L2M NSERC - Intelligent system for classifying imbalanced data based on three-way Bayesian confirmation

不平衡数据是指数据点不均匀分布在不同类中的数据集。结果，多数类占据了数据点的高比例，少数类占据了剩余的低比例。数据不平衡在实际情况中很普遍，特别是当我们试图检测到一些异常情况时，例如欺诈性交易、垃圾电子邮件和某些疾病。标准分类模型可能无法很好地处理不平衡数据。例如，考虑一个不平衡的数据集，其中多数类占90%的数据点，少数类占10%。一个标准的分类模型可能会了解到，它可以在没有任何条件的情况下简单地预测大多数类别，并且对于90%的情况是正确的。这绝对不能解释分类的内在原因。相比之下，在大多数情况下，很难准确预测少数族裔。然而，少数类通常由欺诈性交易、垃圾邮件和我们实际想要学习和检测的疾病案例组成，比多数类更重要。我们应用贝叶斯确认理论构建了一个智能系统，以学习对不平衡数据进行分类的有效特征和规则。通过比较某一特征或一组特征在观察到该特征的值之前的先验概率(S)和在观察值之后的后验概率(S)来评估该特征或一组特征的重要性。这一变化显示了特征(S)对我们分类决策的真实影响，并有助于检测有效的特征。例如，在检测某种疾病时，我们可能会通过不同的测试来测量一组特征。某些特征的值可能会显著增加或降低疾病的概率。因此，可能会要求患者进行相应的测试，以实现有效的诊断。与标准分类模型相比，我们能够学习到检测多数类和少数类的实际有效的特征和规则，解决了在分析不平衡数据时的上述问题。

Imbalanced data refers to a data set where the data points are not evenly distributed across different classes. As a result, there are majority classes taking high proportions of the data points and minority classes taking the remaining low proportions. Imbalanced data is prevalent in practical situations, especially when we try to detect something abnormal such as fraudulent transactions, spam emails, and certain diseases. Standard classification models may not work well with imbalanced data. For example, consider an imbalanced data set where a majority class takes 90% of the data points and a minority class takes 10%. A standard classification model will likely learn that it can simply predict the majority class without any condition and be correct for 90% of the cases. This definitely does not explain the inherent reason for the classifications. In contrast, it is hard to predict a minority class correctly in most cases. However, the minority class usually consists of fraudulent transactions, spam emails, and cases of diseases that we actually want to learn and detect, more importantly than the majority class.We apply the Bayesian confirmation theory to build an intelligent system to learn effective features and rules for classifying imbalanced data. The importance of a certain feature or a group of features is evaluated by comparing the prior probability of a class before observing the value(s) of the feature(s) and the posterior probability after observing the value(s). The change shows the real impact of the feature(s) on our classification decisions and helps detect the effective features. For example, in detecting a certain disease, we may have a group of features measured through different tests. The values of some features may significantly increase or decrease the probability of the disease. Accordingly, the patient may be asked to do the corresponding tests to achieve an effective diagnosis. Compared to the standard classification models, we are able to learn the actual effective features and rules for detecting the majority and minority classes, addressing the aforementioned issues in analyzing imbalanced data.