Message Passing Aggregators

Aggregators combine neighbor node features in graph neural networks. Each aggregator implements a different reduction over the features of a node's neighbors, and is passed to a layer that consumes one, such as layer_sage().

Usage

Aggregator(name = character(0), learnable = logical(0))

SumAggregator()

MeanAggregator()

MaxAggregator()

MinAggregator()

ProductAggregator()

VarAggregator()

StdAggregator()

Arguments

name

Character scalar. Short identifier for the reduction, such as "sum" or "mean".

learnable

Logical scalar. Whether the aggregator holds parameters that are updated during training.

Value

An S7 object inheriting from Aggregator, with properties name and learnable.

Details

Available aggregators:

• SumAggregator(): Sum of neighbor features

• MeanAggregator(): Mean of neighbor features (with row normalization)

• MaxAggregator(): Element-wise maximum of neighbor features

• MinAggregator(): Element-wise minimum of neighbor features

• ProductAggregator(): Element-wise product of neighbor features

• VarAggregator(): Variance of neighbor features

• StdAggregator(): Standard deviation of neighbor features

Aggregator() is the abstract parent class and cannot be instantiated directly. It is exported so that user-defined aggregators can subclass it and register a forward() method.

See also

layer_sage(), which takes an aggregator.

Examples

MeanAggregator()
#> <torchgnn::MeanAggregator>
#>  @ name     : chr "mean"
#>  @ learnable: logi FALSE

SumAggregator()
#> <torchgnn::SumAggregator>
#>  @ name     : chr "sum"
#>  @ learnable: logi FALSE

# Aggregators are passed to the layers that consume them
S7::prop(MaxAggregator(), "name")
#> [1] "max"