A campus-scale mobile crowd-tasking platform

Hoong Chuin Lau, Nikita Jaiman, Randy Tandriansyah, Thivya Kandappu, Archan MisraUbicomp Adjunct 2016 · DOI

TL;DR

This paper introduces TA$Ker, a mobile crowd-tasking platform designed to address the challenges of opportunistic mobile crowdsourcing. By employing a centrally-coordinated approach that considers worker movement patterns and behavioral dynamics, TA$Ker enables more efficient and effective task completion on a large scale.

What Problem Does This Paper Address?

Mobile crowdsourcing, which leverages a decentralized pool of mobile users to perform location-specific tasks, holds significant promise for urban operations like logistics and monitoring. However, existing models often rely on a purely opportunistic approach, leading to uncertainties in task completion rates and worker retention. This paper tackles the problem of making mobile crowdsourcing more reliable and efficient by moving towards a centrally-coordinated system. The authors highlight the need to account for predicted worker movement trajectories and human behavioral responses to incentives and deadlines, moving beyond purely rational agent models often used in synthetic studies.

Methods

The core of the paper is the introduction and evaluation of the TA$Ker platform. TA$Ker consists of a mobile app for workers and a backend system with several key components:

  • Route Predictor: Uses historical movement traces to predict individual user trajectories.
  • Task Recommender: Suggests tasks that best align with a worker's predicted path, minimizing detour overhead.
  • Task Management Portal: Allows administrators to manage available tasks.
  • Bundle Engine: Groups nearby tasks into bundles, requiring completion of all tasks for payment.
  • Incentive Engine: Determines task rewards based on factors like location popularity.

The platform was deployed on a university campus and used to conduct experiments on:

  • Task Bundling: Comparing worker productivity with bundled vs. individual tasks.
  • Differential Pricing: Investigating inverse density-based pricing strategies.
  • Cheating: Analyzing the factors influencing task reporting without being at the location.
  • Dynamic Collaboration (Task Referral): Evaluating the impact of peer-to-peer task referrals.

Datasets

The paper describes experiments conducted on a campus-scale deployment of TA$Ker.

  • Initial study: 80 student participants performing ~1000 tasks over 4 weeks.
  • Main study: 900 undergraduate students performing 30,000+ tasks over 2 months. The data collected includes task completion reports, worker movement traces, and responses to various experimental conditions. No specific external datasets are mentioned; the data is generated from the TA$Ker platform itself.

Key Insights and Ideas

  • Workers prefer bundled tasks: Bundling tasks significantly improves worker productivity, suggesting that workers find it more efficient and potentially more rewarding to complete a set of related tasks together.
  • Differential pricing can mitigate spatial skews: Inverse density-based pricing, where rewards are inversely proportional to location popularity, effectively counters uneven task completion across different areas and promotes fairer reward distribution.
  • Cheating is influenced by both intrinsic and contextual factors: A worker's tendency to cheat is not solely due to personality but also depends on task properties, such as tighter deadlines and the position of a task within a bundle.
  • Dynamic collaboration boosts efficiency: The introduction of a task referral feature among peer crowd-workers demonstrably increases the overall task completion rate.

Key Results

  • Task Bundling: Worker productivity increased by 77% when performing bundled tasks compared to individual tasks, earning 20% more per minute of additional detour.
  • Differential Pricing: A popularity-based differential pricing mechanism reduced the variation in spatial skewness of task completion rate by 5-fold.
  • Cheating:
    • When performing the first two tasks of a bundle, most responses were received within 2 minutes of the task location. For later tasks in a bundle, this distance increased to 6 minutes.
  • Tasks with tighter time windows (less than 90 minutes validity) led to disproportionately more cheating (80% of cheating occurred on these tasks).
  • Dynamic Peer Referrals: The task referral feature increased the overall task completion rate by 14%.
  • Smart Campus Sensing: The platform demonstrated its ability to collect timely data for campus operations, e.g., obtaining restroom reports with 90% latency of 30 minutes or lower, and rubbish bin reports with 100% latency of 1 hour or lower.

Limitations and Future Work

  • Limitations: The study was conducted on a university campus, which may not fully represent the complexities of a general urban environment. The authors acknowledge that their model of worker behavior is based on observed patterns and may not capture all nuances of human decision-making.
  • Future Work: The paper suggests further exploration of how to incorporate social factors into per-user task preference models, which resulted in a significant 31% improvement in prediction accuracy. The authors also highlight the potential of TA$Ker for developing smart campus applications, particularly in improving operational efficiency for services like cleaning staff. Further research could explore more sophisticated incentive mechanisms and task allocation strategies.

Summary updated

BibTeX
@inproceedings{lau2016campus,
  author = {Hoong Chuin Lau and Nikita Jaiman and Randy Tandriansyah and Thivya Kandappu and Archan Misra},
  title = {{A campus-scale mobile crowd-tasking platform}},
  booktitle = {Ubicomp Adjunct},
  year = {2016},
  doi = {10.1145/2968219.2971388},
}
Figure