Endorsement in Referral Networks.
EUMAS(2018)
摘要
Referral networks is an emerging research area in the intersection of Active Learning and Multi-Agent Systems where experts—humans or automated agents—can redirect difficult instances (tasks or queries) to appropriate colleagues. Learning-to-refer involves estimating topic-conditioned skills of colleagues connected through a referral network for effective referrals. Proactive skill posting is a learning setting where experts are allowed a one-time local network advertisement of a subset of their top skills. The learning challenge is exploiting partially available (potentially noisy) self-skill estimates, including adversarial strategic lying to attract unwarranted referrals. In this paper, we introduce the notion of endorsement typically found in professional networks where one colleague endorses another on particular topic(s). We first augment proactive skill posting with endorsements and propose modifications to existing algorithms to take advantage of such endorsements, penalizing subsequent referrals to agents with bogus skill reporting. Our results indicate that truthful endorsements improve performance as they act as an additional cushion to early failures of strong experts. When combined with truthful endorsements, extensive empirical evaluations indicate performance improvement in proactive-DIEL and (epsilon )-Greedy in both market-aware and market-agnostic skill posting setting while retaining desirable properties like tolerance to noisy self-skill estimates and strategic lying.
更多查看译文
关键词
Active Learning, Referral networks, Proactive skill posting
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要