ReTriM: Reconstructive Triplet Loss for Learning Reduced Embeddings for Multi-Variate Time Series

Advances in sensor networks have allowed for capturing complex patterns (over time) spread across multiple sensors generating multi-variate time series (MVTS). Real-World time series are not often represented in the form best suited for analytical tasks such as classification. AutoEncoders(AEs), an unsupervised mechanism, have shown promise at learning a low-dimensional embedding, as a consequence, the class associations become intractable in the learned embeddings. Contrastive learning (CL) demonstrated that one can learn class-aware distance-based embedding that can maximize the distance between the embeddings if they belong to the same class, or minimize the distance if they do not belong to the same class. This leads to representation being lossy, i.e. some amount of information from the input is lost. Relying on these observations, this paper presents the ReTriM Framework to map multi-variate time series to vectorized embeddings. ReTriM observes that one can learn a universal embedding, independent of the classifier model when learning with a hybrid strategy that pushes for both, retention of information from original input and separation between the classes. As shown in the experimental evaluation, ReTriM can outperform the embeddings learned from only using CL and AEs for various classifiers such as MLP, One-Shot Classifier, SVM, KNN1, Random Forest, and Logistic Regression for 10 benchmark multi-variate time-series datasets.