GAEAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.DeepLearning

GAE — Graph Autoencoder for causal discovery.

Inheritance: CausalDiscoveryBase<T> → DeepCausalBase<T> → GAEAlgorithm<T>

For Beginners

A Graph Autoencoder compresses data through a “graph bottleneck.” The connections in this bottleneck represent causal relationships — the autoencoder is forced to find the minimal set of connections needed to recreate the data.

How It Works

GAE uses an autoencoder architecture where the encoder produces a latent graph representation and the decoder reconstructs the data through the learned graph. The encoder maps each variable to separate source/target latent embeddings via shared MLP, then computes edge probabilities as sigmoid(Zs_i^T * Zt_j) using asymmetric dot products to encode directionality. The decoder reconstructs X_hat = X * A where A is the soft adjacency. NOTEARS acyclicity constraint h(A) = tr(exp(A∘A)) - d ensures a valid DAG.

Reference: Kipf and Welling (2016), “Variational Graph Auto-Encoders”, NeurIPS Workshop.

Constructors

Constructor Summary
GAEAlgorithm(CausalDiscoveryOptions options)

Properties

Property Summary
bool SupportsNonlinear { get; }
string Name { get; }