For LightGCN, mean pooling (#934) is basically layer 1 of the user embedding stack (where layer 0 is the trained embedding. Therefore, in LightGCN, instead of pooling the final embeddings, we can just compute the user's embedding using an all-0 embedding for layer 0, and compute layers 1+ as usual for LightGCN, and exclude layer 0 from the final average of layers.
For LightGCN, mean pooling (#934) is basically layer 1 of the user embedding stack (where layer 0 is the trained embedding. Therefore, in LightGCN, instead of pooling the final embeddings, we can just compute the user's embedding using an all-0 embedding for layer 0, and compute layers 1+ as usual for LightGCN, and exclude layer 0 from the final average of layers.