Synthetic Classification Datasets for Metamorphic Testing of Machine Learning Models

Description This collection gathers multiple synthetic classification datasets generated using the make_classification function from the scikit-learn library.The datasets were designed to support the training and evaluation of machine learning classification models under a wide variety of controlled conditions, and were specifically used to assess model behavior through multiple metamorphic relations. The collection includes datasets with varying: numbers of classes (from 2 to 5); numbers of samples and features; noise levels (no noise, 10% and 60% of noisy samples); distribution (homogeneous, and heterogeneous). Dataset Structure Each dataset contains at least the following files: original.csv: Coordinates of the data points (feature values). class.csv: Class labels associated with each data point. config.csv: Metadata indicating which attributes are contextual and which are behavioral, as used in the metamorphic testing framework. Some datasets additionally include the following directories: MR_points/: CSV files containing points generated for metamorphic relations applied to the full dataset, including deliberately generated outliers. splited_convexes/: CSV files containing points produced for three convex-based metamorphic relations, applied after splitting the original dataset into smaller subsets. Two splitting strategies are provided: chained and independent. Special Datasets 4-classes_1000x2A specific 2D dataset composed of four classes arranged into four symmetrically placed clusters.The clusters contain identical point distributions, modulo geometric transformations applied to enforce symmetry. Small intersection-focused datasets: 3_classes_close 3_classes_far 5_classes_4_superposed 5_classes_close 5_classes_far These datasets were intentionally constructed with manually positioned clusters in order to produce intersections of different shapes and complexities, enabling fine-grained analysis of classifier behavior in overlapping decision regions. Related Publications This dataset collection has been used in previously published research as well as in a doctoral dissertation, and continues to support ongoing work. Conference paper: Colonval, J., & Bouquet, F. (2025). New Convex-Based Metamorphic Relations and Large-Scale Machine Learning Model Evaluation. In Lecture Notes in Computer Science, Springer.DOI: https://doi.org/10.1007/978-3-032-05188-2_23 Doctoral thesis: Colonval, J. (2024). Validation de composants d'intelligence artificielle issus d'apprentissages sur la base de détection d'aberration et de tests métamorphiques. PhD Thesis.Available at: https://theses.hal.science/tel-05016629v1/ Ongoing work:An additional research article based on these datasets is currently under preparation.### 数据集描述 本数据集集合汇集了多个使用scikit-learn库中`make_classification`函数生成的合成分类数据集。此类数据集旨在支撑多种可控条件下的机器学习分类模型训练与评估工作，并专门用于通过多种**变质关系（metamorphic relations）**评估模型行为。 本集合涵盖了以下多维度可变参数的数据集： 1. 类别数量（2至5类） 2. 样本与特征数量 3. 噪声水平（无噪声、10%噪声样本与60%噪声样本） 4. 数据分布（同分布与异分布） ## 数据集结构 每个数据集至少包含以下文件： - `original.csv`：数据点的坐标（即特征值） - `class.csv`：各数据点对应的类别标签 - `config.csv`：元数据文件，用于标注变质测试框架中所使用的上下文属性与行为属性 部分数据集还额外包含以下目录： - `MR_points/`：存储用于对完整数据集应用变质关系时生成的点的CSV文件，其中包含刻意生成的异常值 - `splited_convexes/`：存储针对三种基于凸性的变质关系生成的点的CSV文件。此类关系需先将原始数据集划分为若干小子集后再应用，共提供两种划分策略：链式划分与独立划分 ## 专用数据集 ### 4-classes_1000x2A 一款特定的二维数据集，包含4个类别，对应4个对称排布的簇。各簇的点分布完全一致，仅通过施加几何变换以实现对称排布。 ### 聚焦交叉分布的小型数据集 包含`3_classes_close`、`3_classes_far`、`5_classes_4_superposed`、`5_classes_close`、`5_classes_far`。 上述数据集均通过手动设置簇的位置构建，旨在生成不同形状与复杂度的交叉区域，从而支持对分类器在重叠决策区域中的行为进行精细化分析。 ## 相关出版物 本数据集集合已被应用于已发表的研究与一篇博士学位论文中，并将持续支撑当前正在开展的研究工作。 ### 会议论文 Colonval, J. 与 Bouquet, F. (2025). 新型基于凸性的变质关系与大规模机器学习模型评估. 收录于《计算机科学讲义》（Lecture Notes in Computer Science），Springer出版社.DOI: https://doi.org/10.1007/978-3-032-05188-2_23 ### 博士学位论文 Colonval, J. (2024). 基于异常检测与变质测试的人工智能学习组件验证. 博士学位论文. 可访问地址：https://theses.hal.science/tel-05016629v1/ ### 在研工作 基于本数据集集合的另一篇研究论文目前正在筹备中。