation, tail)
triples = [
    ("Paris", "isCapitalOf", "France"),
    ("Berlin", "isCapitalOf", "Germany"),
    ("Madrid", "isCapitalOf", "Spain"),
]

entities = {"Paris", "France", "Berlin", "Germany", "Madrid", "Spain"}
relations = {"isCapitalOf"}

# Generate positive triples (those in the graph)
positive_triples = triples.copy()

# Generate negative triples by corrupting either head or tail entity
def generate_negative_triples(triples, entities, num_negatives=3):
    negatives = []
    for h, r, t in triples:
        # Corrupt head
        corrupted_head = random.choice(list(entities - {h}))
        negatives.append((corrupted_head, r, t))
        # Corrupt tail
        corrupted_tail = random.choice(list(entities - {t}))
        negatives.append((h, r, corrupted_tail))
        if len(negatives) >= num_negatives:
            break
    return negatives[:num_negatives]

negative_triples = generate_negative_triples(triples, entities, num_negatives=4)

print("Positive triples:")
for triple in positive_triples:
    print(triple)

print("\nNegative triples:")
for triple in negative_triples:
    print(triple)

1. Why is negative sampling important in knowledge graph learning?

2. What is a negative triple?

War alles klar?

Danke für Ihr Feedback!

Abschnitt 3. Kapitel 1

Fragen Sie AI

Fragen Sie alles oder probieren Sie eine der vorgeschlagenen Fragen, um unser Gespräch zu beginnen

Suggested prompts:

Can you explain why negative sampling is important in link prediction?

How does the negative sampling function work in this code?

Can you show how to generate more negative samples or change the sampling strategy?

Swipe um das Menü anzuzeigen


              1234567891011121314151617181920212223242526272829303132333435363738
            
import random

# Define a toy knowledge graph as a set of triples (head, relation, tail)
triples = [
    ("Paris", "isCapitalOf", "France"),
    ("Berlin", "isCapitalOf", "Germany"),
    ("Madrid", "isCapitalOf", "Spain"),
]

entities = {"Paris", "France", "Berlin", "Germany", "Madrid", "Spain"}
relations = {"isCapitalOf"}

# Generate positive triples (those in the graph)
positive_triples = triples.copy()

# Generate negative triples by corrupting either head or tail entity
def generate_negative_triples(triples, entities, num_negatives=3):
    negatives = []
    for h, r, t in triples:
        # Corrupt head
        corrupted_head = random.choice(list(entities - {h}))
        negatives.append((corrupted_head, r, t))
        # Corrupt tail
        corrupted_tail = random.choice(list(entities - {t}))
        negatives.append((h, r, corrupted_tail))
        if len(negatives) >= num_negatives:
            break
    return negatives[:num_negatives]

negative_triples = generate_negative_triples(triples, entities, num_negatives=4)

print("Positive triples:")
for triple in positive_triples:
    print(triple)

print("\nNegative triples:")
for triple in negative_triples:
    print(triple)

1. Why is negative sampling important in knowledge graph learning?

2. What is a negative triple?

War alles klar?

Danke für Ihr Feedback!

Abschnitt 3. Kapitel 1