ATD: Augmenting CP Tensor Decomposition by Self Supervision.

Area of Science:

• Multidimensional data analysis
• Machine learning
• Signal processing

Background:

• Tensor decompositions are vital for dimensionality reduction and feature interpretation.
• Traditional objectives may not align with downstream classification tasks, as raw data can contain irrelevant information.

Purpose of the Study:

• To propose Augmented Tensor Decomposition (ATD) for enhanced downstream classification.
• To incorporate data augmentations and self-supervised learning (SSL) into tensor decomposition.

Main Methods:

• Developed an iterative alternating least squares (ALS) method to address non-convexity.
• Integrated data augmentation and SSL within the tensor decomposition framework.

Main Results:

• Achieved 0.8%–2.5% accuracy gain over tensor-based baselines.
• Demonstrated comparable or superior performance to self-supervised and autoencoder baselines (up to 15% accuracy gain).
• Required less than 5% of the learnable parameters used by baseline models.

Conclusions:

• ATD effectively boosts downstream classification performance by leveraging data augmentation and SSL.
• The proposed method offers a parameter-efficient alternative to existing deep learning approaches for tensor data.