Robust Adaptive Learning for Real-World Data with Changing Features and Imperfect Labels

In many machine learning tasks, environments are complex and unstable. Features in the data change over time. The labels in the data contain errors. These issues make model training very difficult. This paper focuses on a situation in which both problems happen at the same time. Learning in such a case is very challenging. It is even harder when the data in the new stage is limited. A new method termed ALDN is proposed. This stands for Adaptive Learning for Dynamic Features and Noisy Labels. The method has two main stages. The first stage learns a model from clean data. The second stage uses that model to help train a new model. The feature space has changed in this second stage. To do this, a tool titled optimal transport is used. This helps map the old model into the new feature space. Then, a special rule is applied to align the behavior of the old model with the new one. This rule is named a regularizer. It helps the new model learn better, even with noisy labels. Two versions of the method are presented. One uses a direct rule to keep the models close. The other uses an indirect rule that compares predictions. It is proven that the method works well in theory. The method is tested through many experiments. Comparisons are made with other identified methods. Better performance is shown in most cases. It works well even when the features change, and the labels are not reliable. This paper presents a useful way to deal with both dynamic features and noisy labels. It reuses past models and adapts them to new data. This helps when there is little clean data and the new data is messy. The method is practical and backed by strong theory and experiments.