#aggregation

Question 1: Single choice

A message-passing layer updates node $i$ as

h_i' = \gamma\Big(h_i,\ \bigoplus_{j \in \mathcal N(i)} \phi\big(h_i, h_j, e_{ji}\big)\Big),

where $\mathcal N(i)$ is the set of neighbours of $i$ and $e_{ji}$ the feature of the edge $j \to i$ . What is $\bigoplus$ , and which property must it have?

The aggregation; it must be permutation invariant and take any number of messages1
The update function; it must be a trainable multi-layer perceptron2
The message function; it must be a linear function of $h_j$ 3
The aggregation; it must concatenate all the messages so that no information is lost4

Timed test