This paper presents a comprehensive framework for expert-augmented supervised feature selection, addressing pre-processing, in-processing, and post-processing aspects of Explainable Artificial Intelligence (XAI). As part of pre-processing XAI, we introduce the Probabilistic Solution Generator through the Information Fusion (PSGIF) algorithm, leveraging ensemble techniques to enhance the exploration and exploitation capabilities of a Genetic Algorithm (GA). Balancing explainability and prediction accuracy, we formulate two multi-objective optimization models that empower expert(s) to specify a maximum acceptable sacrifice percentage. This approach enhances explainability by reducing the number of selected features and prioritizing those considered more relevant from the domain expert's perspective. This contribution aligns with in-processing XAI, incorporating expert opinions into the feature selection process as a multi-objective problem. Traditional feature selection techniques lack the capability to efficiently search the solution space considering our explainability-focused objective function. To overcome this, we leverage the Genetic Algorithm (GA), a powerful metaheuristic algorithm, optimizing its parameters through Bayesian optimization. For post-processing XAI, we present the Posterior Ensemble Algorithm (PEA), estimating the predictive power of features. PEA enables a nuanced comparison between objective and subjective importance, identifying features as underrated, overrated, or appropriately rated. We evaluate the performance of our proposed GAs on 16 publicly available datasets, focusing on prediction accuracy in a single objective setting. Moreover, we test our multi-objective model on a classification dataset to show the applicability and effectiveness of our framework. Overall, this paper provides a holistic and nuanced approach to explainable feature selection, offering decision-makers a comprehensive understanding of feature importance.

中文翻译：

---

通过专家增强的监督特征选择迈向可解释的人工智能

本文提出了一个专家增强监督特征选择的综合框架，解决了可解释人工智能（XAI）的预处理、处理中和后处理方面的问题。作为 XAI 预处理的一部分，我们通过信息融合 (PSGIF) 算法引入概率解生成器，利用集成技术来增强遗传算法 (GA) 的探索和利用能力。平衡可解释性和预测准确性，我们制定了两个多目标优化模型，使专家能够指定最大可接受的牺牲百分比。这种方法通过减少所选特征的数量并优先考虑从领域专家的角度认为更相关的特征来增强可解释性。这一贡献与正在处理的 XAI 相一致，将专家意见作为多目标问题纳入特征选择过程中。考虑到我们以可解释性为中心的目标函数，传统的特征选择技术缺乏有效搜索解决方案空间的能力。为了克服这个问题，我们利用遗传算法（GA），一种强大的元启发式算法，通过贝叶斯优化来优化其参数。对于后处理 XAI，我们提出了后验集成算法 (PEA)，用于估计特征的预测能力。 PEA 能够对客观重要性和主观重要性进行细致入微的比较，识别被低估、高估或适当评价的特征。我们在 16 个公开可用的数据集上评估了我们提出的 GA 的性能，重点关注单个目标设置中的预测准确性。此外，我们在分类数据集上测试我们的多目标模型，以展示我们框架的适用性和有效性。总体而言，本文提供了一种全面且细致入微的可解释特征选择方法，使决策者能够全面了解特征重要性。