How to get Structural disorder probability in keras API about dspp-keras

[ktamiola]

1. shift the propensity scores by -2.0, so that you are working in -1.0; 1.0 range
2. compute the 1.0-abs(shifted_propensity_score)

That's it! The probability of disorder is simply proportional to absolute propensity score.

Please read our preprint: http://biorxiv.org/content/early/2017/06/01/144840 We have explained how to use the disorder probability in the results section.

@ktamiola sorry, I have read the paper today and understood the formulation, but suddenly forgot it for a while. My silly mistake, and thanks for remembering...

[ktamiola]

Pleasure buddy! It's a relatively long thing to read! Allow me to tick this off the list and close the issue. Please reopen if more disordered questions come to your mind.

But what about missing values, which are zero without the -2 shift?! How does your formula handle it? By this formula, the probability of missing values is -1! Too complicated indeed!
Of course, we can filter such missing values, but process is a bit untidy.

[ktamiola]

@ErmiaAzarkhalili missing values are always a pain in the neck. One of the approaches is to back-compute and extrapolate missing values (obviously this is up to user), or use binary weights during model training in Keras.