Contrastive Representation Learning for Highly Imbalanced Multivariate Time Series With Extreme Instance Strategy

Time series data refers to a sequence of data points collected or recorded at regular time intervals. In many fields including space weather, healthcare, and finance, predicting events from such data is crucial because these predictions can help protect infrastructures, improve healthcare outcomes, and forecast financial trends. However, one challenge in working with time series data is the presence of rare events, which are often underrepresented in the data. This imbalance makes it difficult for traditional prediction methods to provide accurate results, as they tend to focus more on the more frequent events and overlook the rare ones. To tackle this challenge, this research introduces a new approach using contrastive representation learning. This method improves the ability to predict rare events by learning more effective representations of the data. Specifically, contrastive learning helps group similar data points together in a way that emphasizes the important differences between classes, such as rare and common events. By learning to focus on these distinctions, the approach provides better predictions, even in situations where the data is imbalanced. In this work, the proposed method first extracts key features from time series data, then creates meaningful representations that help distinguish between the different types of events. These representations are refined by training a model to recognize patterns over time, capturing the dynamic nature of the data. A final classification step uses these learned features to make accurate predictions. The method was tested on several datasets, including one focused on predicting solar flare events, which is an important task in space weather forecasting. The results show that the approach effectively addresses the challenges of class imbalance, providing more reliable predictions that can be applied to a wide range of time series prediction tasks. By offering a powerful solution to the problem of imbalanced time series data, this research has the potential to improve predictions in areas where rare but impactful events can have serious consequences, ultimately benefiting society by enhancing our ability to forecast and respond to critical events in real time.