Spatial On-Demand Economy Robustness: A Provider-Centric Utilization Aware Approach

Transcription

1 Spatial On-Demand Economy Robustness: A Provider-Centric Utilization Aware Approach Reem Ali Department of Computer Science and Engineering University of Minnesota Co-investigators: Shashi Shekhar *, Shounak Athavale #, Eric Marsman # * Computer Science & Engineering Dept., University of Minnesota # FORD Motor Company This material is based upon work supported by FORD University Research Program (URP), the National Science Foundation under Grant No , IIS , and IIS , the OVPR U-Spatial and Minnesota Supercomputing Institute (MSI) at the University of Minnesota (

2 Outline Motivation Basic Concepts & Problem Definition Challenges Related Work Proposed Approach Evaluation Conclusion & Future Work 2

3 Motivation Increasing proliferation of mobile technologies led to emergence of the on-demand and sharing economies, with $50+ billion in spending More growth expected with emergence of self-driving cars Service Providers Static Moving Static Consumers Moving (Our Focus) People heading to: Lunch in busy hour Dinner on Friday night Stores on Black Friday Movies on popular nights (Future Work) Raises several questions: How to meet time-varying demand with a fixed supply? How to ensure balanced provider utilization to keep ecosystem functioning and robust against variations in market balance e.g. Uber encourages drivers to stay on road using earning target, next fare alerts 3

4 Outline Motivation Basic Concepts & Problem Definition Challenges Related Work Proposed Approach Evaluation Conclusion & Future Work 4

5 Basic Concepts (1/2) A spatial network G = (V,E): consists of a node set N and an edge set E Each node is associated with a pair of real numbers (latitude, longitude) Edge set E is a subset of the cross product N N. Each edge is associated with a travel time cost A Service Provider: A service provider registered in the spatial broker system. A Consumer Request: A request from a mobile consumer. Each is associated with: Service Provider Provider id id p, Node in the spatial network loc p Service rate per hour s p Consumer Request Consumer id cid r Request arrival time a r Node in the spatial network loc r Max. acceptable travel time d max, r Max. acceptable waiting time w max,r 5

6 Basic Concepts (2/2) A Service Provider Proposition: provided by the broker in response to a consumer service request. Defined as a quadruple (r, p, d, w) where: r represents a consumer request p represents a service provider proposed for servicing request r d represents the shortest travel time between r and p w represents the waiting time for c before it is served at provider p Utilization of provider p: u P = No. propositions accepted by consumers from provider p servicerateper hour total time in hours 6

7 Problem Definition: On-Demand Spatial Service Propositions (OSSP) Input: A spatial network G A set P of service providers in G A set R of consumer requests arriving dynamically in G A number of required propositions K per request A timeout interval length t timeout Output: K service provider(s) propositions for each request r ϵ R Objectives: Broker-centric: Maximize number of matched requests Provider-centric: Constraints: Maximize number of matched providers Minimize the variance of provider utilization For each output proposition (r, p, d, w), we have d d max,r and w w max,r For each provider p, the number of propositions with which p is associated per hour s p 7

8 Example Input: Toy Example X Y Example Output: K=1, t timeout =1 time unit, & unit edge costs Assume that both providers are valid candidates for each consumer request Heuristic Final Match # matched requests # matched providers Utilization u X u Y σ Least Travel Cost (LTC) X A B C D E F Y 6 1 6/8 =

9 Challenges Need to satisfy many conflicting requirements Broker: maximize No. matched requests while balancing provider utilization. Service providers: Maximize No. assigned consumers Consumers: Minimizing travel and waiting times Consumers requests are unknown in advance Finding the set of K-propositions that maximizes no. matched requests at any given time instant is an NP-hard problem. 9

10 Limitations of Related Work Balances provider assignments to keep provider eco-system functioning? No Yes Least Travel Cost (LTC) (ridesharing [1,2,3,4,5,6,], spatial crowdsourcing [7, 8,9,10] ) Least Location Entropy Priority (LLEP) (spatial crowdsourcing [7] ) X A B C D E F Y Utilization Aware? No Preliminary work Least Accepted First (LAF) Least Appearance as Candidate First (LCF) X A B E Y C D F Yes Proposed work Least Utilized First (UA-LUF) X A B E F Y C D 10

11 Example Input: Possible Outputs: Toy Example X Y K=1, t timeout =1 time unit, & unit edge costs Assume that both providers are valid candidates for each consumer request Heuristic Final Match # matched requests # matched providers Utilization u X u Y σ Least Travel Cost (LTC) X A B C D E F 6 1 6/ Y Least Location Entropy Priority (LLEP) X A B C D E F Y 6 1 (arbitrary) 6/ Least Accepted First (LAF) X A B E Y C D F 6 2 3/8 3/

12 Toy Example: Matching Details Example Input: Possible Outputs: K=1, t timeout =1 time unit, & unit edge costs Assume that both providers are valid candidates for each consumer request Heuristic\Matching t=0 t=1 t=2 t=3 Final Match Least Travel Cost (LTC) (A, X) (B, X) (C, X) (D, X) (E, X) (F, X) X A B C D E F Y Least Location Entropy Priority (LLEP) (A, X) (B, X) (Arbitrary) (C, X) (D, X) (Arbitrary) (E, X) (Arbitrary) (F, X) (Arbitrary) X A B C D E F Y Least Accepted First (LAF) (A, X) (B, X) (C, Y) (D, Y) (E, X) (F, Y) X A B E Y C D F 12

13 Outline Motivation Basic Concepts & Problem Definition Challenges Related Work Proposed Approach Evaluation Conclusion & Future Work 13

14 Taxonomy of Proposed Work Objective Maximize No. Matched Consumers Maximize No. Matched Providers + Minimize Provider variance Proposed Work Fewer Candidates First Consumer Prioritization (FCF) Conflict-aware Prioritization (CAP) Least utilized first (UA-LUF) A consumer-priority-based Greedy Matching Approach 14

15 Proposed Approach (1/3): A provider centric utilization-aware Heuristic Consumer-centric Broker-centric Provider-centric Related/Prelim Work Least Travel Cost (LTC) Least Location Entropy Priority (LLEP) Least Accepted First (LAF) Related and preliminary work is utilization agnostic: May not balance utilization Least Utilized First Heuristic (UA-LUF): Provider-centric, utilization-aware Prioritize least utilized providers Score of candidate propositions provider utilization 15

16 Input: Toy Example X Y K=1, t timeout =1 time unit, & unit edge costs Assume that both providers are valid candidates for each consumer request Possible Outputs: Heuristic\Matching t=0 t=1 t=2 t=3 Final Match Least Utilized First (UA-LUF) u X, u Y Match 0,0 2/8, 0 2/8, 2/4 3/8, 2/4 (A, X) (B, X) (C, Y) (D, Y) (E, X) (F, X) X A B E F Y C D 16

17 Heuristic Final Match # matched requests Least Travel Cost (LTC) Toy Example: Performance Measures Comparison X A B C D E F Y # matched providers Utilization u X u Y σ 6 1 6/ Least Location Entropy Priority (LLEP) Least Accepted First (LAF) X A B C D E F Y X A B E Y C D F 6 1 (arbitrary) 6/ /8 3/ Least Utilized First (UA-LUF) X A B E F Y C D 6 2 4/8 = 0.5 2/4 =

18 Proposed Approach (2/3): A consumer-priority-based Greedy Matching Approach Phase 1: Candidate Evaluation Identify candidate providers for each request Compute score of each candidate based on provider utilization (UA-LUF) consumer priority queue candidates Phase 2: Consumers Prioritization Fewer Candidates First: Sort consumers by no. candidates Tie breaker: favor shorter waiting times Phase 3: Matching For each consumer, match to first K candidates in its priority queue Reserve service time slots for matched consumers 18

19 Proposed Approach (3/3): Conflict-aware Consumer Prioritization (CAP) Definitions: Provider Conflict Score: No. of occurrences of provider as a candidate proposition at current time instant Non-conflicting candidate: a candidate proposition with a provider conflict score of 1 Candidate evaluation Phase Compute provider conflict scores Consumer Prioritization Phase Prioritize consumers with smaller no. of non-conflicting candidates Tie breakers: smaller no. candidates, larger sum of conflicts, shorter waiting times Matching Phase: Apply UA-LUF heuristic Tie breaker: prioritize propositions with smaller conflict score 19

20 Outline Motivation Basic Concepts & Problem Definition Challenges Related Work Proposed Approach Evaluation Conclusion & Future Work 20

21 Experimental Setup (1/2) Experimental Goals: 1. Comparative Analysis: Does the utilization-aware provider centric approach improve provider fairness/equity? Does a provider-focused approach reduce business volume? 2. Self Analysis: How does the proposed approach scale with the number of consumers? Does consumer-prioritization improve the business volume? How do the (online) proposed approach compare with offline optimizers? 21

22 Experimental Design: Experimental Setup (2/2) 120 restaurants in Minneapolis, MN Simulated 10 hrs Market-Balance : Ratio between aggregate supply and demand (SDR) SDR = sum of servicerate per hour for all providers total no. requests in that hour 22

23 Experimental Setup: Performance Measures Solution Quality: Provider Fairness:» Standard deviation of provider utilization» Average utilization for lowest 10% providers» % of Matched Providers Business Volume:» % of matched requests Performance:» Average response time of consumer request 23

24 Comparative Analysis: Candidate Algorithms Utilization-Unaware (Related/Preliminary Work):» Least Travel Cost (UN-LTC)» Least Location Entropy Priority (UN-LLEP)» Least Accepted First (UN-LAF) Utilization-Aware (Proposed Work):» Least Utilized First (UA-LUF) 24

25 Synthetic Dataset Details Supply-Demand Ratio (SDR) No. consumer requests No. service providers , , , , ,

26 Does the supply-aware provider centric approach improve provider fairness? Fixed parameters: V = 1173, E = 3340, m=120 providers, K=3, t timeout = 2 min, Service rates ϵ [5,15] req/hr, max. travel time ϵ [8, 25] min, max. wait ϵ [10, 25] min, Grid cell length = 2000m Trends: UA-LUF achieved lowest STDEV. Gap increases as SDR increases 26

27 Does the supply-aware provider centric approach improve provider fairness? Fixed parameters: V = 1173, E = 3340, m=120 providers, K=3, t timeout = 2 min, Service rates ϵ [5,15] req/hr, max. travel time ϵ [8, 25] min, max. wait ϵ [10, 25] min, Grid cell length = 2000m Trends: When demand > supply, proposed heuristics and LTC achieve high average. At supply demand, proposed UA-LUF greatly outperform related/preliminary work. 28% higher than UN-LAF at SDR=1 and 40% higher at larger SDR 27

28 Does the provider-focused approach reduce business volume? Fixed parameters: V = 1173, E = 3340, m=120 providers, K=3, t timeout = 2 min, Service rates ϵ [5,15] req/hr, max. travel time ϵ [8, 25] min, max. wait ϵ [10, 25] min, Grid cell length = 2000m Trends: As SDR increases, % matched requests increases Proposed heuristics provided comparable business volume Even outperformed related/prelim work for supply demand. 28

29 Does a provider centric approach increase % of matched providers? Fixed parameters: V = 1173, E = 3340, m=120 providers, K=3, t timeout = 2 min, Service rates ϵ [5,15] req/hr, max. travel time ϵ [8, 25] min, max. wait ϵ [10, 25] min, Grid cell length = 2000m Trends: For demand supply, all algorithms can engage 100% of providers For supply> demand, only provider centric heuristics engage 100% providers. 29

30 Experimental Goals: Self Analysis How does the proposed approach scale with the number of consumers? Does consumer-prioritization improve the business volume? How do the (online) proposed approach compare with offline optimizers? Candidate Algorithms: Proposed Algorithms (online):» Least Utilized First Fewer Candidate First (UA-LUF-FCF)» Least Utilized First Conflict Aware Prioritization (UA-LUF-CAP)» Least Utilized First No Consumer Prioritization (UA-LUF-NoSort) Offline Integer Programming Approaches: OIPMaxReq: Maximize no. matched requests OIPMaxProv : Maximize no. matched requests OIPMinSTD: Minimize standard deviation of provider utilization 30

31 How does the proposed approach scale with the number of consumers? Fixed parameters: V = 1173, E = 3340, m=120 providers, K=3, t timeout = 2 min, SDR=1, max. travel time ϵ [8, 25] min, max. wait ϵ [10, 25] min Heuristic n ,000 10,000 UA-LUF-FCF , ,428.9 UA-LUF-CAP , ,347.5 Cost Model: O(n E log V + n m log m) Trend: Query response time grows linearly with average queue length 31

32 Does consumer-prioritization improve the business volume? Fixed parameters: V = 1173, E = 3340, m=120 providers, K=3, t timeout = 2 min, Service rates ϵ [5,15] req/hr, max. travel time ϵ [8, 25] min, max. wait ϵ [10, 25] min Trends: Consumer prioritization based strategies result in increase in % matched requests. Effect is most significant at balanced supply and demand 32

33 Evaluating solution quality compared to offline Integer Programming: Synthetic Dataset Details Supply-Demand Ratio (SDR) No. consumer requests No. service providers

34 How do the proposed online heuristics compare with offline optimizers? (% Matched Requests) Fixed parameters: m = 5 providers, K=1, t timeout = 0 min, Service rates ϵ [4,8] req/hr, max. travel time ϵ [20, 30] min, max. wait ϵ [20, 30] min, 60 time steps Trends: When demand supply proposed algorithms achieved within 4.7% to 6.7% of the offline upper bound. When supply> demand, algorithms achieved the offline upper bounds 34

35 How do the proposed online heuristics compare with offline optimizers? (STDEV of Provider Utilization) Fixed parameters: m= 5 providers, K=1, t timeout = 0 min, Service rates ϵ [4,8] req/hr, max. travel time ϵ [20, 30] min, max. wait ϵ [20, 30] min, 60 time steps. Trends: When demand > supply, UA-LUF achieves within 1.34x to 2.6x σ off Using Least Local Supply-Demand Ratio as a tie-breaker can reduce it to 1.7xσ For demand supply, gap decreases and all algorithms are 1.39x σ off When supply>demand, UA-LUF achieves 1.3x σ off (tie breaker can achieve bound) 35

36 How do the proposed online heuristics compare with offline optimizers? (% Matched Providers) Fixed parameters: m=5 providers, K=1, t timeout = 0 min, Service rates ϵ [4,8] req/hr, max. travel time ϵ [20, 30] min, max. wait ϵ [20, 30] min, 60 time steps. Trends: For all values of SDR, proposed algorithms engaged 100% of providers (i.e. the offline upper bounds). 36

37 Summary of Experimental Results Proposed approach consistently achieved lowest utilization variance while matching all providers even when supply >> demand. Higher average utilization of lowest 10% of providers when supply demand. Achieved larger matching size when supply demand. Comparison with offline integer programming solution show that, for our experiments involving 5 providers and 60 time steps: Achieved within 4.7% to 6.7% of the offline matching size Matched same no. providers. Achieved within 1.3x to 2.6x offline stdev. Tie breaker heuristic allows achieving lower bound when supply > demand. Trends: Query response time grows linearly with avg. queue length

38 Conclusion and Future Work We formulated the problem of On-demand Spatial Service Propositions. We proposed: A Provider-centric supply-aware heuristic (Least Utilized First) A consumer-priority based greedy matching approach Our experiments using synthetic datasets with real-world characteristics show that the proposed approach outperforms preliminary/related work on multiple performance measures: utilization variance, matching size and average utilization of lowest 10% providers when supply demand. In the future we plan to: Scale-up proposed approach (road network size, no. providers) Investigate a stochastic dynamic programming approach with assumptions on data distribution Extending the OSSP problem to allow mobile service providers 38

39 References (1/2) 1. Y. Huang, et al., Large scale real-time ridesharing with service guarantee on road net- works, Proc. of the VLDB Endowment 7 (14) (2014) B. Cici, A. Markopoulou, N. Laoutaris, Designing an on-line ride-sharing system, in: Proceedings of the 23rd SIGSPATIAL International Conference on Advances in Geographic Information Systems, SIGSPATIAL 15, ACM, New York, NY, USA, 2015, pp. 60:1 60:4. 3. M. Ota, et al., A scalable approach for data-driven taxi ride-sharing simulation, in: IEEE Intl. Conf. on Big Data, IEEE, 2015, pp N. A. Agatz, et al., Dynamic ride-sharing: A simulation study in metro atlanta, Transportation Research Part B: Methodological 45 (9) (2011) P. M. d Orey, M. Ferreira, Can ride-sharing become attractive? a case study of taxisharing employing a simulation modelling approach, IET Intelligent Transport Systems 9 (2) (2014) M. Mahmoudi, X. Zhou, Finding optimal solutions for vehicle routing problem with pickup and delivery services with time windows: A dynamic programming approach based on state space time network representations, Transportation Research Part B: Methodological 89 (2016) L. Kazemi, C. Shahabi, Geocrowd: enabling query answering with spatial crowdsourcing, in: Proc. of the 20th SIGSPATIAL Intl. Conference on Advances in Geographic Information Systems, ACM, 2012, pp L. Kazemi, et al., Geotrucrowd: trustworthy query answering with spatial crowdsourcing, in: Proc. of the 21st SIGSPATIAL Intl. Conf. on Advances in Geographic Information Systems, ACM, 2013, pp D. Deng, C. Shahabi, L. Zhu, Task matching and scheduling for multiple workers in spatial crowdsourcing, in: Proceedings of the 23rd SIGSPATIAL International Conference on Advances in Geographic Information Systems, SIGSPATIAL 15, ACM, New York, NY, USA, 2015, pp. 21:1 21: H. To, C. Shahabi, L. Kazemi, A server-assigned spatial crowdsourcing framework, ACM Trans. Spatial Algorithms Syst. 1 (1) (2015) 2:1 2:28. 39

40 References (2/2) 11. C. Colby and K. Bell. Harvard Business Review. The On-Demand Economy Is Growing, and Not Just for the Young and Wealthy. Apr Spinlister StyleBee TaskRabbit. 40

41 Related Publications Related publications: 1) Ali, R. Y., Eftelioglu, E., Shekhar, S., Athavale, S., & Marsman, E. (2016, October). Supply-demand ratio and on-demand spatial service brokers: a summary of results. In Proceedings of the 9th ACM SIGSPATIAL International Workshop on Computational Transportation Science (pp. 7-12). ACM. 2) Ali, R. Y., Li, Y., Shekhar, S., Athavale, S., & Marsman, E. (2017, November). Supply and demand aware synthetic data generation for on-demand traffic with real-world characteristics. In Proceedings of the 10th ACM SIGSPATIAL International Workshop on Computational Transportation Science. ACM. 3) Ali, R. Y., Shekhar, S., Athavale, S., & Marsman, E. Supply-demand ratio and ondemand spatial service brokers. Under review in the Future Generation Computer Systems Journal Special Issue on Scheduling Algorithms for Cyber-Physical-Social Systems. 41

42 Thank you.