实时空间时间预测 @ Lyft

Real-time spatial temporal forecasting is often used to predict market signals, ranging from a few minutes to a couple of hours, at fine spatial and temporal granularity. For example, we can predict rideshare demand and supply at geohash-6 level for every 5 minute interval in the next hour for an entire city or region. The forecast also runs at a high frequency (e.g. every minute) with real-time input of the most refreshed data to capture the latest marketplace conditions in fine detail, including local spikes and dips.

实时时空预测通常用于预测市场信号，时间范围从几分钟到几个小时，具有细粒度的空间和时间分辨率。例如，我们可以在接下来的一个小时内以geohash-6级别预测整个城市或地区每5分钟的拼车需求和供应。该预测还以高频率（例如每分钟）运行，实时输入最新数据，以捕捉最新市场条件的细节，包括局部的高峰和低谷。

Lyft currently operates in hundreds of North American cities and regions. Our spatial temporal forecasting models predict forecast values for 4M geohashes per minute, per signal.

Lyft目前在北美数百个城市和地区运营。我们的时空预测模型每分钟为每个信号预测400万个geohash的预测值。

Use Cases

用例

Predictions from real-time forecasting are often used for inputs into levers that balance real-time demand and supply across space. For example:

实时预测的结果通常用于平衡空间中实时需求和供应的杠杆输入。例如：

• Dynamic pricing: As a platform, maintaining marketplace balance is one of Lyft’s top missions. Dynamic pricing maintains balance in real-time by raising prices to dampen demand when drivers are scarce and lowering prices to encourage rides when there are more drivers.
• 动态定价：作为一个平台，维护市场平衡是Lyft的首要任务之一。动态定价通过在司机稀缺时提高价格以抑制需求，并在司机较多时降低价格以鼓励拼车，从而实时维持平衡。
• Real-time driver incentives: In contrast to dynamic pricing managing demand, real-time driver incentives achieve marketplace balance by raising drivers’ earnings to attract drivers to come online during undersupply or relocate from oversupplied to undersupplied areas. Similarly, understanding the current and near-future marketplace conditions in different locations is essential to the incentive model.
• 实时司机激励：与动态定价管理需求不同，实时司机激励通过提高司机的收入来吸引司机在线，以在供给不足时平衡市场，或从供给过剩地区转移到供给不足地区。同样，了解不同地点当前和近期的市场状况对激励模型至关重要。

Challenges

挑战

The high dimension, high frequency nature of real-time spatial temporal fo...