GigFitter: Tailored Scheduling Recommendations for RidesharingWorkers

Cameron Cabo, Jeffrey Chen, Minjae Tyler Kwon, Hao Chuan (Jordan) LeiDr. Gad Allon, Dr. Sangeeta Vohra

Abstract

Recent years have seen a rise in the preva-lence of freelance work, facilitated by largecorporate platforms who mediate contractor-consumer interactions. The distributed na-ture and scale of this so-called “gig economy”makes it difficult for gig economy workers tomobilize, negotiate, or adequately plan theirlivelihoods. The GigFitter Team has designedan application to enable gig economy workersto lead more predictable work schedules whilemaximizing pay and better fitting their workschedules to their preferences.

1 Introduction

1.1 Motivation

The workforce in the gig economy has been steadilygrowing in recent years. According to the NationalAssociation of Counties (NACo), the share of gigeconomy workers in the U.S. labor market has sky-rocketed from 10.1 percent in 2005 to nearly 16percent in 2015 (Harris, 2017) to nearly 36 percentin 2018 (McCue, 2018). Companies like Uber andLyft hold a major presence in the market, fulfill-ing a market demand for ridesharing both acrossthe country and around the world, at least in thecase of Uber. The National Association of Coun-ties delineates two primary forms of gig economywork. The first is labor providers (“contractors”),which includes drivers, handymen, and delivery-men, among others. These workers tend to be low-income, less-educated workers who rely on gigeconomy work for the majority of their livelihood.The other form of gig economy work involves goodproviders (“freelancers”), including artists, crafts-men, and bespoke clothing retailers. These workerstend to be higher-income, educated workers whorely on gig work as a subset of their earnings (Har-ris, 2017). For the purposes of this paper, we willlargely be focused on the former category for our

user base. We chose to focus on contractors be-cause a larger proportion of their daily lives arecentered around working in the gig economy.These workers typically do not enjoy the samerights and benefits that salaried employees enjoy,including minimum wage, overtime pay, paid sickleave, and other benefits (Asmelash, 2019). Theresult is that a growing proportion of the U.S. laborforce is working in a space that is, by virtue of itsdistributed nature, flexible work style, and lack ofregulation, highly volatile and unpredictable. As-melash (2019) states that approximately 1 in 10workers depend on low-paying gig work as theirprimary source of income, leading to criticism ofthe industry as exploitative of its workforce.This view of the gig economy has pressed legisla-tive bodies to intervene, as exemplified by Califor-nia’s 2020 law requiring the state’s gig work forceto be treated with the same rights as those of con-ventional employees. Several gig economy tech-nology platforms have skirted the issue by definingtheir business as a marketplace between peoplewho demand services and those who provide them(Irwin, 2019).Also unique to this space is the ability of work-ers to switch seamlessly, at essentially no cost, be-tween multiple platforms (known as multi-homing),a practice extremely common amongst drivers inthe ridesharing space and in related gig economyspaces like food delivery (Gad Allon, Maxime Co-hen, and Wichingpong Sinchaisri, 2018). Accord-ing to (Campbell, 2017), ridesharing contractorsfor Uber and Lyft represent nearly 70 percent of theon-call driving work force, and 25 percent of thesedrivers drive for more than those apps exclusively.

1.2 Business Implications

The gig economy space is now a significant partof the global economy, with total spending esti-mated at 4.5 trillion USD worldwide and 1.3 tril-

lion USD in the U.S. alone in 2018 (Arcuni, 2019).The global ridesharing market alone was valued at51.3 billion USD in 2017 and is projected to growsteadily at over 20 percent per year until 2025, atwhich point it will reach a value of approximately220.5 billion USD (Costello, 2019). This massivegrowth has come hand in hand with a large upsurgein the value of platforms which operate within thisspace. In particular, Uber is valued at an estimated120 billion USD, and Lyft is valued at an estimated23 billion USD (Delventhal, 2019). These techsuccess stories have come at the heels of an in-creasingly dissatisfied work force which has seenitself excluded and alienated from the success thesecompanies now celebrate. According to McCue(2018), only 34 percent of contingent workers andregular workers said the hours that they work were“good for them,” and gig economy workers are lesslikely to report being paid as timely and accuratelyas their more traditionally-employed counterparts.This points to the growing concerns among gigeconomy workers that the space is no longer play-ing to their benefit, and that the flexibility that wasonce appealing to them is now leading to high lev-els of uncertainty.Exacerbating this issue is the fact that several keyplayers in this space are not yet profitable. In itssecond quarterly earnings as a public company,Uber reported losses totaling up to 5.2 billion USD.Lyft’s earnings report was arguably better, but stillnegative, at a loss of 644 million USD (Hawkins,2019). Given the low profitability in this space,Uber, Lyft, and other companies are in a constantstruggle against the bottom line, meaning they havelittle incentive to share earnings with the driverswho facilitate the ridesharing process. In the future,we anticipate that this tension is likely to persist,as it is unlikely that ridesharing companies will beprofitable for a long time without a major paradigmshift in technology (like self-driving cars) or regu-lation (like subsidies and partnerships). Our goal isto cater to the workers in this space in spite of thisoverarching uncertainty and flux.

1.3 Goal

We believe that there is a strong need in the gigeconomy for stability and predictability. On onehand, the flexibility associated with the gig econ-omy and ridesharing in particular makes it an ap-pealing proposition for many workers (Gad Al-lon, Maxime Cohen, and Wichingpong Sinchaisri,

2018). However, this flexibility also introduces agreat deal of uncertainty, which we hope to mit-igate. Our first iteration of the app focuses on asubsection of the gig economy: ridesharing, withthe hope that in the future we may expand to otherdomains. We chose this space because of its wealthof data and accessibility of contractors for demandtesting.We have identified three primary areas of greatestneed in the gig economy which currently are notavailable to the workforce:

• Past projection. Workers in the industry haveno centralized place to share or record infor-mation about past trips taken.

• Predictive recommendation. While work-ers are engaged in instantaneous price-comparison, there is currently no mechanismfor them to project future earnings days orweeks in advance.

• Financial recommendations. We would liketo provide high-level information and tips togig economy workers who rely on sustainedlong-term contract work in this space.

In this paper, we introduce the methods, key com-ponents, evaluation, and business implications 1 ofour technology.

2 Business Model

2.1 Stakeholder AnalysisWe have identified several key stakeholders relatingto our product.The first stakeholders are our end-users: work-ers in the gig economy and drivers for rideshar-ing apps. As described in the introduction section,these users largely suffer from the unpredictablenature of the gig economy, as evidenced by irregu-lar pay and overall worker dissatisfaction (McCue,2018). Many workers fail to account for depreci-ation and opportunity costs which make their realwages far below minimum wage. Our product isa solution to this problem because it gives work-ers in this space the ability to forecast and betterunderstand how they should segment their timebetween apps. Our goal as an app is to provideresources that tackle the 3 areas of need identifiedin the goals section. By providing a scheduling app,

1For the business analysis of our work, refer to the follow-ing sections: Business Implications, Business Model, DemandScoping

we allow users to record their rides and the revenuethey received from them. By providing predictiverecommendations through our app, we can increasecertainty in scheduling work plans, allowing work-ers to stabilize their schedules. Our app will alsoprovide resources and financial recommendationsto end users through job referrals and aggregatedcompany information to allow users to make in-formed decisions about their work life.The companies in the gig economy will also bemajor stakeholders. Companies such as Uber andLyft may see our app as an attempt to drive downtheir revenues, but we are not directly at odds withgig economy companies. In fact, Uber and Lyftmay want to better understand driving habits, forwhich we can provide information regarding driversatisfaction and behavior, as well as multi-hominginformation. Other companies in this space maywant to advertise their services on our platform–wewill touch on this more later in our discussion ofthe GigFitter revenue model. It is also possiblethat our app can also act as a platform connectinggig economy workers with alternative revenue op-tions, which is an attractive proposition both forthe workers who are seeking opportunities as wellas the companies seeking more workers.Another stakeholder segment is the end-consumerfor gig economy workers. For example, in rideshar-ing it would be the riders who benefit from the ser-vice and pay for the service itself. While we don’tanticipate significant changes to this user base, weacknowledge the possibility that large-scale deploy-ment of our service may lead to higher prices to theend user as both workers and companies optimizefor higher earnings. We believe that on our currentscale, the impact will be negligible, and any pricedisparities resulting from our app will be mitigatedby price competition.The final major stakeholder segment we identifiedwas that of governments, particularly municipaland state-level legislature. As mentioned in theintroduction, the growth of the gig economy hasprompted increased regulation in favor of workerprotection and benefits, as exemplified by the newCalifornia legislation described by Irwin (2019).We believe that while increased legislation forworkers will increase stability, our service providespredictability in other ways that will continue to bea value add even after the introduction of worker-friendly legislation.

2.2 Market Opportunity and ResearchTotal spending in the gig economy approached 4.5trillion USD in 2018 (Arcuni, 2019). The globalridesharing market alone was valued at 51.3 billionUSD in 2017 and is projected to grow at over 20%per year until 2025 (Costello 2019). Within theridesharing space, two companies, Uber and Lyft,have dominated the US market and competition haslargely consolidated. Despite the success of thesecompanies (and in part because of it), contractorsin this space have increasingly voiced dissatisfac-tion with their treatment and the inability to havepredictable work lives. These issues include, butare not limited to, those of scheduling, pay, andworker’s benefits. According to McCue (2018),only 34% of contract workers said they enjoyedtheir work schedules. This is contrary to the veryobjective of working in the gig economy relative toa more traditional employment space.

2.3 Customer SegmentsWe conducted several live interviews with on theorder of 50 Uber and Lyft drivers throughout thesemester. We identified broader trends within thedriver base and used them to inform our high-leveldecisions in the design of our applications.We segmented the driver base into three primarydriver types:

• Optimizer Drivers. These consist of highlycommitted drivers who actively engage inmulti-homing. These drivers typically takean active approach to optimize pay, and workup to 40 hours per week.

• Scheduled Drivers. These consist of driverswho largely adhere to a set schedule, or whoset a daily quota or window for when to drive.

• Spontaneous Drivers. This driver segmentconsists largely of students or retirees whotend not to follow a particular schedule. Mostof these drivers tend to drive part-time andleverage this job as additional income andspending money rather than core earnings.

Note that of these three groups, optimizing driverstended to have the highest pay per hour by virtueof when and how they drive. Next, we set out toidentify common strategies within the optimizersegment to inform our recommendations to otherusers. We identified the following common strate-gies among the optimizer segment which led toincreased pay per hour in ridesharing:

• Multi-homing. Optimizers were overwhelm-ingly more likely to be multi-homers. Theydescribed how they used both Uber and Lyftsimultaneously and concurrently waited forrequests from both. Upon receiving a request,they would silence the other app until the tripwas finished. This allowed workers in thisspace to maximize the utilization of their carand minimize waiting time between rides.

• Strategic Time Allocation. Optimizers werealso more likely to have more flexible timeallocations, choosing to drive during peak de-mand times such as rush hour or during peakevent times, such as Friday evenings or aftersporting events.

• Strategic Geographic Positioning. Finally,we found that optimizers were likely to posi-tion themselves strategically in locations thatwould allow them to have the highest demandwith shorter trips, such as downtown Philadel-phia. They would also position themselves inlocations with high anticipated demand, suchas sports stadiums and airports.

Among the scheduling drivers and spontaneousdrivers, few incorporated multi-homing or strate-gic time allocation or geographic positioning intheir decisions behind when and where they drove.Notably, even optimizers tended not to considerlong-term demand planning for the future.

2.4 Related Work and Existing Services

We have identified two primary services that alsooperate in the gig economy space. Gigworker 2 is acentralized website that shows various openings forgig economy jobs and their expected hourly pay-out, and although it does a good job of enumeratinga multitude of available opportunities, it does notgive recommendations or forecast profit at a granu-lar level (e.g. expected profit hourly). Its dataset islimited and the product is built on top of Wordpresswith limited engineering and data scalability.Ridester 3 is a mobile application specific to Uberand Lyft that switches between the two dependingon which service will provide a higher payout byleveraging a screen reader on users’ devices. It alsotakes into account other user preferences, such asautomatically rejecting rides when the destination

2https://gigworker.com/3https://www.ridester.com/

is out of range of the user’s preferred area of op-eration. However like Gigworker, Ridester doesnot provide any predictive forecasting, and insteadonly recommends the highest paying option at thetime of use, not the future.Neither of these services combine a high-level anal-ysis of the gig economy and provide future recom-mendations around scheduling work for extendedperiods of time. We believe that these aspects arecrucial in allowing ridesharing (and more generally,gig economy work) to approach the level of stabil-ity afforded to individuals in regularly employedwork.Our service differs from these existing services bycovering all three of the needs that we describedearlier. Our app is a scalable, data-driven recom-mendation system that provides recommendationsfor the future, not just instantaneous price compar-isons. Our application also provides services whichallow users to track previous and future rides. Fi-nally, our app integrates company information andaggregated news to allow gig economy workers tomake financial decisions relating to their work.

2.5 Value PropositionWe view ourselves as a provider and aggregatorof information and opportunity for gig economyworkers. Our goal is to address the needs for (1)past projection, (2) predictive recommendation and(3) financial stability among gig economy workers.Each of our components is tailored to address oneor several of the needs outlined here, which webelieve are central to the overall predictability ofworking in the gig economy space.

2.6 Revenue and Cost ModelWe considered several revenue models for our ser-vice, including subscription and ad-based revenuemodels. We concluded that subscription modelswere not aligned with our value proposition as aservice provider for the workers in the gig econ-omy (it did not make sense to charge them for ascheduling app), and that generic ad-based revenuemodels would be innocuous and counterproductive,decreasing overall user satisfaction with our appli-cation.Our revenue model centers around our identity asan aggregator of information and opportunity. Ourapp will generate revenue by referring registeredusers to job sites. When users sign up through ourplatform we will set up affiliate accounts with eachof these companies, and we will do the same for

financial and educational service providers in thisspace. Being a work-life planning platform enablesus to partner with a variety of financial planningservices and offer more tailored, monetizable ser-vices as workers continue to use the platform.Our costs are relatively lean. We plan to acquirecustomers through paid advertising campaigns andword of mouth. Web based storage and computingcosts are low recurring costs. Upfront developmentcosts and costs to adapt the product to new geogra-phies and verticals are our highest costs.

3 Technical Components

3.1 Data Collection

Our data collection consisted of two main com-ponents: (1) data collection for broader frontendrecommendations and (2) data collection for ourpredictive revenue model.The data we collected for our model came froma variety of sources. One dataset originated fromthe Uber TLC-FOIL response released by FiveThir-tyEight on Github 4. The most significant datasetwhich we used to train our preliminary modelscame from a set of Boston ridesharing data releasedon Kaggle consisting of Uber and Lyft rides withinthe city 5. This data consisted of information suchas price, distance between destinations, cab type,source and destination information, ride type (e.g.UberXL, Uber Black), surge multipliers, and times-tamps.

3.2 Model

The initial model we chose used the Boston Datasetto train a decision tree regression model of deptheight. The surge multiplier was dropped in order toensure that our model interacted with as little priceinformation as possible. The output variable was

4https://github.com/fivethirtyeight/uber-tlc-foil-response

5https://www.kaggle.com/ravi72munde/uber-lyft-cab-prices

the price of the ride. Our initial choice of a decisiontree architecture was informed by the prevalenceof categorical information. The decision tree wastrained on an ID3 algorithm using scikit-learn. Thecode for the model and the output of the decisiontree in a graphical format is available in our modelGitHub repository 6.Over the course of development we realized thismodel did not give the functionality we neededgiven that, when a user is planning their week, wedo not have control over features like location oncethey actually start driving. Thus, we migrate to arules-based model for segmenting users based ontheir preferences (like willingness to drive in traffic,number of hours, desire for a consistent schedule,etc.) and providing noisy recommendations to caterto their preferences. A discussion of the modelefficacy is included in the following section.

3.3 App

Our final deliverable is an application that is acces-sible to gig economy workers via the Internet bothon mobile phones and desktop computers. The appprovides recommendations to ridesharing and otherfreelance contractors on available opportunities andways to optimize their work. The application con-sists of several components. These components arealigned with the primary areas of greatest need inthe gig economy as well as the strategies we de-rived from our demand scoping work as outlinedearlier in this paper:

• Learn. Centralized news curation relating togig economy work. This page will allow gigeconomy workers to stay up-to-date on the lat-est trends and changes in legislation regardinggig economy work.

• Companies. A compilation of companies thatare active in the freelance / gig economy space,as well as information about pay, wage, re-quired skills, and required assets.

• Scheduling. Scheduling functionality forridesharing drivers using a combination ofpredictive analytics and information gatheredthrough demand scoping. Drivers can sub-mit their availability for any week and canreceive recommendations within their avail-ability windows for when to work.

6https://github.com/jordanlei/gig-economy

4 Evaluation

4.1 Preference-based RecommendationDifferentiation

In this section, we evaluate the fact that differentuser profiles would result in distinct recommendedschedules. In the absence of a user-study, we’vegenerated a set of 100 users, split into two user-profile groups consistent with possible ranges weobserved in preferences for traffic. Users weregenerated with availability from 7 AM to 11 PMwith a weekly driving time normally distributedhoursweek ∼ N(20, 3). 50 users were selected toemulate traffic-loving drivers on a 5-point LikertScale wtraffic ∼ N(4, 1) and 50 were assigned tobe traffic-hating drivers wtraffic ∼ N(1, 1).

4.1.1 Results

4.1.2 DiscussionBased on our aggregate results, it is clear that ouralgorithm generates divergent recommendationsfor users with different traffic preferences. Basedon this proof of concept, we can extend this workto other preferences including weather and events

to allow for more holistic user representations anduser segments in the future.

4.2 Simulation

In the absence of a user study and due to social dis-tancing limitations, we have proceeded to conduct alarge-scale user simulation based on possible ridesfrom the Boston Kaggle Dataset. More specifically,this simulation is designed to evaluate our hypoth-esis that for profit-maximizing users, the realizedwages from our recommended driving times wouldbe higher than a baseline (defined by spontaneousscheduling decisions). In the following subsection,we describe the parameters of the simulation, theresults, and a brief discussion.

4.2.1 Simulation DesignThe simulation was designed with the followingstrong limitation in mind: drivers can choosewhere they start driving but cannot choose wherethe ridesharing app takes them. The design ofthe simulation is specifically tailored to addressthis parameter; drivers are assigned the nextavailable ride based on where they currently are(the source) and taken to some secondary location(the destination), which will then become thesource for their next drive. This is consistent withhow drivers are assigned rides in the real world.Note that in addition to simulating realisticdriving assignment scenarios, our simulationalso realistically represents drives taken. Eachsimulated drive is taken from the Boston KaggleDataset, representing a real drive taken by an Uberor Lyft driver in the past. This data also includesthe true distance, source, destination, and price ofthe ride, so these values are true-to-life.For a given weekly schedule, a driver may beassigned to several different shifts, with eachshift having a start and end time. By simulatingtheir earnings on each shift, we can aggregatetheir earnings to estimate a set of possible weeklyearnings.As mentioned earlier, ride simulations make use ofthe Boston Kaggle Dataset, consisting of 693,071unique rides conducted over several months inBoston. The dataset consists of a few importantfields, including but not limited to: Start Time,End Time, Price, Start Location, End Location 7

7Locations are listed within the city of Boston, consistingof the following: ’Back Bay’, ’Beacon Hill’, ’Boston Uni-versity’, ’Fenway’, ’Financial District’, ’Haymarket Square’,’North End’, ’North Station’, ’Northeastern University’,

There are a few primary inputs to the simulation–astart time, start location, and an end time 8.We construct a simulated ride in a frameworkdescribed by Algorithm 1.In simple terms, the algorithm describing a ride

Algorithm 1 Ride Shift SimulationRequire: startTime, startLoc, endTime, rideList;Ensure: rideList is sorted by start time

ridesTaken = {}runningPrice = 0possibleRoutes = {x in rideList s.t. x.start >startTime, x.end < endTime, x.source = start-Loc}while len(possibleRoutes) > 0 do

wait = genWait()takenRoute = drives[0]ridesTaken = ridesTaken ∪ {takenRoute}

runningPrice += takenRoute.priceduration = takenRoute.durationstartLoc = takenRoute.deststartTime += duration + wait

possibleRoutes = {x in rideList s.t. x.start> startTime, x.end < endTime, x.source =startLoc}

end whilereturn runningPrice, ridesTaken

simulation takes data points from the BostonKaggle Dataset which fit the source, start time,and end time descriptions. Then it generates a setof rides where the destination of the prior ridebecomes the source or starting point of the nextride. If we view the city of Boston as a graph wherethe nodes are the districts listed in the dataset andthe edges are the time it takes to travel betweendistricts, we may view this simulation alternativelyas taking several possible contiguous paths throughthe graph. The pseudocode description obscuressome of the minute details required to model thetraffic, ride duration, and wait times, associatedwith a more accurate simulation. Our model alsotakes into account the difference between theprice earnings and the wage received, which arenot described here for simplicity; however, thepseudocode provides a reasonable approximation

’South Station’, ’Theatre District’, ’West End’8Uber/Lyft riders have limited control over where they

are routed to–our choice of not defining an end location isconsistent with this reality

for what our current simulation does on a largescale.The following image shows a sample outputof a single simulated ride with the followingspecifications:startLoc = "Boston University",startTime = 04:40:56 2018-11-26,endTime = 06:40:56 2018-11-26

In the full simulation, each schedule would consistof several shifts, which would each be representedin a similar way.

4.2.2 Experiment DesignThe experiment was conducted as follows. 100,000simulated users were generated with traffic prefer-ences normally distributed wtraffic ∼ N(4, 0.5)with availability from 7AM to 11PM and weeklyhours normally distributed hoursweek ∼ N(20, 3).Users were either assigned to be in the controlgroup (n = 5000) or in the treatment group(n = 5000). Those in the control group wereassigned to drive spontaneously, a random set ofdriving times within the specified time windowthat matched the hours they were willing to drive.Those in the treatment group were assigned tofollow the recommendations given by our model.We tracked simulated user earnings over the courseof one week.

4.2.3 ResultsWe found that the treatment group, which wasassigned to adhere to our model, outperformed the

baseline. On average, the treatment group earned$335.37 (SD = 58.88), compared to the controlgroup $318.79 (SD = 53.06). The average usergained $16.58 compared to the baseline, a 5.20%improvement within a given week.Because we observed large variances in our data,we conducted a one-sided t-test for independentsamples to determine the statistical significanceof our effect. Our null hypothesis, H0, is thatx̄control ≥ x̄treatment, and our alternative hy-pothesis H1 is that x̄treatment > x̄control. Afterour t-test, our t-statistic was found to be 14.78,meaning our p-value was well below our thresholdof 0.01.

4.2.4 DiscussionOur simulation realistically modeled driving situa-tions that would be faced by a ridesharing worker inreal-time scenarios. Our experiments demonstrateda statistically significant (p < 0.01) increase inearnings from the treatment group compared to thecontrol, demonstrating that our offering truly addsvalue to ridesharing workers who are willing orable to optimize for earnings.

5 Societal Impact and Ethical Concerns

GigFitter is designed to bring stability to the livesof ridesharing workers, who often come frommarginalized communities and many of whom relyon ridesharing services as their primary source ofincome.We believe our offering provides stability on theindividual level in several key ways:

• Economic Stability. A significant subsectionof the ridesharing workforce depends solelyon the industry to make a living. Our appprovides a way for them to anticipate futureearnings and maximize utilization, based ontheir preferences.

• Bookkeeping. Working in this space is associ-ated with high uncertainty that stems from alack of job security, insufficient tools to recordand track earnings, and no ability to union-ize or share information. Our app provides away for them to track past expected earningsand learn relevant information pertaining toindustry-wide changes.

• Scheduling. Despite being marketed as a lib-erating option for work, only 34% of contract

workers said they enjoyed their work sched-ules (McCue, 2018). Our app provides a wayfor them to schedule their work and plan aweek or more in advance.

Beyond these metrics, we also believe that ourservice is essential to the long-term health of theridesharing service industry:

• Turnover. 68% of workers stop driving within6 months of starting (Brown, 2019). Our appmakes it possible to sustain long-term workin this uncertain environment, which woulddrive down turnover rates.

• Social obligation. Companies, users, and con-tractors who depend on this space have anobligation to ensure the stability of the livesof the drivers. Our app fulfills this obligationby giving drivers the negotiating power to settheir own hours and plan ahead.

• Worker benefits. Our role as a stabilizing forcein the industry actively supplements effortsmade in legislation to give contractors morelegal rights and representation by spreading in-formation about such initiatives and providingalternative sources of value.

5.1 Data PrivacyAll location or route data is and will be provideddirectly by users. No data is kept without the con-sent and knowledge of the users, and no data canbe used to trace the driving patterns directly backto the user. Users know what data is entering oursystem. For scaling our recommendations beyondthe rules based model, we will take advantage ofride data. This data will be in similar format toour base ride datasets from Kaggle namely in thatwe remove as much information about drivers andrides as possible, keeping only essential informa-tion like a generated driver ID, time, and start andend approximate location of ride.

5.2 Algorithmic BiasOne concern that was brought up was algorithmicbias in favor of certain drivers or geographic loca-tions. In this section we respond to both of theseconcerns.With respect to driver bias, we do not priori-tize different drivers differently. The drivers areanonymized to the recommendation system, sothere is not any bias we insert in terms of recom-mending when to drive. In the updated algorithm

where we reroute certain drivers, it will come ata “first come first serve” basis–after a threshold isreached, we will reroute drivers once we predictdropping prices.With respect to different geographic locations, themain concern here is that the app will route usersaway from under-served locations. While we agreethat this is a point of concern, we believe it is out ofthe scope of this course project. We have no controlover where drivers are routed or where they chooseto drive. Our recommendation system only recom-mends when they should drive. While we agreethat this is a larger issue, we neither recommendnor exacerbate any part of the routing process. Thiscould be a larger conversation about the ethics ofthe gig economy in general, and we believe we canaddress other latent needs for low-income individ-uals in the gig economy space while this debateplays out on a different stage.

5.3 ScalingOne possible concern is that if all drivers use ourservices, this would hypothetically drive up sup-ply with constant demand, driving down pricesand wages for drivers—-the outcome of this sce-nario is that drivers go back to earning the samewages they did before. Our app has several miti-gation strategies to prevent this. First, we segmentusers based on preferences: different users will begiven different recommendations, based on theirown unique preference which they set in the GigFit-ter web app. Thus, different user segments will notcompete directly in terms of when they supply theirlabor based on our recommendations. Right nowwe have two user segments as a proof of concept.Second, as we scale, we can alter recommendationsto make sure that users do not interfere with oneanother. Once we have aggregate user data it willbe possible to make more granular recommenda-tions based off our user base beyond informationthey currently supply through surveys.

5.4 Future WorkOur recommendation model can also expand be-yond its current form by incorporating weatherdata (e.g. many rideshare users are more likelyto request rides in increment weather, though manydrivers per our interviews perfer only driving innice weather or under predictable conditions) andevents data (e.g. Valentines, Christmas, festivals,etc.). Our model allows for easy integration withthese data sources. While we have generated pre-

dictions on short term events data and weather dataalready, we look forward to implementing morelong term and automated sourcing of this data ingenerating our recommendations. This, in tandemwith basing recommendations on data collectedfrom users of our app, will enable us to scale tomore geographies and to more granular levels ofrecommendations.

5.5 Lessons LearnedThe gig economy is a rapidly moving space on allfronts: the software and products are forever shift-ing, consumers demand new things every week,and workers in the space are all caught somewherein the middle. COVID-19, if anything, has accel-erated trends in this space and has highlighted thedisproportionate impact it has on different groupsof people. To this point, we have truly loved ex-ploring how we can use technology to play to thebenefit of workers in the gig economy and to betterunderstand how the space is evolving. There iscertainly a lot more work to be done by us, by othercompanies, and by regulators.

6 Conclusion

The gig economy space is currently seeing a mas-sive upsurge in growth across a variety of differentindustries, from ridesharing to food delivery to out-sourcing. While these jobs confer a high degree offlexibility in work hours, this work also comes witha high degree of uncertainty, both in schedulingand financial security.Our app, GigFitter, addresses a social need amonggig economy workers to allow for a more stableand predictable work life without sacrificing earn-ings. Our application combines insights drawnfrom drivers who optimize earnings as well as pre-dictive analytics tools to make curated recommen-dations for drivers in the ridesharing space. We alsoprovide a centralized website which compiles newsarticles and company information to allow users tomake informed decisions about their involvementin the industry. Our hope is that our service willallow the gig economy to become a sustainableand stable industry for the growing population ofworkers who depend on it.

ReferencesJennifer Arcuni. 2019. Global gig economy reaches

usd 4.5 trillion. Staffing Industry Analysts sia.https://www2.staffingindustry.com/row/

About/Media-Center/Press-Releases/Global-Gig-Economy-Reaches-USD-4.5-Trillion.

Leah Asmelash. 2019. Your Questions Aboutthe So-Called Gig-Economy, Answered.Cable News Network. CNN. https://www.cnn.com/2019/09/11/business/gig-economy-explainer-trnd/index.html.

Harry Campbell. 2017. The RideshareGuy 2017 Reader Survey. The RideshareGuy. The Rideshare Guy. https://docs.google.com/document/d/1QSUFSqasfjM9b9UsqBwZlpa8EgqNj6EBfWybFBSHj3o/edit#heading=h.14a26bf76t1u.

Hector Costello. 2019. Global Ride SharingMarket 2019, By Type, Expanse, Owner-ship, Business Model, Demographic andGrowth Opportunities to 2025 Reuters. Reuters.https://www.reuters.com/brandfeatures/venture-capital/article?id=83120.

Shoshanna Delventhal. 2019. Read This BeforeInvesting in Uber and Lyft IPOs Investopedia. In-vestopedia. https://www.investopedia.com/read-this-before-investing-in-uber-and-lyft-ipos-4584464.

Gad Allon, Maxime Cohen, and Wichingpong Sin-chaisri. 2018. The Impact of Behavioral and Eco-nomic Drivers on Gig Economy Workers.

Jonathan Harris. 2017. The Future of Work:The Rise of the Gig Economy. NationalAssociation of Counties. NaCo. https://www.naco.org/featured-resources/future-work-rise-gig-economy.

Andrew J. Hawkins. 2019. Uber lost over $5Billion in One Quarter, but Don’t Worry, itGets Worse The Verge. The Verge. https://www.theverge.com/2019/8/8/20793793/uber-5-billion-quarter-loss-profit-lyft-traffic-2019.

Neil Irwin. 2019. Maybe We’re Not All Goingto be Gig Economy Workers After All. TheNew York Times. The New York Times. https://www.nytimes.com/2019/09/15/upshot/gig-economy-limits-labor-market-uber-california.html.

TJ McCue. 2018. 57 Million U.S. Workers Are Part OfThe Gig Economy. Forbes. Forbes. https://www.forbes.com/sites/tjmccue/2018/08/31/57-million-u-s-workers-are-part-of-the-gig-economy/#6ea8e5157118.

A Appendix

Figure 1: Home Page

Figure 2: Traffic Page

Figure 3: Articles Page

Figure 4: Companies Page

Figure 5: Gig Page

Figure 6: Registration Page

Figure 7: Linked Article Page