Lyft and urban mobility

Two pictures of how we move across the metropolis

From: Alex Chin, Michael Jancsy, Shilpa Subrahmanyam, and Mark Huberty

Lyft moves people through space and time. But where those people move, and why, is up to them. Lyft’s riders use our services to get to and from work, go out to dinner, visit family, and get to the airport. When and where they do so tells us a lot about urban mobility — whether and how the notions of neighborhood, geography, and landscape shape how people move through space.

Here we share what we can learn from long-run patterns on Lyft’s operations in major US cities. We see that cities vary a lot internally in how people travel, where, and when. That diversity implies a need for a diverse range of products and services. But, strikingly, we also see how cities resemble each other — that sometimes, common patterns, like urban downtowns, look more like other cities’ downtowns than they do their companion suburbs.

This paints a rich picture of how we live in time and space, and helps to shape a complex understanding of the transportation services that modern metropolitan areas need to function socially, economically, and culturally.

Two visions of mobility: where you start, and where you want to go

Lyft rides paint many different pictures of urban mobility depending on what part of the trip we emphasize.

If we choose only pickup information — both where a trip starts, and when — we get a picture of the demand for mobility. We’ll notice, for instance, that demand in an office district clusters heavily towards the end of the day; while demand in a residential area may vary based on weekdays (mornings, for the commute) versus weekends (evenings, for going out).

But if we link pickup and dropoff information, we start to uncover patterns of connection between different parts of the city. For instance, Manhattanites stereotypically loathe leaving the island: but is that true for Lyft riders? Does Lyft mobility show us that we do a lot of rides that start and end on the island, but rarely venture to the uncharted regions of Brooklyn or Queens?

Here we showcase two different ways of thinking about how these patterns emerge from rideshare demand, and how those patterns in turn reflect the way we live and move in American cities.

Where you start: pickup locations and a picture of fluctuating demand

We begin by examining the geographic distributions of rideshare demand. By fitting a mixture model to the locations of rider sessions, we can obtain probabilistic clusters that highlight patterns of demand.

In many areas, the demand clusters faithfully represent actual neighborhoods, such as the Mission, North Beach, and downtown Oakland. The SFO and OAK airports are also clearly visible, representing tight pockets of passengers requesting airport pickups.

More notably, the clusters follow not just neighborhoods, but arterials as well. Cluster boundaries hew close to arterials — suggesting that major urban roads tend to break up underlying patterns of economic and social activity. That theory would be consistent with work noting that transportation networks — particularly postwar highways — split up and divided historic neighborhoods in permanent ways. For instance, the construction of the Civic Center BART station in downtown San Francisco divided the mid-Market neighborhood in critical ways that remain to this day, 40 years on.

Localized clusters in the San Francisco Bay Area, overlaid with official neighborhood boundaries. We can see that neighborhoods appear readily: the outer Sunset (pink, far left) separates from the Richmond (grey); the Mission stands out in pink and purple at the center.

Looking eastward, we can see how demand has shifted over time. Lyft’s airport business often stands out from other business patterns — airports are very unique places. But in April 2020, as the world went into lockdown, we saw that New York City’s two major airports faded into the surrounding neighborhoods. But by April 2020, as vaccines started to make air travel palatable again, we see the airport areas re-emerge as unique clusters of their own, reflecting the unique travel patterns created by airport arrivals and departures.

Demand in portions of Queens and Brooklyn, in April 2020 (left) and April 2021 (right). In addition to the difference in intensity indicating the drastic difference in overall demand, the clusters highlight changes in demand patterns. In April 2021, JFK and LaGuardia airports are prominent clusters, representing the recovery of airplane traffic; in April 2020, JFK represents a faint cluster whereas LaGuardia is not even visible on the map.

Where you go: a picture of interconnectedness

Clustering demand for mobility over time gives us a picture of how demand varies in time and space. But it doesn’t intrinsically tell us about what trips those individuals demand. To better understand how people move about an area, we want a clustering of connectedness, of where people start and where they go.

Conventional transportation networks, like buses or trains, provide easily-understood connection patterns between two points. A bus line or a train track clearly demonstrates how transportation connects two parts of a city. Rideshare, however, is different: with no fixed lines, those connections emerge from the structure of demand.

We can infer that structure using ideas from graph theory. We construct a graph — a set of nodes, linked by edges — from the geohashes that represent the pickup and dropoff points of a ride. Each ride that connects a pair of geohashes — either way — adds 1 to the weight we place on that edge connecting the two points in space. Using this graph, we can then run community detection to discover how different parts of a city are linked together. To continue the social media analogy, we want to know if 2 different parts of the city are connected to each other more intensely than they are to other parts of the city.

The picture below starts to tell a story about mobility in different urban areas. In the Bay Area, San Francisco separates out from the Peninsula. Oakland separates from Berkeley. On the other side of the country, we see that, apparently, Manhattanites really don’t leave Manhattan. But, to be fair, Brooklynites aren’t much better — at least on Lyft. Perhaps New Yorkers save their subway rides for trips into other boroughs. Finally, Chicago splits north and south — reflecting longstanding divisions in the city. But we also see that connectivity tends to stream in and out of the downtown area, like spokes emanating from a hub. This reflects not so much the geography of the city, but the very structured way that zoning and historical development separate out work, entertainment, and living.

Finally, we note that airports don’t show up here — and for good reason. Our discussion so far has centered on the idea that there’s geographic structure in the connections between pickup and dropoff. But hubs like airports serve a wide range of areas from a single place. An early version of these clusters, for instance, showed LaGuardia Airport linked to Manhattan but not its surrounding neighborhoods. While that’s of interest for us as geographers, Lyft’s broader interest — in how we build products and services — usually treats airport trips as distinct from the rest of our business.

Spatial clustering in the greater San Francisco Bay Area. Each color reflects one ‘community’ uncovered from weeks of Lyft trips. We see that San Francisco is a community unto itself. The water barriers naturally dissuade people from crossing between San Francisco and Oakland (to the east) or Marin (to the north).

Communities of travel in the greater New York area. Notice that Manhattan below Central Park separates from Harlem and the area north of the East River. Brooklyn separates from Queens. But a tiny bit of western Manhattan is more closely tied to eastern New Jersey than Manhattan itself.

Connected communities in Chicago. Chicago is infamously segregated north to south. Here we see that reproduced in mobility patterns: despite the continuity of Lake Shore Drive up and down the lakefront, mobility splits north-to-south around downtown. We also see how the trunk road network structures work and leisure — communities spread outward along the major trunk roads from downtown and are often bounded by them.

Compared with the pickup-based clusters, these clusters provide more a picture of large areas. They reproduce, to a striking degree, our intuitive labels for living areas. The boroughs in New York stand out clearly. The city of San Francisco separates cleanly from the Peninsula, almost exactly at the official city boundary. All of this happened despite the algorithm having no actual spatial information in it — geohashes are just labels. Instead, human mobility connects areas within these “official” areas more than it connects across them. In a sense, just as Roman transport networks continue to influence patterns of European economic activity, so too does a flexible, dynamic rideshare service reproduce existing political and social boundaries.

Where do we go from here

Lyft has a novel perch from which to observe patterns of urban mobility. Here we’ve highlighted just a few. Our motivations, though, stem not just from curiosity — but also better, more efficient, more effective business operations.

We’ve shown that we can demonstrate what seems an obvious truth: that cities aren’t homogeneous in their needs for transportation. We can use that for all kinds of improved operations — for instance, knowing that some neighborhoods demand more rush hour transportation than others can influence how we incentivize drivers to ensure our riders can get rides.

Can we extend that curiosity from looking within cities to between them? Chicago and New York, will, of course, tell you that they are nothing like each other. But how true is that? Having decomposed our cities into more coherent elements, we can now start thinking about whether those elements more closely resemble the city they are from — or equivalent places elsewhere.

Again, we find things that are compelling but strangely obvious. The heatmap below illustrates the similarity, for Lyft’s business, of clusters labeled for downtown versus suburban Los Angeles and Dallas. These two cities have very different climates, economic models, state regulatory frameworks, and cultures. Yet despite these forces, their downtown urban areas tend to resemble each other in their rideshare needs, more than they resemble their respective suburbs. Conversely, the suburbs resemble each other more than their respective downtowns.

Within- and between-city similarity for different neighborhood types. We see that downtown LA and Dallas resemble each other (on mobility dimensions observed in Lyft’s operational data) more than either downtown resembles its companion suburbs.

Beyond proving that we can recover what’s generally known, this kind of cross-city similarity is hugely valuable to logistics providers like Lyft. It means we can develop strategies to serve different kinds of customers in space and time, and share these strategies across cities. Automated discovery and then exploitation of variation within and between cities improves overall service levels for riders and drivers alike.

As always, Lyft is hiring! If you’re interested in urban mobility, check out our careers page.