Design of an Iterative Method for Efficient Streaming Data Analysis Using Dynamic Sparse Temporal Convolutional Networks and Sparse Meta-Adaptive Learning

Resumen

Efficient real-time data processing in streaming environments is a key requirement for applications in financial forecasting, autonomous driving, and other scenarios. Traditional Temporal Convolutional Networks (TCNs) are good at modeling temporal dependencies but suffer from high computation, low interpretability, and being not adaptive to data variability over time. In this work, we propose a novel architecture, Dynamic Sparse Temporal Convolutional Networks (DSTCNs), working toward improving the computational efficiency of TCNs by incorporating sparse activations and meta-learning techniques. DSTCNs use dilated convolution, extending the receptive field without increasing computational complexity, alongside sparse activation functions, that selectively activate the neurons and reduce the computational overhead. In addition, we introduce meta-learning strategies that can adaptively adjust the parameters in response to concept drift, hence keeping model accuracy through changing data streams. Another approach, Sparse Meta-Adaptive Learning (SMAL), addresses the issue of concept drift by dynamically adjusting the model parameters in real-time, according to the dynamics of data distributions, resulting in optimized computational resources and smaller latency. Empirical results show that DSTCNs consume about 30% fewer computational resources than traditional TCNs with the same level of accuracy. SMAL could reduce prediction errors by up to 25% compared to baseline methods, thus proving its applicability in dynamic scenarios. These innovations present a new benchmark for the analysis of streaming data and would have significant implications for both academia and real-world applications to streamline real-time data-driven decision-making systems.