How Lyft Uses Machine Learning to Optimize Rides in 2024

Image Credit: Lyft.com

 

When a user opens their phone and types a few words, they can quickly secure a Lyft ride to their destination. This process is designed to feel effortless, as if a friend is picking them up. Jason Laska, a principal software engineer on Lyft’s mapping team, emphasizes that the goal is for the system to be so seamless that users don’t even notice it working.

This seemingly simple process is driven by advanced machine learning algorithms. Here’s an inside look at how Lyft utilizes machine learning to generate faster, more accurate routes and more convenient pickup spots, and even predict user destinations.

Getting Accurate ETAs

Accurately predicting estimated times of arrival (ETAs) is crucial for managing Lyft’s complex network. The system must know how long it will take a driver to reach a pickup point to efficiently pair drivers with riders. Minimizing this time across all rides helps optimize the entire system.

Machine learning plays a key role here. By learning from aggregated data of past rides, it determines precisely how long a journey will take, considering factors such as traffic conditions at different times. For example, a ride through a traffic-heavy urban area will have a different ETA on a Thursday afternoon compared to a Sunday morning. Real-time data is also factored in. If a ride is moving slower than usual, the model understands there’s likely heavy traffic and adjusts the ETA accordingly. Laska explains, “Seasonality, changes to the road, changing customer behaviors, all of those things impact predictions.”

Not all data is treated equally. For instance, ETA models had to disregard much historical data to adapt to new traffic patterns during the pandemic and will continue to adjust as more companies require employees to return to the office. Choosing which data to consider is a key task for Lyft’s technical team.

In 2019, Lyft engineers developed their own in-app maps, generating more rideshare-specific data points. For example, drivers tend to slow down when picking up and dropping off passengers, and Lyft can account for this when predicting average speed.

The effort has paid off. Serdar Colak, Lyft’s mapping data science manager, notes that ETAs are highly accurate, improving continuously over the years. In San Francisco, Lyft’s average pickup ETA is accurate to within a minute, and the average dropoff ETA within two minutes.

Optimizing Routes

Lyft’s algorithms not only estimate travel times but also find the fastest routes based on current conditions. By ingesting real-time data from many drivers, Lyft gets an up-to-the-minute snapshot of heavily used roads. This allows for quick adaptation of routes. For instance, if a road is blocked, Lyft will detect the change and reroute accordingly. When the road reopens, the model will return to its original route recommendations.

Laska mentions that this process also applies to complicated pickup areas like apartment complexes, where rules about car movement can cause confusion. Lyft detects unusual patterns and prompts drivers to provide feedback, improving future routes for new drivers.

Selecting the Right Destination

When users start typing a destination into the Lyft search box, a list of suggestions appears. These recommendations are generated by machine learning algorithms based on customers’ locations and past behaviors. On a Friday evening, the app might suggest a bar frequently visited on weekends, while on a Tuesday it might prompt a favorite weeknight taco spot. According to Lyft engineer Murat Aydos, this technology can predict destinations with 60 to 70 percent accuracy.

For new destinations, such as a newly opened restaurant, users may need to look up and input the address manually. Lyft’s algorithms then alert employees to add this new destination to the database for future convenience.

Coordinating Rider Pickup

Complex pickup locations, like malls or hospitals with multiple exits, require more guidance. For instance, after a concert at a large venue, coordinating the pickup spot can be challenging. To address this, Lyft’s app suggests the most common pickup point, based on historical data and the most intuitive spots for both rider and driver.

The app's algorithms use historical behavior to avoid inconvenient or unsafe pickup spots. If certain areas are consistently avoided, the model learns to suggest better locations.

In the future, Lyft aims to improve not only where riders and drivers meet but also when they meet. At airports like LAX and JFK, a new preorder experience allows passengers to book rides as soon as they deplane. An algorithm estimates walking times to the pickup area, making the process smoother as the models continue to adapt.

Sometimes, however, the best solution is not machine learning but direct communication. For instance, drop-off locations are often managed by the rider in the car, who can guide the driver to the precise spot. This person-to-person dialogue can sometimes be more effective than relying solely on AI.