Abstract:
The communication technology revolution in this era has increased the use of smartphones in the world of transportation. The public transit system wishes to predict the expected ridership at a specific time. The bus system's dream of dynamic routing (i.e., having no more near-empty buses) is close to reality. Our study collected data from the transit bus system in two different countries (USA and Japan). This paper studies the correlation between ridership and weather conditions using a deep learning algorithm. Empirically, an experiment with several bus routes in Japan, was utilized to confirm a high precision level. We compared our deep learning model against three baseline models: linear regression, regression trees, and a wide neural network in two sub-experiments. First, the deep learning model had 23%, 21%, and 14% better mean squared error (MSE) scores compared to linear regression, regression trees, and wide neural networks when trained on three months of historical data. Second, we ran the models on a larger dataset of two years. The result was that the gap between the baseline models and the deep learning model was almost doubled to 56%, 50%, and 35% better MSE scores for predictions in favor of the deep learning model compared to linear regression, regression trees, and wide neural networks, respectively.
