Theoretically, we can implement fix seed on the machine learning model to get the same results every run (reproducible)but it may leads to bias. So, in order to prevent bias, I gonna run the model for several times without fix the seed/state. How many time should I run the machine learning model or normally done by researchers to get the result (10 runs,20.. or more)?
It depends on your problem, start with a big number (32 ?) then decrease if the model is stable / when you deal with stability.
You need to know that having multiple runs may be part of the solution: you can create an ensemble of models to get more stable models. Here again, the way to create an ensemble and the number of models you need to run to build an ensemble is determined experimentally. The main alternative to add stability to your model is trough regularisation (that really depend on the model you want to use).