Perturbation-Based Explanations of Prediction Models
Perturbation-Based Explanations of Prediction Models
复制标题
DOI：
10.1007/978-3-319-90403-0_9
复制
发表时间：
2018-01-01
期刊：
影响因子：
--
通讯作者：
Bohanec, Marko
中科院分区：
文献类型：
--
作者：
Robnik-Sikonja, Marko;Bohanec, Marko
当前对预测模型的算法解释方法的研究可分为两种主要方法：仅限于神经网络的基于梯度的方法和可用于任意预测模型的更一般的基于扰动的方法。我们概述了基于扰动的方法，重点关注最流行的方法（EXPLAIN、IME、LIME）。这些方法支持对单个预测的解释，但也可以将模型作为一个整体进行可视化。我们描述了它们的工作原理、它们如何处理计算复杂性、它们的可视化以及它们的优点和缺点。我们通过公司 B2B 销售预测的实际用例说明了在业务环境中应用解释方法的实际问题和挑战。我们演示如何将解释用作假设分析工具来回答相关业务问题。
Current research into algorithmic explanation methods for predictive models can be divided into two main approaches: gradient-based approaches limited to neural networks and more general perturbation-based approaches which can be used with arbitrary prediction models. We present an overview of perturbation-based approaches, with focus on the most popular methods (EXPLAIN, IME, LIME). These methods support explanation of individual predictions but can also visualize the model as a whole. We describe their working principles, how they handle computational complexity, their visualizations as well as their advantages and disadvantages. We illustrate practical issues and challenges in applying the explanation methodology in a business context on a practical use case of B2B sales forecasting in a company. We demonstrate how explanations can be used as a what-if analysis tool to answer relevant business questions.