Social learning with questions
Social learning with questions
复制标题
带着问题进行社交学习
DOI：
--
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
V. Subramanian
中科院分区：
文献类型：
--
作者：
G. Schoenebeck;Shih;V. Subramanian
在社交网络中，参与者利用以下信息来指导自己的决策：（a）他们所拥有的私人信号（知识）；（b）对过往参与者行为的了解（观察）；（c）来自可联系的有经验参与者（专家）的评论。当参与者在观察历史后为了决策而忽略其私人信号是最优选择时，在信号的似然比有界且历史完全可观测的情况下，几乎必然会发生信息级联。尽管这对个体而言是最优的，但可能会导致社会次优的结果。研究社会学习（即做出社会最优决策）的文献主要关注贝叶斯理性参与者利用上述（a）和（b）渠道，要么放宽信号强度有界的假设，要么允许历史被扭曲。在这项工作中，我们设定有限的通信能力，让贝叶斯理性参与者询问其前辈，这是受到人们在做决策前通常会咨询几个朋友这一现实行为的启发。我们允许每个贝叶斯理性参与者向一部分前辈中的每个人提出一个单一的、私人的且容量有限（用于回复）的问题。请注意，最大后验概率（MAP）规则对个体而言仍然是最优的，并且每个参与者都会将其用于自己的决策。有了给定的通信能力，我们想回答以下两个问题：1）什么是合适的框架来模拟问题对信息聚合所提供的帮助？2）我们能否构建一组问题，通过以最低的容量要求（以比特为单位）询问最少的参与者来实现学习？
In social networks, agents use information from (a) private signals (knowledge) they have, (b) learning past agents actions (observations), and (c) comments from contactable experienced agents (experts) to guide their own decisions. With fully observable history and bounded likelihood ratio of signal, Information Cascade occurs almost surely when it is optimal for agents to ignore their private signals for decision making after observing the history. Though individually optimal, this may lead to a socially sub-optimal outcome. Literature studying social learning, i.e., making socially optimal decisions, is mainly focused on using channels (a) and (b) above for Bayes-rational agents by either relaxing the assumption of bounded signal strength or allowing the distortion of the history. In this work, we enable a limited communication capacity to let Bayes-rational agents querying their predecessors, motivated by the real-world behavior that people usually consult several friends before making decisions. We allow each Bayes-rational agent to ask a single, private and finite-capacity (for response) question of each among a subset of predecessors. Note that the Maximum Aposteriori Probability (MAP) rule is still individually optimally and will be used by each agent for her decision. With an endowed communication capacity, we want to answer the following two questions: 1) What is the suitable framework to model the help that questions provide on information aggregation? 2) Can we construct a set of questions that will achieve learning by querying the minimum set of agents with the minimum capacity requirements (in terms of bits)?