New Longitudinal Classification Methods with Automated Detection of Monotonicity Constraints

Longitudinal datasets contain repeated measurements of the same attributes measured at different points in time. They can be used to discover interesting knowledge based on time-based trends and create prediction models that can predict future values of target attributes, like class attributes in the case of the classification task of data mining (or machine learning).

Monotonicity is a special type of relation between attributes in which their values have either a strongly positive or a strongly negative relationship. These relations can occur in real-world data and can be used to enhance the predictive performance and the acceptability of prediction models by users.

In recent years, there has been an increasing interest in using longitudinal data and monotonicity relations in the field of classification, resulting in the emergence of the fields of Longitudinal Classification and Monotonic classification.

While there has been a lot of development in each of these fields separately, there have not been any studies that united these two fields. The aim of this study is to develop new approaches to longitudinal and monotonic classification and develop a unified approach for using monotonicity relations found in longitudinal datasets in order to construct highly accurate and useful classification models.

As a result of this study, several contributions were made to both fields, as follows. First, a novel approach to automated monotonicity detection has been created. Second, a novel longitudinal classification algorithm was developed, based on the idea of nested decision trees. The basic idea is that each internal node of the outer tree consists of a decision tree built using as input all the temporal variations (in different time points) of the same longitudinal attribute. Third, two new methods for deriving monotonicity-based longitudinal attributes were created, and a unified algorithm was created that united all of these features into one monotonic longitudinal classification algorithm, the first algorithm to combine the two approaches.

The proposed methods were evaluated on longitudinal datasets derived from the English Longitudinal Study of Ageing (ELSA), and the results confirmed that the Nested Tree algorithm outperformed a conventional decision tree algorithm regarding predictive performance.