Abstract: For a public transit authority located in Houston, TX, providing accurate bus arrival times plays a critical role in the organization’s quality of service and customer satisfaction. Faced with the task of predicting accurate bus’ arrival times, the Metropolitan Transit Authority in Houston, TX and EastBanc Technologies harnessed the power of AI – in particular long short-term memory (LSTM) artificial recurrent neural network architecture. LSTM based models have been shown to provide highly accurate predictions based on a series of data observations. When using LSTM to predict public transport arrival time it is common to enrich the training data with external information such as surrounding traffic, maps, and city structures.
However, the Houston, TX Metropolitan Transit Authority faced complicated challenges, including limited access to structured data of such nature. The team had to seek alternate data sources and leveraged seemingly less related, but easily accessible data (i.e. weather information, social activity data, city events) to improve predictions.
This session will examine in detail what data (structured and unstructured) was gathered, the LSTM architecture best suited for the solution and how the additional data can be incorporated into LSTM model to improve accurate predictions. The audience will learn how to assess the predictive power of each dataset, how to train a LSTM model, and the approach used for each dataset to train a separate model and apply ensemble algorithm to predict arrival time.
Bio: Polina Reshtova, a Data Scientist at EastBanc Technologies, received her PhD in complex systems data analysis. Over the past 5 years, she has been developing machine learning algorithms and predictive analytical techniques. Today, Polina focuses on implementing iterative approaches (e.g. Minimal Viable Prediction) that break complex, challenging assumptions into small, digestible chunks that can be tested in weeks, not months. These methods minimize the time needed to garner actionable insights while maximizing user feedback loops which progressively increases the accuracy of the developed algorithms.