We need to implement a time series problem by LSTM model.
But, while implementing the same, the main challenge we are facing is feature selection issue. Because, our data-set contain 5000 observations and 30 features. And we already know there are so many features which are not at all relevant. But, no one of has domain expertise to confirm which feature is relevant. The results obtained with use of all predictors are good ( accuracy about 98% ). Our goal is to reduce the number of predictors and keep the accuracy in good marge (> 85% for example)
Check out the shap library. This offers the potential to get a better understanding of the interactions of the features and you will get a sensitivity analysis.
With this information you will be able to reduce the features in the next step.
https://github.com/slundberg/shap
Abderrahim Abbassi after training your model, you can use the print(model.feature_importances_) in python. This will give the scores of the features sorted regarding their importance in getting 98% of accuracy.
Khaoula Belhaj Soulami Thanks for your answer !
in Keras there is no function that calculates the feature importance as far as i know.
Abderrahim Abbassi Python's sklearn library has the SelectKBest method which ranks the top K most significant features in a dataset. For supervised learning, you can use the the chi squared (X^2) function, which has proven to be one of the most effective in feature selection. Check it out.
José Joaquín Peralta Abadía I've read some articles stating that Chi-squared for feature selection is only used for categorical datasets! Otherwise, you can use some correlation methods such as Matrix with heatmap or Pearson's correlation if your data is linear Good luck.
Maybe, there is another way to solve your problem. In context of deep neural networks, there are some promising machanisms like droupout or attention.
You can identify the importance level of each feature by performing an ablation study. I also recommend to look into the work of Karim et al. (2019), "Multivariate LSTM-FCNs for time series classification".
Did you find a solution to this issue? I'm facing the same situation
- Badges
- Science topic