Intelligently Augmented Contrastive Tensor Factorization: Empowering Multi-dimensional Time Series Classification in Low-Data Environments

TL;DR

Multi-dimensional time series classification under low-data conditions is hampered by the need to learn cross-factor interactions and intra-class variability. The authors propose ITA-CTF, a data-efficient framework that combines intelligently targeted augmentation (ITA) with a class-aware contrastive tensor factorization (CTF) extractor and an MLP classifier, jointly optimized via $L_{rec}$, $L_{con}$, and $L_{reg}$. ITA uses a dynamically sampled soft class prototype to guide DTW-based augmentations, enabling targeted pattern-mixing that preserves class characteristics; CTF learns discriminative, intra-class-aware representations through a contrastive objective integrated into the CPD reconstruction. Across five tasks spanning robotics, traffic, and power grids, ITA-CTF achieves up to 18.7% improvements over tensor factorization and deep baselines, while maintaining real-time inference. The framework offers practical benefits for real-world systems with limited labeled data and complex cross-dimensional dependencies, and its task- and domain-agnostic design suggests broad applicability, including smart healthcare and other sensor-rich domains.

Abstract

Classification of multi-dimensional time series from real-world systems require fine-grained learning of complex features such as cross-dimensional dependencies and intra-class variations-all under the practical challenge of low training data availability. However, standard deep learning (DL) struggles to learn generalizable features in low-data environments due to model overfitting. We propose a versatile yet data-efficient framework, Intelligently Augmented Contrastive Tensor Factorization (ITA-CTF), to learn effective representations from multi-dimensional time series. The CTF module learns core explanatory components of the time series (e.g., sensor factors, temporal factors), and importantly, their joint dependencies. Notably, unlike standard tensor factorization (TF), the CTF module incorporates a new contrastive loss optimization to induce similarity learning and class-awareness into the learnt representations for better classification performance. To strengthen this contrastive learning, the preceding ITA module generates targeted but informative augmentations that highlight realistic intra-class patterns in the original data, while preserving class-wise properties. This is achieved by dynamically sampling a "soft" class prototype to guide the warping of each query data sample, which results in an augmentation that is intelligently pattern-mixed between the "soft" class prototype and the query sample. These augmentations enable the CTF module to recognize complex intra-class variations despite the limited original training data, and seek out invariant class-wise properties for accurate classification performance. The proposed method is comprehensively evaluated on five different classification tasks. Compared to standard TF and several DL benchmarks, notable performance improvements up to 18.7% were achieved.

Figures (10)

• Figure 1: Illustration of tensor factorization applied to a slice (n-th sample) of multi-dimensional tensor data, decomposing it into $R$ components (latent factors). Darker areas in vectors $\mathbf{a}_r$ and $\mathbf{b}_r$ represent higher contributions from sensors or time steps to the respective latent component. The weight coefficients $z_r$ quantify the importance of each component.
• Figure 2: Contrastive learning, where embeddings of similar samples are pulled closer together and embeddings of dissimilar samples are pushed further apart.
• Figure 3: The proposed framework consists of three simple but highly effective modules: ① an intelligently targeted augmentation module, ② a class-aware contrastive tensor factorization-based feature extractor that is also cognizant of important intra-class variations via the intelligent augmentations, and ③ a downstream MLP classifier.
• Figure 4: Two random samples from same terrain class label can have considerable intra-class variations (as shown by the correlations of sensors) due to varied operating conditions such as different robot walking speeds.
• Figure 5: Visualization of the inter-connected electric grid network in PSML datasetzheng2022multi, red symbols represent different fault types at different locations.