This patent refers to attention mechanisms applicable to the analysis of multivariate time series using recurrent neural networks (RNN). The invention presents a multi-convolutional attention unit capable of generating an independent attention vector for each variable in the multivariate time series data.
The objectives are:
- Enhanced Time Series Analysis: The developed attention unit improves the ability of neural networks to analyze complex time series data, allowing for more precise identification of relationships between variables over time.
- Capturing the Importance of Temporal Steps: By utilizing one-dimensional convolutional operations, the invention enables the capture of the relevance of each temporal step within a sub-pattern, which is essential for understanding how variables interact with one another.
- Flexibility and Scalability: The modular structure of the attention unit, composed of different blocks (division, attention, concatenation, and scaling), provides flexibility in configuring and adapting the system to different datasets and analytical problems.
- Improved Prediction Accuracy: With the ability to focus on different variables and temporal steps, the model is more effective at predicting future behaviors based on historical patterns, which is valuable in various applications such as finance, healthcare, and environmental monitoring.
- Innovation in Machine Learning Models: The patent represents an advancement in how neural networks can be utilized to process and analyze complex temporal data, contributing to the development of more robust and accurate machine learning models.
The patent offers an innovative solution for the analysis of multivariate time series, enhancing the ability of neural networks to capture the dynamics of variables over time and, consequently, improving the accuracy of predictions.
Publication:
- WO2021255515A1, December 23, 2021 (filed on 11/27/2020 - PCT/IB2020/061239)
