For Uber, the arrival of peak travel days like New Year’s Eve can strain its service to the limit, with the number of rides nearly tripling its average rate. Consequently, it’s critical for Uber to develop accurate forecasting methods to forecast the volume and location of ride demand. To improve prediction precision, Uber has developed a single, flexible neural network that can model all kinds of data from multiple cities simultaneously.
Demand for ridesharing can fluctuate dramatically. For example, Uber estimated that it would have more than 15 million rides on New Year’s Eve of 2016. This compared company average of 5.5 million rides per day during the first half of 2016.
Such a massive change in demand requires Uber to plan ahead in order to allocate resources. However, since Uber has been in operation only eight years, the company has only a small amount of historical data to use to as a basis for future predictions. Furthermore, the difficulty in planning for such events can be compounded by a slew of external variables, such as weather and demographic changes.
Uber noted that a common approach to predicting such events involves the use of classical time-series models, such as those employed by the R software environment. However, Uber said this method doesn’t provide the flexibility or scalability required to meet its needs.
Instead, Uber said it has used thousands of time-series to train a multi-module neural network. The network accounts for all kinds of factors including completed trips, app views, riders, temperature, wind speed and wind bearing. The neural network also employs data from multiple cities where it operates, compensating for the company’s lack of historical data.
Once the neural network’s weights are computed, they can be exported and implemented in a programming language. Uber’s pipeline first trains the network offline using Tensorflow and Keras and then exports the produced weights into native Go code, according to the company.
The company said testing suggests its new approach yields an 18 percent improvement in accuracy in prediction of completed and canceled trips compared to the company’s previous prediction technique that used a proprietary model.
Beyond improved accuracy, the use of neural network prediction has yielded some interesting insights on forecasting rider demand. For example, the forecasting technique suffers its highest error rate on Christmas Day. This corresponds to the greatest error and uncertainty in rider demand.
Uber said it intends to continue working with neural networks and will create a general prediction model for heterogeneous time series, perhaps as an element of larger automated forecasting system.
Tyler Schulze is vice president, strategy & development at Veritone. He serves as general manager for developer partnerships, cognitive engine ecosystem, and media ingestion for the Veritone Platform. Learn more about our platform and join the Veritone developer ecosystem today.