Q&A

[Barry57]

This post summarizes the issues raised and provides answers. Further questions can also be commented here.

[Barry57]

1. About scalar_ge
   This version of scalar_ge may not converge to a stable solver. This is caused by the stability of the neural network. Try to use grid_scalar_ge to find opt parameters. Our team is trying to enhance the stability of the neural network.

[Barry57]

2. About data imbalance
   In GENetLib, dealing with imbalanced data is actually pretty similar to what you’d do in a regular neural network classification task.

[Barry57]

3. About running time
   Generally speaking, the running speed of func_ge will be faster than that of scalar_ge, while grid search will be slower than running a single model. Especially in the scalar case, it is very likely to result in slower execution, which is related to your computational power and the size of your data. We are also working on integrating GPU support to handle more cases, as the data size we currently encounter is mostly around 500–1000, and we have not yet considered scenarios with larger datasets.

[Barry57]

4. About import error
   The structure of our package was adjusted in version v1.2.2. Some of the module names have been changed, so the import code also needs to be modified. The documentation was also updated accordingly. For details, please refer to https://github.com/XMU-Kuangnan-Fang-Team/GENetLib/releases/tag/v1.2.2

[Barry57]

5. About the sample size
   An excessively large sample size may lead to slow execution. We are currently working on providing a GPU version to enable parallel computing. In GE, we recommend a sample size of no more than 1500, under which the code can run smoothly.

[Barry57]

6. About fda
   Regarding the expansion of FDA(Functional Data Analysis), our paper introduces the FuncGE model, which is based on functional data analysis methods. Consequently, we have incorporated a variety of FDA functions into the package. These functions are designed for analyzing functional data and are stored in the fda_func.py file, making it easy to expand.

[Barry57]

7. About bug report local variable 'opt_Learning_Rate2' referenced before assignment
   As our neural network training has two steps: first training without MCP to optimize the l2 parameters, then training the MCP parameters based on the optimal l2 parameters. This error shows there are no optimal l2 parameters, that is, the first step of training reported an error and there's no proper model.