Think about the things you or your business own. Are those things highly utilized? When you’re at a stoplight, count the empty seats in the vehicles around you. When you’re in a parking lot, think about the fact that most of the surrounding cars get an hour of use a day. Or consider the duplication of effort when both you and your neighbor make individual grocery trips. Now think of resources your company owns. Is your lab equipment used all the time? What about your warehouse – is it full to capacity 12 months a year? How many empty miles does your fleet of trucks accumulate each month? These example all represent underutilized capacity, and with the right systems, data, and technology, individuals and companies could start accessing these underutilized resources.
Attempts to build these systems are all the rage these days (platform companies such as Uber, Lyft, Flexe all claim to increase underutilized resources). Existing systems to match supply and demand in platforms can be organized broadly into two categories.
The first is a centralized model, which prioritizes meeting demand commitments and enabling a quick time to match. However, centralized approaches do not allow for decision making by the resource owners, so it dampens their participation. Examples of a centralized approach are the default settings in ride sharing. For example, did you know that Uber drivers don’t know where they’re taking you until you get in the car? The driver instead has 15 seconds to accept a recommended request (again without knowing its destination). This means drivers are not provided with enough information to make decisions based on their needs and schedule. This limits who can participate as a driver and utilizing underutilized capacity is not prioritized. If you’re interested in learning more about drivers’ experiences in ride sharing applications – I highly recommend a recent book on the topic: Uberland by Alex Rosenblat
The other approach is a “jobs board” system, where all requests are displayed to all suppliers. This can lead to myopic decision making: no one supplier has the whole picture of the marketplace, which results in reduced system performance because some requests receive multiple selections and others are left unfulfilled. Also, this is a passive system – in which individuals must take the initiative (and invest the time) to find preferable demand requests. Recently, a number of interesting empirical work captures reduced systematic performance in such systems, see Cullen and Farronato (2018); Fradkin (2017) and Horton (2014).
Our research is interested in a new way to match underutilized capacity owned by freelance suppliers to demand requests that combines the advantages of the two approaches. This is a hard problem, because we have two different sets of stake-holders tugging us in different directions. On the one hand, we want to entice the owners of these resources to provide access, this means creating pro-supplier policies, taking into account individual preferences (for example, that these freelance suppliers may have planned tasks that we want to tag along for). However, on the other, we need to ensure high levels of service for the demand-side.
Our innovation is an approach enabling “Supplier Choice”. The platform – without control nor perfect knowledge of suppliers’ preferences – uses choices estimated from suppliers’ past behavior to increase participation and resource utilization. To illustrate, meet Millennial Millie. Millennial Millie has gone through training and signed up to be an on-demand volunteer. She gets a notification on her phone asking if she’d like to deliver groceries to shut-in residents. Millie clicks yes. Two requests appear. She chooses the one that fits with her plans that day.
Before this example can become a reality, research is needed to discover new ways to provide choices and quantify the impact of those choices on suppliers and demand requests. Our research, led by Shahab Mofidi, recently published in Transportation Research Part B: Methodological is an important starting point for platforms that must coordinate demand requests with decentralized owned resources. For all the details, please check out our paper:
Mofidi, Shahab, and Pazour, Jennifer A., “When is it Beneficial to Provide Freelance Suppliers with Choice? A Hierarchical Approach for Peer-to-Peer Logistics Platforms,” to appear in Transportation Research Part B: Methodological. [PrePrint] https://doi.org/10.1016/j.trb.2019.05.008
This link provides free access until July 23, 2019: https://authors.elsevier.com/a/1Z9wHhVEAswhb
In what follows, I walk you through why I am excited about this research, and why I believe it can start to improve crowdsourced delivery and rider sharing, but also impact volunteer management and other applications.
As displayed in the above figure, our research recasts the platform’s role as one providing personalized recommendations (i.e., a menu of requests) to a set of suppliers. This means that to match supply and demand, two decisions processes take place.
- First, the platform must decide how multiple, simultaneous recommendations are made. The same request may be recommended to multiple suppliers to decrease the time to match and to hedge against suppliers’ autonomy to decline recommendations.
- Second, suppliers have autonomy to select requests (if any) from the personalized menus, resulting in some requests not selected and others selected by more than one supplier.
This hierarchical approach enables a quick time to match, and does not require suppliers to explicitly provide preference information for all requests. Suppliers retain autonomy to select requests that can be interleaved with their planned activities. And as the number of choices increases, suppliers have a higher chance to be recommended a request they are willing to select (increasing participation). However, due to misalignment between the suppliers and the platform’s preferences, a larger number of choices may lead to suppliers selecting a request with lower platform benefit. Also, as the number of choices increases, less systematic coordination occurs, and the chance for rejected requests increases. Therefore, we are interested in understanding when is it beneficial for a platform to use personalized recommendations to a set of freelance suppliers to coordinate demand requests? We also want to quantify this benefit to the platform, the suppliers, and the demand requests under different environmental factors.
To be able to answer these questions, we develop a new bi-level optimization framework to explicitly model both the platform’s decisions and the suppliers’ selection decisions. What’s challenging about the model is that the platform’s objective function is influenced by both the platform’s recommendation, but also the suppliers’ interdependent selection outcomes. Our model assumes the platform uses expected values of agents’ utilities to make recommendations. We have a single period model and assume the model is re-solved iteratively with a given set of available requests and available suppliers. Because bi-level optimization problems are computationally hard to solve, we exploit the problem’s structure to reformulate as an equivalent single-level optimization problem. For mathematical and computational details, please see our journal article.
Choice is only useful if the platform does not fully understand what the freelance suppliers want to do. This is typically the case – because eliciting utility values of all requests is time consuming and tedious for suppliers. Instead, in practice a platform is able to only partially estimate suppliers’ selection outcomes. Therefore, we develop a computational study – based on ride-sharing – to investigate whether personalized recommendation sets can be used as a coordination mechanism for distributed resources when the platform has neither perfect knowledge nor control over suppliers. The computational experiments have four phases (see Figure below), which generates 1140 instances and solves 10,260 optimization problems (see journal paper for details).
The goals of our computational experiments are to:
(1) compare the performance of our proposed hierarchical model with existing platform matching methods, namely, centralized, decentralized, and many-to-many stable matching mechanisms.
(2) quantify the platform’s value of providing choices in on-demand environments;
(3) determine under what scenarios it is better to recommend only one alternative, a few alternatives, or all alternatives, and to investigate when it is useful for the recommendations to contain overlapping alternatives.
First, we find that providing choices and recommending alternatives to more than one supplier simultaneously provide value, on average, to the platform, suppliers, and demand requests. When the platform is only partially able to estimate suppliers’ utilities, the hierarchical approach creates recommendation sets with higher average platform performance than centralized, decentralized, and many-to-many stable matching approaches. A centralized approach does not perform well due to a higher chance of suppliers electing not to participate in the platform’s assignment. However, too much choice can be problematic even for the suppliers themselves. As excess discretion creates collisions, the decentralized approach had the worst angry driver rate and lowest serviced ride rate. Thus, in uncertain environments, the proposed hierarchical approach can benefit the platform, drivers, and ride requests.
Second, a platform with uncertainty about suppliers’ selection outcomes can sometimes improve its performance by offering choices to suppliers, but not always. In platforms where uncertainty over suppliers’ selection exists, choices help the platform when (1) suppliers are inflexible with a high no-choice utility; and (2) suppliers’ utility values have higher variance than the platform’s utility values. Choices increase the chance of enticing supplier participation, and in both cases, the increase in participation likelihood outweighs the match’s reduced platform benefit likelihood. However, when (1) the platform’s benefit values have higher variance than the suppliers’ expected utilities or (2) suppliers are flexible, offering additional choices, on average, can reduce the platform’s benefit.
We view the current paper as a “proof of concept” that providing freelance suppliers with choices can sometimes be beneficial. However, this is really just a jumping off point, and much more research is needed. Thankfully I have smart students and funding support to keep researching. Specifically, Hannah Horner (a math PhD student), John Mitchell (Professor in Mathematics at RPI) and I are explicitly capturing the stochastic nature of suppliers’ utilities directly in the optimization formulation, which allows us to extend our analysis beyond answering if choice is beneficial, and to determine optimal menu characteristics (i.e., menu sizes and request overlaps). Rosemonde Ausseil (PhD student in ISE) and I are working to extend the optimization model from a single-period model to one that considers the dynamic properties of the problem directly in the optimization formulation. Safron Smith (MS student at RPI) is interested in application specific formulations for volunteer management. Talented undergrads are also contributing to this work, including Tina Nazario, Brooke Ramlakhan, Wenlin Gong, Parker Shawver, and Karthik Dusi.
We greatly appreciate funding of this work from the National Science Foundation Award# 1751801: CAREER: Distribution Resource Elasticity: A New Hierarchical Approach for On-Demand Distribution Platforms. I would like to acknowledge Dr. Shahab Mofidi – who is the lead author on this work. This was part of his dissertation, which was awarded the 2019 Del and Ruth Karger Dissertation Prize. Also, thank you to John Backman, who wrote up some of this work for the RPI ISE department newsletter
We would appreciate feedback on our work, send questions and inquiries to firstname.lastname@example.org