Inside Uber: Leveraging AI and ML for Enhanced Rider Personalization

In the dynamic realm of on-demand services, Uber is working on convenience of riders and personalization. In an interview Madan Thangavelu, Senior Director of Engineering at Uber, we gain insights into how AI and ML are revolutionizing personalized experiences within the on-demand service industry.

He discusses Uber's responsible AI practices, market predictions, and the competitive landscape, offering a comprehensive view of the company's position and future direction within the evolving on-demand service sector. Excerpts:

How does Uber leverage AI and ML to enhance personalized experiences for riders and drivers?

Our goal is to be as assistive as possible to enable every rider to have a seamless experience each time they book an Uber. We deploy machine learning in key aspects of the booking process such as product selection, where the rider chooses their favoured Uber product (Uber Go, Uber XL, Uber Auto etc). 

The ML system determines the right list of products to show a user depending on their usage history, current marketplace conditions, and their individual preferences. If a rider has been a regular user of the Uber Auto product, the ranking system would continue to surface the product for them in different parts of the app.

For an avid Uber Auto user, products such as Uber Moto and Uber Go may be viable alternatives. In such cases, the system would also recommend those two alternatives on various screens of the app to educate the rider about the variety of offerings on the Uber app.

Can you explain what rider personalization means for Uber and why it's crucial in the on-demand service industry?

The on-demand service industry operates largely based on mobile apps. These mobile apps have limited UI (user interface) surfaces constricted by screen sizes of phones, unlike websites on desktops which have a way bigger surface to interact with their users. As the offerings of these on-demand services grow, the UX (user experience) finds itself limited in terms of being able to offer all available services to its users in an easy-to-find way. Thus, personalizing the UX is highly important to ensure that the app remains relevant to each individual user while also minimizing the effort it takes to navigate within the app.

Depending on the rider's preferences, it is critical to assist each of them with the best product for their needs. For example, if a rider is using Uber during office hours, and they are very sensitive to the time taken for them to get a car, it is important to inform them of booking options that can get their ride faster even as prices may be higher. On the other hand, if the same rider is looking at a casual non-work trip on a Sunday evening, it may be acceptable for them to wait and pick the cheapest option. Surfacing the right ride options depending on rider context greatly enhances the user experience when interacting with the Uber app. Personalization plays a direct role in this.

How does Uber incorporate AI and ML solutions to optimize operational efficiency and streamline its operations?

The usage of AI and ML is across many systems and processes at Uber. Some of them are:

○ Matching - Finding the most optimal driver to assign for a trip so that it maximizes driver earning potential, minimizes kilometres without a passenger in the car and finally, provides the most cost efficient option for a rider. 

○ Batching - Providing the ability to batch multiple trips in package delivery directly improves operational efficiency, while also helping drivers earn more in less time. 

○ Earner Intel - Providing appropriate information to our earners so that they can navigate to locations in the city with expected increased demands is critical for strong operational efficiency. 

○ Data utilization - Utilizing the data from our global operations to train ML models that optimize various aspects of operations helps us improve efficiency and streamline operations. 

Is there a method in place to measure the impact of AI-driven personalization on customer engagement and satisfaction?

There are direct and indirect measurement techniques to observe the benefits of this system. Direct measures like click-through rates on various personalized components provide feedback on how assistive and useful is personalization. Long-term user interaction with Uber such as the number of trips over a few months can be another way to measure the impact.

We have an array of product offerings. A measure of the rider's awareness about the entire range, and the increase or decrease in their utilization of multiple products provides a measure of our personalization tech.

Is the personalization approach region-specific? How does this strategy benefit both Uber and its users across different geographical areas?

The ML algorithm needs to consider the various geographic preferences and differences. This is done by training the ML models with the datasets that incorporate features having regional attributes. Based on the differences in regional behaviors of riders, the ML model automatically develops an ability to weigh each awareness feature differently when applicable.

This benefits users as regional customizations are important. For example, in Asia-Pacific, a rider may prefer a low-cost product such as high-capacity vehicles, while in the US the preferred low-cost option may be to share a ride with UberXShare. Incorporating the different cultural and regional preferences is critical to recommending the right product to every rider.

Can you share insights into Uber's data curation and model development processes, particularly how they ensure responsible AI and ML use?

Our data curation for ML development depends on the strong foundation of privacy and handling of rider data. Appropriate guidelines are adopted to ensure that the right level of PII (personally identifiable information) scrubbing and anonymization is adopted across stages of the development lifecycle.

The model development process starts with careful selection of features in a central storage system offered by Uber's in-house ML platform called Michelangelo. Algorithms are developed and validated against a small segment of this data. Once a model is determined to be viable, it is gradually rolled out to our rider base while carefully monitoring the accuracy and correctness of the developed system.

What is your vision for the future of on-demand services powered by AI and ML, and how does Uber plan to contribute to this evolution?

On-demand services by nature are real-time and transient. However, AI and ML provide the ability to predict many possible scenarios of an on-demand marketplace ahead of time. As the data to make these predictions accumulates, and the predictions become more robust, it will allow for on-demand companies to provide those services much cheaper and more reliably due to improving the efficiency of the operations and planning ahead of time.  As an example, the amount of accuracy we can have while predicting the demand curve in a certain part of the city, is proportional to the confidence to provide a more reliable service by ensuring that enough drivers are available in those parts of the city at any given time of the day.

Over the years, our predictive capability has become extremely robust allowing us to offer reliable services at the right cost efficiency to our customers across the world.

How does Uber utilize AI and ML to implement dynamic pricing models and predictive analytics, and how does this benefit both riders and drivers?

The primary goal of pricing modeling is to ensure that our riders can experience a dependable marketplace where they can get a ride anytime they need. The goal of the pricing is to ensure balancing demand and supply such that 100% of available supply does not get utilized. This ensures availability of supply in a marketplace for the riders who need access to that reliability. When the available supply is plenty, the pricing systems price a product cheaper while they go up if there is limited availability.

How does Uber balance the need for personalized experiences with user preferences while ensuring flexibility and choice for its diverse user base?

Recommendation or personalized systems have an inherent bias towards continuing to recommend experiences that a user prefers. However, advanced methods are adopted to continuously explore other preferences that a user may be interested in. The app's information architecture also provides our riders the ability to look at all options they have beyond the recommended choices. This allows them to never be stuck due to personalization and have a way to access any product of their choice.

Having a good information system means that the rider knows how to get to all options and services that we offer. This allows personalization to exist and tailor the Uber app to each rider’s preferences, while not coming in the way of them being able to optimally utilise the app. When the personalization is accurate, the experience is more delightful for every rider.