Generalizable Low-Resource Activity Recognition with Diverse and Discriminative Representation Learning (2306.04641v2)

Abstract: Human activity recognition (HAR) is a time series classification task that focuses on identifying the motion patterns from human sensor readings. Adequate data is essential but a major bottleneck for training a generalizable HAR model, which assists customization and optimization of online web applications. However, it is costly in time and economy to collect large-scale labeled data in reality, i.e., the low-resource challenge. Meanwhile, data collected from different persons have distribution shifts due to different living habits, body shapes, age groups, etc. The low-resource and distribution shift challenges are detrimental to HAR when applying the trained model to new unseen subjects. In this paper, we propose a novel approach called Diverse and Discriminative representation Learning (DDLearn) for generalizable low-resource HAR. DDLearn simultaneously considers diversity and discrimination learning. With the constructed self-supervised learning task, DDLearn enlarges the data diversity and explores the latent activity properties. Then, we propose a diversity preservation module to preserve the diversity of learned features by enlarging the distribution divergence between the original and augmented domains. Meanwhile, DDLearn also enhances semantic discrimination by learning discriminative representations with supervised contrastive learning. Extensive experiments on three public HAR datasets demonstrate that our method significantly outperforms state-of-art methods by an average accuracy improvement of 9.5% under the low-resource distribution shift scenarios, while being a generic, explainable, and flexible framework. Code is available at: https://github.com/microsoft/robustlearn.

• The paper introduces a DDRL method that mitigates data scarcity and distribution shifts in HAR by leveraging self-supervised learning.
• It employs sensor data transformations to enhance feature diversity while using supervised contrastive learning to enforce class discrimination.
• Experimental results demonstrate an average 9.5% accuracy improvement on three public HAR datasets in low-resource environments.

Overview of Generalizable Low-Resource Activity Recognition

The paper "Generalizable Low-Resource Activity Recognition with Diverse and Discriminative Representation Learning" addresses a critical problem in the field of Human Activity Recognition (HAR): the challenge of creating accurate models when data is scarce and diverse. This paper revolves around two major issues faced in HAR—low-resource environments and distribution shifts among data collected from various individuals. To mitigate these issues, the authors propose a novel approach, utilizing both diversity and discrimination in representation learning, which shows significant improvement over existing methods in low-resource settings.

Problem Definition and Challenges

HAR involves classifying human activities based on sensor data. The primary challenge in this domain is the scarcity of adequately labeled data. Collecting large-scale labeled datasets can be prohibitive due to time and financial constraints. Additionally, the distribution of data can vary significantly across individuals due to differences in lifestyle, body morphology, and demographics. This variability often results in performance degradation when models trained on one dataset are applied to new, unseen data.

Proposed Method: Diverse and Discriminative Representation Learning

The paper proposes a methodology known as Diverse and Discriminative Representation Learning (DDRL) to enhance HAR models' generalization ability in low-resource settings. The novelty of DDRL lies in its two-fold strategy:

1. Diversity Generation: By employing self-supervised learning tasks with strategic sensor data transformations, this approach aims to increase data diversity. These transformations include rotations, permutations, and magnitude alterations, designed to simulate real-world variations in sensor readings. The self-supervised task helps in learning invariant properties of activities by classifying the type of transformation applied to the data.
2. Diversity Preservation and Discrimination Enhancement: Through a diversity preservation module, the method ensures the learned feature representations of original and augmented data are distinct yet expansive, preventing feature space overlap. Additionally, DDRL leverages supervised contrastive learning to strengthen semantic discrimination, wherein it enhances the intra-class compactness and inter-class separability of the feature representations.

Experimental Results and Implications

Through extensive experimentation on three public HAR datasets, the proposed DDRL method demonstrated an average accuracy improvement of 9.5% over state-of-the-art methods in low-resource, distribution-shift scenarios. This improvement highlights the efficacy of using a diversity and discrimination-focused approach in enhancing model robustness and performance.

The experimental results underscore DDRL's potential to generalize beyond existing domains, making it a potentially valuable contribution to fields such as healthcare monitoring and context-aware computing, where sensor data is prevalent but heterogeneous.

Future Directions

This paper opens avenues for further research into the application of self-supervised and contrastive learning techniques within HAR and other domains where data scarcity and distributional shifts pose significant challenges. Future studies could expand upon these methods by exploring additional transformation techniques or integrating federated learning strategies to enhance privacy while utilizing distributed datasets.

In summary, the work presented in this paper marks a significant step toward solving the generalizable low-resource HAR problem, demonstrating a feasible approach that holds promise for broader real-world applications.