Ecommerce ops · Production

Lyft's ML-powered ride mode recommendation and ranking system

The problem

As Lyft's product portfolio grew, riders in major markets faced 8–10 mode options at once, making it hard to discover the right ride type and risking accidental wrong-mode bookings. New mode launches suffered from cold-start bias due to limited training data, limiting their early visibility.

First attempt

Earlier preselection heuristics were flawed: selecting the last-taken mode introduced exploration bias, and selecting the most-frequently-used mode was unstable for infrequent users. Static ranking by price or ETA did not account for user preferences.

Workflow diagram · grounded in source
1
Rider opens app or sets destination
trigger
“The recommendation system covers user experiences throughout the ride journey”
2
ML model predicts mode propensity
ai_action
“the recommendation system leverages a set of machine learning models to predict a rider's propensity of converting into each mode and customizes the rankings based on it. Rich information has been considered in building these models, inc…”
3
Post-processor adjusts cold-start rankings
ai_action
“a post-processor was introduced as an additional layer to adjust the machine learning model results, mitigating the natural bias created by the lack of sufficient training data”
4
Ranked mode list displayed
output
“the app presents a ranked list of product offerings based on the user's travel preferences and the current marketplace conditions. Some visual highlights are also displayed to help clarify the tradeoffs across different options. For inst…”
5
Highest-propensity mode preselected
ai_action
“preselecting the mode with the highest predicted propensity score. This solution is by design real-time, dynamic and has proven to be more accurate and effective”
6
Post-request cross-sell shown
routing
“In select sessions, these changes can be captured and made beneficial to riders in our post request cross-sell experience, where an interstitial prompt is introduced detailing upgrade options with a better ETA or price”
Reported outcome

ML propensity models now serve about 90% of use cases and have proven to drive more desirable outcomes than the static system, with model-based preselection proven to be more accurate and effective.

Reported metrics
ML model coverage of use casesabout 90%
Outcomes vs static systemmore desirable outcomes compared to a static system
Model-based preselection accuracymore accurate and effective
Reported stack
LightGBM
Source
https://eng.lyft.com/the-recommendation-system-at-lyft-67bc9dcc1793
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

ML propensity models now serve about 90% of use cases and have proven to drive more desirable outcomes than the static system, with model-based preselection proven to be more accurate and effective.

What tools did this team use?

LightGBM.

What results were reported?

ML model coverage of use cases: about 90%; Outcomes vs static system: more desirable outcomes compared to a static system; Model-based preselection accuracy: more accurate and effective (source-reported, not independently verified).

What failed first in this deployment?

Earlier preselection heuristics were flawed: selecting the last-taken mode introduced exploration bias, and selecting the most-frequently-used mode was unstable for infrequent users.

How is this ecommerce ops AI workflow structured?

Rider opens app or sets destination → ML model predicts mode propensity → Post-processor adjusts cold-start rankings → Ranked mode list displayed → Highest-propensity mode preselected → Post-request cross-sell shown.

Related ecommerce ops cases