1 of 21

Leveraging Statistical Experiment Designs in the Rideshare Industry

Danni Lu

DAE 2024

2 of 21

Danni Lu

  • 2021/03 - now

Sr. Applied Scientist, Uber. San Francisco, CA.

  • 2015/08 - 2021/02,

Ph.D. student, Virginia Tech, Blacksburg, VA

3 of 21

DAE 2024

01 Uber and road safety

02 Experimentations

03 Testing a product

3

4 of 21

DAE 2024

Uber

“ Uber Technologies, Inc., or Uber, is an American multinational transportation company that provides ride-hailing services, courier services, food delivery, and freight transport. ”

From :

4

5 of 21

DAE 2024

Rides and beyond

In addition to helping riders find a way to go from point A to point B, we're helping people order food quickly and affordably, removing barriers to healthcare, creating new freight-booking solutions, and helping companies provide a seamless employee travel experience. And always helping drivers and couriers earn.

Stand for safety

Whether you’re in the back seat or behind the wheel, your safety is essential. We are committed to doing our part, and technology is at the heart of our approach. We partner with safety advocates and develop new technologies and systems to help improve safety and help make it easier for everyone to get around.

5

6 of 21

Road Safety

6

7 of 21

DAE 2024

01 Uber and road safety

02 Experimentations on the platform

03 Testing a product

7

8 of 21

02 Experimentations on the platform

    • A/B Experiments
    • Switchback Experiments
    • Synthetic control experiment and others

8

9 of 21

02 Experimentations on the platform

  • A/B Experiments

While the statistical underpinnings of A/B testing are a century old, building a correct and reliable A/B testing platform and culture at a large scale is still a massive challenge!

Core experiment structure

  • Randomization: the way units are mapped into treatment groups
  • Treatment plan: a mapping from context and unit’s randomization (treatment group) into actions (values for parameters)
  • Logs: an auxiliary construct that records additional information in the experiment

9

10 of 21

02 Experimentations on the platform

  • A/B Experiments

Core experiment structure

  • Randomization:

The experiment units (eg. riders, drivers, eaters) are randomized by hashing their identifiers with a salt determined by the experiment key in a given experiment. Experiment keys are unique, which ensures that all experiments are randomized independently of each other and anything else,

10

11 of 21

02 Experimentations on the platform

  • A/B Experiments

Core experiment structure

  • Treatment plan:

The treatment plan specifies what is done in each context to each treatment group.

    • Context: the knowledge we have about an experimentation unit. It could include geographical information such as the city or country the user is located in, device characteristics such as the operating system (iOS or Android), as well as potentially anything else available at the time of experiment evaluation.
    • Action: the parameter values we return, for the parameters impacted by the experiment

11

12 of 21

02 Experimentations on the platform

  • A/B Experiments

Other experiment structures

  • Parameter Constraints
    • Blocking factors for Traffic splitting, Holdouts, Dependent experiments/feature flags, etc.
  • Generalized Analysis Engine
    • Standard experiment analysis report and ad-hoc analysis support.

12

13 of 21

02 Experimentations on the platform

  • A/B Experiments

13

14 of 21

02 Experimentations on the platform

Sometimes, the experiment we want to test naturally introduce competition/interference between experiment units and network effect is inevitable.

14

15 of 21

02 Experimentations on the platform

  • Switchback Experiments

In cases where network effect is inevitable.

    • Regular Switchback

Switching between control (incumbent algorithm) and treatment (challenger algorithm) that impact all users in the marketplace with a given frequency over a predetermined time period.

    • Quick backs

A switchback with a switch frequency of thirty minutes or less. Used principally for dispatch experimentation to get a quick read on what changes are likely to do.

    • Controlled Lookback

A time-randomized switchback with 80% of time treated and remaining 20% plus prior observations as control.

15

16 of 21

DAE 2024

01 Uber and road safety

02 Experimentations on the platform

03 Testing a product

16

17 of 21

Percent of Intersection Related Traffic Fatalities occurred at unsignalized intersections

67.3

Percent of Traffic Fatalities in the US are intersection related

27.2

The idea

According to the Federal Highway Administration Safety Report:

18 of 21

Headroom analysis

Estimate the impact of the product

Product design

Experiment plan

Experiment type

Country/City selection

Duration

Stakeholders

Metrics

Patent application

Model R&D

Product Launch

Model research, development, and production

How to surface the product to App users

Experiment

XP deployment

Monitor

Data analysis

The idea → product

Improving intersection safety by making app users aware of the upcoming non-all-way-stop intersections.

19 of 21

The product:

Uncontrolled Intersection Alert

Hypothesis

  • Displaying information when approaching an uncontrolled intersection where cross traffic does not stop will reduce safety incidents at those intersections.

Product design: Navigation reminders

  • Alert drivers of upcoming cross traffic in the Navigation

Geography

  • Phase I: Pilot XP - City of San Francisco
  • Phase II: Expansion - All cities in the US

Experiment type: Driver A/B

  • Treatment = Alert in navigation banner
  • Control = No additional Information

20 of 21

Uncontrolled Intersection Alert

Experiment monitor & analysis

  • Improvement in Road safety metrics
  • Improvement in Navigation metrics
  • No degradation in Marketplace metrics

Launch decision

  • Launch the product in the experiment cities.
  • The product is live in all US cities.

21 of 21

DAE 2024

Thank you!

And… we’re hiring! 😄

Opportunities to:

  • identify and solve industry problems and make direct impact on business decision-making, product development, and customer experiences.
  • collaborate cross-functionally with teams from different departments such as marketing, product development, and operations.
  • work with massive timely industry data with flexibilities to research open-ended questions; Access to cutting-edge technologies and tools for data analysis, machine learning, and AI, etc.

See more in career website.

21