Node selection for LSTM models is typically arbitrary, with values such as 128 or 256 being commonly used. However, if we can develop a heuristic approach for selecting nodes, we can avoid random node assignment for model parameter selection. This will lead to a more efficient utilization of resources and the creation of better models.
I proposed a metric approach to node selection, which is second to none. By leveraging carefully crafted metrics, I analyze and evaluate its performance on a diverse range of two real datasets using various LSTM architectures.
Unfortunately, we had to abandon the project as some experimental results did not support the proposed metric.
For more details, please check the full report ``Number_of_node_calculation.pdf''.
Note: I want to highlight the last result obtained for stack 3 (Power data) does not support the node calculation; however, the overall performance of the analysis supports it.
Feel free to reach out if you have any questions or feedback.