Training a model on multiple road segments of traffic data to predict another road segment

Hello,

I have a massive traffic dataset that contains noisy measurements of traffic in 15 minute bins on a few hundred road segments. My goal is to train a model on 80% of the segments and then predict the traffic on the other segments (their full timeseries). For this, it is fair to assume that the segments are very similar.

Although this is a timeseries forecasting question, it is different than normal forecasting models (like LSTM) because I am not using ground truth historical data to predict the future. Rather, I am using the sequences of other road segments to predict a new road segment.

My idea was to train a DNN including time of day, day of week, etc. as prediction columns to help incorporate the seasonality (daily, weekly, monthly) in the traffic predictions (along with my other noisy measurements). However, I wasn’t sure if this was the best option. Anyone have any ideas/suggestions? Would really appreciate any insight!

Hi @Alexander_Augustine
In my understanding for this problem, for learning spatial relationships between road segments and modeling the seasonality of traffic data (daily, weekly, monthly).
Using a DNN with time features, a CNN or GNN for spatial relationships, or a combination of both should help the model generalize well across different road segments.
Transfer learning can also be used for better generalization from known road segments to new ones.
Thank you