1. Supply-Demand Ratio and On-Demand Spatial Service Brokers
A collaboration between Ford University Research Program and University of Minnesota
University PI: Shashi Shekhar
Ford PI:
Shounak Athavale
Eric Marsman

2. Outline IWCTS Slide Review Date
Extensions for the ACM Transactions on Intelligent Systems and Technology (ACM TIST) - Special Issue on Urban Intelligence

3. Extensions for ACM TIST
IWCTS 2016
ACM TIST 2016
Problem Definition
Objective: max. matched requests
Consumers had 2 constraints:
Max. travel distance
Max waiting time
Proof of NP-hardness
Add a “Basic Concepts” subsection
Modify consumer constraints
Formalize a measure of fairness
Proposed Approach
Proposed 2 service-provider centric heuristics:
Least Accepted First
Least appearance as Candidate First
Incorporating ratings
Supply-demand ratio-aware broker (moving providers?)
Add example for each different heuristic
Experimental Evaluation
Four experiments:
% matched requests
% matched providers
avg. requests/provider
Stdev of requests/provider
Expand experimental evaluation section by adding more experiments.
from original paper
tie-breaker heuristics
Discussion
None
Add discussion section (further related work, choice of simulation framework)
Special Issue: Urban Intelligence - ACM TIST.  

4. Basic Concepts
A Service Provider: A provider registered in the system is defined using its location and service rate per hour over the day (e.g. 15 requests per hour)
A Consumer Request: A request for from a mobile consumer, including the consumer’s current location, max. acceptable travel distance and max. acceptable waiting time before service.
A Service Provider Proposition: A quadruple (r, p, d, w) where:
r ϵ set of available consumer requests
p ϵ set of registered service providers
d: distance between r and p
w: waiting time before r is served by p
Example: (C1, P1, 2 miles, 5 min)

5. Problem Definition
Add computational complexity subsection with proof of NP-hardness
Modify consumer constraints
Maximum travel time vs. travel distance (longer distance on a freeway can be faster)
Maximum waiting time at provider
Formalize a measure of fairness:
Previously communicated informally as “Keeping the eco-system functioning by engaging many service providers and balancing their assignments as evenly as possible”
Need to propose a formal measure

6. Problem Definition (cont’d)
Provider Balance Score:
A weighted sum to:
Maximize ratio of matched providers
Minimize STDEV of number of assigned requests per provider (normalize)
Used as a secondary objective function (tie-breaker)
Another alternative:
Keep a single objective (maximizing number of matched requests), but model providers leaving the system if not matched. (unbalanced assignment → less matched requests)
Drawbacks:
Makes more sense only in case of moving providers i.e. problem changes to moving providers and fixed consumers. (a reviewer’s comment)
No need for several propositions since for moving providers, only a single assignment is enough.
Need to simulate a long duration for providers to shift in space or leave system

7. Problem Definition: On-Demand Spatial Service Propositions
Input:
A set P of service providers
A set R of consumer requests arriving dynamically
A number of required propositions K
Output: K service provider(s) propositions for each request
Objective: Maximize number of matched requests
Secondary Objective: Maximize provider-balance score
Constraints:
Each returned proposition satisfies the consumer’s max. travel time and waiting time constraints and does not violate the provider’s service rate.

8. Proposed Approach (1/?): Incorporating Ratings
Reviewer Comment:
“clearly LAF results in higher % matched providers as SDR increases, by its very definition, but this alone does not make it a good choice as a heuristic, as it doesn't consider the possibility that certain providers could be "least accepted" for a reason (e.g., bad reputation, historically-bad service, etc.). Forcing consumers to choose from commonly-unaccepted options solely for the sake of "balance" seems unlikely to be acceptable in certain "spatial service" use cases (such as the restaurant scenario considered).”
To avoid bias towards low-rating providers:
Define an “Average Consumer Loss” function (ACL):
Add ACL to the Provider Balance Score:

9. Proposed Approach (2/?): Incorporating Ratings
Approaches to compare:
For least assigned providers, select highest-rated K candidates
Challenge: Are we penalizing new providers?
For least assigned providers, diversify selected ratings
Challenge: May penalize consumers by adding low ratings
For least assigned providers, return when average ratings of propositions ≥ avg. ratings of candidates

10. Interaction between price modeling, incentives, moving providers
Proposed Approach (3/?)
Reviewer Comments:
“The technical contribution is limited to a number of heuristics and a rather straightforward evaluation framework.”
“It is unclear why the consumer is travelling in this application framework and not the provider. e.g. for most ridesharing applications, the providers are travelling and the consumers location is fixed (for pickup).”
“It is assumed that all providers equally charge for their services; in practice, this is not the case. One could argue that the distance from consumer’s current position to the location of the service provider captures to some extend the consuming cost but not entirely. For example, if the providers are restaurants or supermarkets, a person might decide to travel a bit more instead of paying extra for her meal or her grocery.”
Interaction between price modeling, incentives, moving providers

11. A Supply-Demand Ratio Aware Broker for On-demand Services (4/?)
Motivation: Supply-demand ratio exhibits spatio-temporal heterogeneity
Balanced Zone: A zone where Demand ≈ Supply
Jammed Zone: A zone where Demand >> Supply
Sparse Zone: A zone where Supply >> Demand
A zone type can also change over time: a moving horizon is used to infer zone type as time changes
Propose a supply-demand ratio-aware broker that changes matching strategies in space and time based on local supply-demand ratios.
A separate queue is maintained for each zone and matching strategy is selected based on zone type.
Balanced zone: focus on making consumers happier (since we should be able to handle most requests)
Apply Consumer-centric heuristics (e.g. NN)

12. A Supply-Demand Ratio Aware Broker for On-demand Services (5/?)
Jammed zone: focus is maximizing matched requests
Apply a broker-centric heuristic (i.e. highest matching heuristic)
Incentivize consumers:
Model hurried vs. relaxed vs. indifferent consumers
Suggest longer waiting/travel time to consumers (maybe with lower prices) to shift demand to sparser zones (relaxed)
Suggest higher pricing to consumers to get service now (hurried) in case of mobile providers
Incentivize (mobile) providers with recently few matches to move to current zone for a higher price.
Provider cost model: price > cost of travel to current zone + trip cost
Sparse zone: main focus is keeping eco-system alive until demand rises again
Apply provider-centric heuristics
Mobile providers are more realistic when discussing keeping the eco-system alive.
Incentivize (mobile) providers:
Suggest higher price for providers to move to jammed zones

13. A Supply-Demand Ratio Aware Broker for On-demand Services (6/
A Supply-Demand Ratio Aware Broker for On-demand Services (6/?): Requests crossing zone boundaries
Challenges:
Price modeling
Sharing of ride for a number of consumers
How to learn the boundary values between different zone types?
Need to model (mobile) providers leaving the system to influence matched requests:
e.g. if cost of travel time since last ride > expected price of next ride
Two types of requests:
local: max travel distance extends only inside the origin zone
focal: max travel distance extends beyond origin zone
For focal requests, may sort candidates in neighbor zones based on their zone types or current number of real-time requests.

14. Experimental Evaluation
Extend experimental evaluation section to include:
Evaluating effect of supply-demand ratio on other metrics:
Avg. consumer waiting time,
Avg. consumer travel time
Total execution time
Evaluating the effect of varying other parameters: K
ttimeout
Service times, consumer’s maximum waiting time and travel time constraints
Neighborhood size for LLEP
Combining heuristics to solve ties and enhance matching quality
Evaluate other proposed work

15. Thank you.