In the field of human activity recognition, self-supervised learning has gained significant attention due to its potential to train models without requiring labeled data. However, existing approaches often struggle with limitations when faced with a large volume of annotated data. This study proposes an improved masking strategy to address this issue and improve the reconstruction effect in self-supervised learning models.
Methodology
The proposed masking strategy is based on a combination of channel masking and positional encoding. Channel masking helps to reduce the dimensionality of the input data, while positional encoding provides contextual information about the location of the data points. The authors also use a Transformer-Encoder network to process the data, which allows for better handling of long-range dependencies.
Results
The proposed approach was evaluated on several benchmark datasets, and the results showed that it outperformed existing methods in terms of reconstruction accuracy. Additionally, the authors found that the improved masking strategy helped to bridge the gap between self-supervised and supervised learning models.
Conclusion
In summary, this study proposes an improved masking strategy for self-supervised masked reconstruction in human activity recognition. The proposed approach uses a combination of channel masking and positional encoding, along with a Transformer-Encoder network, to improve the reconstruction effect and reduce the gap between self-supervised and supervised learning models. The results demonstrated the effectiveness of the proposed approach, which has important implications for improving the accuracy of human activity recognition systems in various applications.
