通过深度概率模型增强Uber的指导热图

At Uber, giving high quality guidance to drivers is crucial for smoothing the learning curve for new drivers and improving driver retention. Our internal research shows that a major pain point drivers face is that it often takes weeks of trial and error for drivers to figure out the nuances of their particular market, resulting in churn and frustration.  

在Uber，为司机提供高质量的指导对于平滑新司机的学习曲线和提高司机留存率至关重要。我们的内部研究表明，司机面临的一个主要痛点是，司机通常需要数周的反复试验才能弄清楚他们特定市场的细微差别，导致流失和挫败感。  

To help address this issue, our AI team has developed probabilistic prediction models that power guidance tools like the Heatmap (Figure 1), giving drivers information for making decisions about when and where to drive. These insights can help significantly improve the driver experience and enhance overall platform efficiency by highlighting areas with greater demand and opportunities.

为了解决这个问题，我们的人工智能团队开发了概率预测模型，这些模型为像热图（图1）这样的指导工具提供支持，帮助司机获取关于何时何地驾驶的信息。这些洞察可以显著改善司机体验，并通过突出需求和机会更大的区域来提高整体平台效率。

Figure 1: In-app screenshot of the earnings heatmap.

图1：收益热图的应用内截图。

The probabilistic models which power the Heatmap use a Deep Neural-Network architecture which outputs a distribution of forecasted earnings outcomes. Our work in this area allows us to capture real-world variability in demand and provide useful insights for drivers.

驱动热图的概率模型使用深度神经网络架构，输出预测收益结果的分布。我们在这一领域的工作使我们能够捕捉到现实世界中需求的变动，并为司机提供有用的见解。

In this blog, we dive into the technical challenges we faced, such as handling noisy data, creating well-fitting probabilistic models, and deploying ML models at scale in a real-time production system that serves millions of users.

在这篇博客中，我们深入探讨了面临的技术挑战，例如处理噪声数据、创建良好拟合的概率模型，以及在实时生产系统中大规模部署机器学习模型，以服务数百万用户。

Our probabilistic forecasting models power the Heatmap, which provides granular, location-based information for drivers. The heatmap updates every 10 minutes, ensuring drivers receive up-to-date estimates highlighting areas forecasted to have above-average earnings opportunities. To do this, we designed our model to predict earnings per hour (EpH), since this allowe...