Negative sampling

As explained in Graph Neural Networks in Action

A training technique that selects a subset of nonexistent relationships as negative examples instead of evaluating all possible negative examples.

Negative sampling trains link-prediction models with a manageable subset of nonexistent edges instead of evaluating every possible negative relationship. It addresses the class imbalance between existing and nonexistent edges. section 5.2.3

A graph autoencoder training step follows this pipeline:

• Encode the graph into a latent representation. section 5.2.3
• Draw new nonexistent edges each epoch with negative_sampling, matching the number of training edge labels. section 5.2.3
• Concatenate positive and negative edge indices, decode them, and assign labels 1 and 0 respectively. section 5.2.3
• Compute binary cross-entropy, backpropagate, and update the model; evaluate validation and test edges without resampling. section 5.2.3

PyTorch Geometric training step

Call negative_sampling on train_data.edge_index, use the returned neg_edge_index with the positive edge_label_index, and create matching zero labels before calculating BCEWithLogitsLoss. section 5.2.3

Figure 5.6 Overall steps for training our model for link prediction

Follow the end-to-end link-prediction pipeline from graph encoding through sampled training edges and evaluation. section 5.2.3

How Graph Neural Networks in Action covers it

How deep it goes

Explains per-epoch sampling, edge concatenation, zero-label construction, loss calculation, and evaluation without sampling.

Examples

Uses a PyTorch Geometric training function that samples sparse negative edges and trains for repeated epochs.

What it assumes

Assumes graph edges, encoders, decoders, binary labels, and training loops while demonstrating their interaction.

What you can do after

Apply it via PyTorch Geometric to construct sampled edge-label batches for graph-autoencoder training.

How it connects

Connects graph autoencoders, link prediction, binary cross-entropy, latent representations, and ROC/AUC evaluation.

liveBookmanning.com

Read Graph Neural Networks in Action on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  Knowledge Graphs and LLMs in ActionNegative sampling is a training technique that selects unobserved or candidate relationships as negative examples instead of evaluating every possible relationship.

Compare how 2 books explain negative sampling →

Related concepts

• Word2Vecused by · Word2Vec is a family of shallow neural models that learns vector representations of words from their contexts in text.
• cross entropy lossuses · Cross-entropy loss measures the expected negative log probability that a predicted distribution assigns to outcomes drawn from a target distribution.
• class imbalanceaddresses · Class imbalance is a condition in a classification dataset where classes have substantially unequal numbers of examples.

This explainer was generated by Manning's knowledge graph from the text of Graph Neural Networks in Action. The section links point to the parts of the book it draws on.