---

Summary

Sample subset optimization (SSO) [1] can be used as an intelligent method for sampling from imbalanced data [2]. It works by ranking and selecting most informative samples from the majority class to form a balanced dataset with samples from minority class.

GUI version

The GUI version SSOgui.jar will run by a double-click assuming you have Java installed in your computer. The dataset exampleData.txt could be used to test the program. When using your own dataset, please follow the way each column and row is defined as in the exampleData.txt. For the example dataset, the instances from 31 to 35 are noise introduced to the majority class but generated by minority class distribution. Therefore, the goal is to give them a fairly low rank when sampling is applied to select from the majority class. SSOgui.jar is straightforward to use and self-explanatory. Let me know if you need further detail on how to use it.

Command line version

To try out the command line version, please going through the following steps:

---

Background Knowledge

Data sampling [3] is a useful procedure when the data to be analyzed is of imbalanced class distribution (samples from one class (referred to as majority class) outnumber samples from another class (referred to as minority class)). This is because the classification algorithms (classifiers) may be biased towards the majority class. Random over sampling (randomly increases the minority class so as to match the majority class) and random under sampling (randomly decrease the majority class so as to match the minority class) are the simplest ways to balance the class distribution. The disadvantage of random over sampling is it duplicates samples in the minority class which may be completely ignored by some classifiers. As to random under sampling, it throws information away! Currently, the most popular and intelligent (as oppose to random) over sampling strategy is SMOTE [4] which produce "new" samples instead of directly duplicate samples from minority class. However, there is no (few) good proposal for intelligent under sampling. Here, we propose an intelligent under sampling approach by using sample subset optimization technique. The usefulness of each sample from majority class are evaluated, and the most useful samples from the majority class are selected to combine with samples from the minority class. This result in a balanced under sampling dataset. The following two figures display the decision boundary of a kNN classifier build on an imbalanced dataset without using SSO-based sampling and with SSO-based sampling.

---

References