Deep Neural Shipping Network

An article about the global shipping network on Motherboard caught my attention. Doing a first order analysis of shipping behaviors, where ports are nodes, and the shipping between them the links (number of ships weight of links) shows that first order dependencies are poor predictors of global shipping.

They found this out when they examined the spread of zebra mussels. Zebra mussels are spread from port to port in the ballast of cargo ships. The spread of zebra mussels does not correspond to the behavioral model of shipping.

Current behavior models only look at first order dependencies. For example, if more ships go from  Tokyo to San Francisco, the model says a ship usually ends up in San Francisco instead of Los Angeles. That's not the case.

Higher order dependencies, such as where the ship came from before Tokyo, give better predictors of where it goes. Thus a ship going from Shanghai to Tokyo may be more likely to end up in Los Angeles (more properly the terminus of Long Beach).

A slightly more detailed report can be found here. Now, what I found interesting was the diagram they used.

See, what that looks like to me is not a static network, but a dynamic neural network. One of those, what they do call it? Encoder-decoders? You know, where they do a nested or sandwiched net. The input layer of a net feeds an input layer of another net sandwiched in, and the output layer of the inner sandwiched layer feeds the output layer of the outer net.

Next logical step in logistics, I suppose.