Causality is getting more and more attention, and its core idea is counterfactual and intervention. However, the current intervention model requires some prior knowledge, and lacks universality. The paper presents a novel solution called Search the Intervention Sample in Sparse Operation Space (SISSOS). SISSOS introduces variational inference and realizes intervention, that’s feature manipulation at the attribute level. SISSOS is for tabular data and uses sparse space to solve attribute coupling. SISSOS is applied to counterfactual and model interpretation in experiments. In the counterfactual experiment, the proposed solution was proven to find the correct causal effect without any prior knowledge. In the model interpretation experiment, a trained time series neural network with high accuracy was proved by the proposed solution to conform to prior knowledge. Compared with the previous method, the proposed method does not require prior knowledge and its intervention effect is better.

因果关系越来越受到关注，其核心思想是反事实和干预。但是，当前的干预模型需要一些先验知识，并且缺乏通用性。本文提出了一种新颖的解决方案，称为“在稀疏操作空间中搜索干预样本”（SISSOS）。SISSOS引入了变分推理并实现了干预，即在属性级别进行功能操纵。SISSOS用于表格数据，并使用稀疏空间来解决属性耦合。SISSOS用于实验中的反事实和模型解释。在反事实实验中，在没有任何先验知识的情况下，所提出的解决方案被证明可以找到正确的因果关系。在模型解释实验中，所提出的解决方案证明了经过训练的时间序列神经网络具有很高的准确性，以符合先验知识。与以前的方法相比，该方法不需要先验知识，干预效果更好。