Network Isolation Mask restores accuracy in batch road-network inference

This experiment builds on a series of posts that first built a traffic flow dataset from OpenStreetMap and UXsim, then trained a Graph Embedding and Transformer model, and later tested transfer learning to a road network with a different number of nodes. In the fifth installment, a Shared Transformer was co-trained on data from the 浅間下 intersection area (122 nodes) and the Kawasaki Municipal Sachi High School area (126 nodes). Inference in that earlier work was run separately for each network, calling the same Shared Transformer twice.

The new step borrows an idea from EdgeLoRA's Batch LoRA Inference, where multiple LoRA adapters share a single base model to process requests together. Here, the two networks' token sequences—1464 and 1512 tokens—are concatenated into a 2976-token input. But naive concatenation lets tokens from one network attend to the other. The Network Isolation Mask blocks those cross-network links, and after separation the outputs go through each network's own Fully layer. With the mask, batch inference matched individual inference almost exactly; without it, error metrics rose noticeably.

The author positions this as evidence that a single Shared Transformer can handle multiple road networks simultaneously, potentially scaling by adding Network Adapters rather than duplicating the Transformer for each network.