# Variation in LSTM Model based on seeds

I am working on a time series project. The results of LSTM model which i am using varies a lot with the variation in seeds. I am wondering how can i make that model stable. Currently to get the reproducible result i am fixing the seed. Is there some way which i can implement to make it stable. The 5 Stratifiedfold MCC (Matthews correlation coefficient) score varies between 0.636 to 0.724 just by using different seeds (for weights initialization).

Any help would be appreciated.

The training dataset has 2904 instances only.

Thanks!

• You can make your question more interesting by specifying what seed you are talking about-- for weights initializer or bias initializer etc., which score you are talking about? Is that score an average across many folds as in $K-Fold$ CV? – naive Jan 29 at 7:25
• I'm not surprised that a model trained with a small number of samples has high variance. Are you? – Sycorax Jan 29 at 15:59
• @Sycorax, I completely understand that, but what shall i do to deal with this. Shall i build n models using different seeds for weight initialization, then do the average or should i do data augmentation. If data augmentation is the key ( do you know what method shall i follow. Thanks! – Harshit Mehta Jan 29 at 16:02
• regularization – Sycorax Jan 29 at 16:03