Dynamical Learning in non-stationary environments

Pattern Recognition (PR) covers all algorithms and methods for classifying patterns into classes. A shape is a simplified observation of a system's state, represented by a set of parameters. These parameters form the representation space that characterizes the system's operation or behavior. Each class is associated with a state or mode characterizing the system's behavior or operating conditions. The classification of a shape is performed by a classification module, called a classifier, using a rule or decision function. The problem of RoF is therefore to characterize the model and label of each class associated with a shape. This requires the use of classification techniques that enable similar shapes to be grouped together by learning.

Classes can be static or dynamic. Static classes are based on stationary data. This means that the parameters and model structure of each class remain unchanged over time. However, as the environment is constantly changing, most real-world data is non-stationary. Non-stationarity is defined as an evolution over time of the parameters and structure of the model or classifier due, for example, to wear and tear or aging of the machine, or to temporal changes in the characteristics of the system's environment. This non-stationarity modifies the static or dynamic behavior of the system, affecting the performance (predictive accuracy) of its model or classifier. Classes in this case are dynamic. The parameters and structure of the model or classifier must be updated to maintain its performance over time.

In this paper, the problem of learning models or classifiers in a non-stationary environment is studied. Suitable methods and techniques for learning effective classifiers for this type of environment are presented and analyzed. Predicting the occurrence of faults, filtering information (documents, e-mails), classifying documents and modeling user profiles and monitoring their evolution over time are just some of the areas of application of these methods.