Despite widespread adoption of machine learning throughout industry, many firms face a common challenge: relevant datasets are typically distributed amongst market competitors that are reluctant to share information. Recent works propose data markets to provide monetary incentives for collaborative machine learning, where agents share features with each other and are rewarded based on their contribution to improving the predictions others. These contributions are determined by their relative Shapley value, which is computed by treating features as players and their interactions as a characteristic function game. However, in its standard form, this setup further provides an incentive for agents to replicate their data and act under multiple false identities in order to increase their own revenue and diminish that of others, restricting their use in practice. In this work, we develop a replication-robust data market for supervised learning problems. We adopt Pearl's do-calculus from causal reasoning to refine the characteristic function game by differentiating between observational and interventional conditional probabilities. By doing this, we derive Shapley value-based rewards that are robust to this malicious replication by design, whilst preserving desirable market properties.

翻译：尽管机器学习在工业界得到广泛应用，但许多企业面临一个共同挑战：相关数据集通常分布在市场竞争者之间，而这些竞争者往往不愿共享信息。近期研究提出数据市场机制，通过货币激励促进协作式机器学习，其中参与者相互共享特征，并根据其对改进他人预测的贡献获得奖励。这些贡献通过相对夏普利值确定，该值通过将特征视为参与者、将其交互视为特征函数博弈来计算。然而，在这种标准形式下，该机制进一步激励参与者复制自身数据并以多个虚假身份参与，从而增加自身收益并削弱他人收益，这限制了其实际应用。在本研究中，我们为监督学习问题开发了一种复制鲁棒的数据市场。我们采用因果推理中的珀尔干预演算，通过区分观测条件概率与干预条件概率来改进特征函数博弈。通过这种方法，我们推导出基于夏普利值的奖励机制，该机制在设计上对这种恶意复制行为具有鲁棒性，同时保持了理想的市场特性。