I built an LSMT model to predict sick cows. I also have risk factors like cow size and height (static risk factor) that I want to combine into the ML model. I found that size is geometrically distributed. My question is how I insert it as a feature to the model? I know that $P(x=K)= p*q^(k-1)$ but I don't know how to combine it as a feature. Thank you.
1
-
1$\begingroup$ Are you sure you want to insert a theoretical distribution instead of the measured value for each cow? $\endgroup$– DaveApr 20, 2021 at 11:43
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
2
As a general approach I would say you need to generate new features, that use your prior knowledge. For example, if you have a known size distribution, then for each specific size you can calculate its probability and use it as a new feature.
As I side-note, the geometric distribution of cow sizes seems very surprising to me, I would expect to see some gamma distribution or just normal (if size is measured in cm/inches).
-
$\begingroup$ I just not understand how to calcalute it. for example if it is geometric so for cow at size 50 the formula is $P(x=K)= p*q^(k-1)$ , how I can get P? $\endgroup$– MorApr 21, 2021 at 10:46
-
$\begingroup$ It's a distribution parameter which you can estimate by training data with methods like maximum likelihood estimation. "I found that size is geometrically distributed" -- could you please add to question the explanation, how did you find that and may be example of the data. $\endgroup$ Apr 21, 2021 at 10:56
$\begingroup$
$\endgroup$
Using a probability distribution as a feature is not possible in most commonly-used machine learning frameworks. Most commonly-used machine learning frameworks only accept scalar-like values as inputs. In the case of height, it would be a single numeric measurement.
If you willing to go outside of established frameworks, you could model the problem in a Bayesian way with probabilistic programming where all quantities are distributions.
answered Jul 6, 2021 at 20:48