I plan to use graph convolution for constructing item embeddings in a recommendation system. When generating the dataset, I found that attaching all features of the items to the nodes would result in an excessively large dataset, causing IO bottlenecks and low GPU utilization. Is there a way to generate a dataset only containing node IDs and dynamically complete the features for nodes in the graph during training?
I plan to use graph convolution for constructing item embeddings in a recommendation system. When generating the dataset, I found that attaching all features of the items to the nodes would result in an excessively large dataset, causing IO bottlenecks and low GPU utilization. Is there a way to generate a dataset only containing node IDs and dynamically complete the features for nodes in the graph during training?