Binary classification with minimum observations is an important task in applications where enough training data are not available. In this paper, we propose a binary classification approach that is based on an unsupervised ranking algorithm for objects with numerical attributes. The ranking algorithm takes normalized attributes of numerical objects as input and returns the weights of attributes. These weights are used to rank the objects. We propose a class labelling algorithm that labels each side of the ranked objects as a class using less than or equal to 15 labeled data objects or observations. Evaluation on six different data sets shows that the proposed approach is comparable to the state-of-the-art binary classification algorithms that use 70 percent observations compared to less than one percent observations (on average) used by the proposed approach.
Article ID: 2022L20
Venue: Canadian Conference on Artificial Intelligence
Publisher: Canadian Artificial Intelligence Association