Instance Seriation for Prototype Abstraction

In instance-based machine learning, algorithms often suffer from high storage requirements because of the large number of training instances. This can result not only in large computer memory usage and long response time, but also very often in oversensitivity to noise. To tackle such problems, various instance reduction algorithms have been developed that remove noisy and redundant patterns. In this work, we discuss the concept of data seriation and its application on instance-based learning, and introduce a new approach, the Instance Seriation for Prototype Abstraction algorithm (ISPA), which is a data condensation method that generates a new set of prototypes. ISPA is evaluated on 10 datasets and its performance is compared to other successfully established pruning algorithms. Our method exhibited competitive results in terms of classification accuracies and reduction rates.