[Paper Review] Exploration vs. Exploitation in Team Formation

Ramesh Johari, Vijay Kamble+2 more|arXiv (Cornell University)|Sep 18, 2018

Mobile Crowdsensing and Crowdsourcing5 references3 citations

TL;DR

This paper studies the exploration-exploitation trade-off in team formation on online labor platforms, modeling team performance as either the weakest or strongest member's quality. It proposes 1-chain and 2-chain algorithms for adaptive team matching, showing that the optimal strategy reverses at p=0.5: 1-chain outperforms for p<0.5, while 2-chain dominates for p>0.5, due to differing regret structures in low- and high-skill regimes.

ABSTRACT

An online labor platform faces an online learning problem in matching workers with jobs and using the performance on these jobs to create better future matches. This learning problem is complicated by the rise of complex tasks on these platforms, such as web development and product design, that require a team of workers to complete. The success of a job is now a function of the skills and contributions of all workers involved, which may be unknown to both the platform and the client who posted the job. These team matchings result in a structured correlation between what is known about the individuals and this information can be utilized to create better future matches. We analyze two natural settings where the performance of a team is dictated by its strongest and its weakest member, respectively. We find that both problems pose an exploration-exploitation tradeoff between learning the performance of untested teams and repeating previously tested teams that resulted in a good performance. We establish fundamental regret bounds and design near-optimal algorithms that uncover several insights into these tradeoffs.

Motivation & Objective

• To address the challenge of matching teams of unknown-ability workers to jobs on online labor platforms.
• To model how team performance depends on constituent worker quality, under two payoff structures: weakest link and strongest link.
• To design adaptive matching algorithms that minimize regret while learning worker types through team outcomes.
• To analyze the fundamental trade-offs between exploration (learning unknown teams) and exploitation (reusing known good teams).
• To derive regret bounds and identify regime-dependent optimal strategies for team formation.

Proposed method

• Models a large pool of workers, each independently high (1) or low (0) quality with probability p.
• Assumes teams of size two, with team payoff determined by min (weakest link) or max (strongest link) of member types.
• Observes only aggregate team performance (1 or 0), not individual worker types, requiring inference over time.
• Proposes two adaptive algorithms: 1-chain and 2-chain, which asynchronously explore matches for known low-quality pairs.
• Uses epoch-based exploration: 1-chain tests one pair at a time, while 2-chain tests two pairs simultaneously to increase information gain.
• Analyzes regret by comparing cumulative payoff loss against an oracle that knows all worker types.

Experimental results

Research questions

• RQ1 How does the optimal balance between exploration and exploitation in team formation depend on the team performance aggregation rule (weakest vs. strongest link)?
• RQ2 What is the fundamental regret bound for team formation under each payoff model?
• RQ3 How do different adaptive matching strategies (1-chain vs. 2-chain) compare in performance across varying worker quality distributions?
• RQ4 At what value of p (probability of high-quality workers) does the relative performance of 1-chain and 2-chain algorithms reverse?
• RQ5 Can simple, myopic strategies be optimal in certain regimes, and if so, under what conditions?

Key findings

• For p < 0.5, the 1-chain algorithm incurs lower regret than the 2-chain algorithm, as it accumulates fewer low-quality matches.
• For p > 0.5, the 2-chain algorithm incurs lower regret than the 1-chain algorithm, due to its ability to reduce low-quality pairings more effectively.
• The 1-chain algorithm consistently results in fewer matches between two high-quality workers across all p ∈ [0,1].
• The 2-chain algorithm results in fewer matches between two low-quality workers than the 1-chain algorithm when p > 0.5.
• The reversal in performance at p = 0.5 is due to a discontinuity in the regret characterization, not in the algorithms’ intrinsic behavior.
• Myopic strategies can be highly suboptimal in some regimes but effective in others, depending on the underlying worker quality distribution.