Bilkent University
Department of Computer Engineering
CS 690 SEMINAR
Online Task Allocation in Crowdsourced Delivery
Fuat Basık
PhD Student
Computer Engineering Department
Bilkent University
For various businesses, crowdsourced delivery is an answer to the growing expectations of customers for faster and cost efficient delivery. These businesses include but not limited to online shopping, on demand local delivery, and on demand transportation. Different than the traditional services, power of crowdsourced delivery is based on the network effect. On the other hand, to engage this powerful network, workers should be attracted by potential income. This situation leads crowdsourced platforms into chicken and egg problem. Therefore, to avoid worker churn, any crowdsourcing platform should try to maximize the quality of service provided to its’ workers as well as to its’ customers. Borrowing from the social psychology literature, fairness is key for crowdsourcing platforms for ensuring participation. However, well known offline assignment solutions felt short on modeling this dynamic fairness metric. They also do not work for online scenarios. In this work, we introduce a new crowdsourced delivery platform that takes fairness among workers into consideration as well as maximum task assignment objective. Since reduntant task assignment is not possible in delivery tasks, we first introduce 2-phase assignment model that maximizes reliability of a worker to complete the given task. This 2-phase model contributes to the taxonomy of crowdsourcing applications as well. To realize effectiveness of our model in practice, we present both offline and online versions of our proposed algorithm F-Aware. Given the task to worker bipartite graph, F-Aware assigns each task to a worker who maximizes fairness locally. We evaluated our work with respect to runtime, task completion ratio(TCR), fairness and assignment ratio(AR). Experiments showed that F-Aware, runs 106 × faster than TCR-optimal solution and assigns 97 % of the tasks that can be assigned by it. Moreover, F-Aware assigns 10 % more tasks and 2 × more fair than the two other baseline algorithms.
DATE: 10 October, 2016, Monday @ 16:40
PLACE: EA-409