GraphAttentionLayerCore<T>

Models & Types · AiDotNet.Finance.Portfolio

The graph attention computation used by GraphAttentionPortfolio (Korangi, Mues and Bravo, arXiv:2407.15532, following Velickovic et al.): a(u,v) = softmax_v( LeakyReLU( a^T [W x_u || W x_v] ) ), aggregated over neighbours and CONCATENATED across heads.

For Beginners

Each company looks only at the companies it is connected to in the network, decides how much attention to pay to each, and builds a summary from them.

How It Works

What separates this from ordinary dense attention is the MASK. Each node attends only over its graph neighbourhood, and the softmax is normalized within that neighbourhood — so the sparsity that TMFG filtering produced is what the attention actually operates on. Attending over all pairs would discard the filtered graph entirely and reduce the model to a plain transformer over assets.

Heads are CONCATENATED, not averaged, following the paper’s || operator. Averaging would collapse the heads into a single representation of the same width and lose the distinct relational views the heads learn.

Constructors

Constructor Summary
GraphAttentionLayerCore(double slope) Creates the attention core.

Properties

Property Summary
double Slope { get; } Gets the LeakyReLU negative slope.

Methods

Method Summary
Tensor<T> Aggregate(Tensor<T> coefficients, Tensor<T> projected) Aggregates neighbour features under the attention coefficients: h_u = sum_v alpha_uv * (W x_v), then applies ReLU.
Tensor<T> AttentionCoefficients(Tensor<T> projected, Vector<T> attentionVector, Tensor<T> adjacency) Attention coefficients over a masked graph, one row per node, normalized within each neighbourhood.
Tensor<T> ConcatenateHeads(IReadOnlyList<Tensor<T>> headOutputs) Concatenates per-head outputs along the feature axis, as the paper’s || operator does.
double LeakyReLU(double x) LeakyReLU.

Fields

Field Summary
double LeakyReLUSlope The paper’s LeakyReLU negative slope.