COIM

Constrain Operator for Inferential Models is a simple tool for pre and pos processing of data to eliminate redunduncy in datasets caused by dependency rules between the variables/columns.

Usage

To start using COIM, import into your code the operator class, which orquestrates the constrains and define an instance.

from COIM import ConstrainOperator
CO=ConstrainOperator()

To add a new constrain, use the add_rule method from ConstrainOperator class.

from COIM import SomeConstrain
SC=SomeConstrain(**parameters)
CO.add_rule(SC)

Each constrain will require their own specific parameters, refer to section Available constrains to know each of them. However, all constrains receive the parameter "labels", which is a list with the new names to be used on the encoded columns.

Then you can encode your dataframe to use the new corrected variables to feed your model.

new_df=CO.encode_dataframe(df)

After running your model, you can regenerate the data in the original format, decoding the acquired values and errors.

decoded_df, decoded_errors=CO.decode_dataframe(predicted_df, errors)

That will yield the predictions for the original variables as if they had been fed to the model themselves, but with rather more consistent results

Available constrains

1. "add_scalar":
   • $a+K=b$
   • base_variable = a
   • target_variable = b
   • constant = K
2. "mul_scalar":
   • $a*K=b$
   • base_variable = a
   • target_variable = b
   • constant = K
3. "const_sum":
   • $\sum W_i\cdot a_i=K$
   • variables = $[a_1, a_2, \cdots, a_n]$
   • reference_variable = $a_j$
   • constant_sum = K
   • weights = $[W_1, W_2, \cdots, W_n]$ or $W$ if $W_1=W_2= \cdots= W_n$
4. "custom_func":
   • to be used when none of the above is applicable and you have to develop your own functions to operate the dataframe
   • variables : list of the variables to be used
   • validate_function: Function to assert if the received dataframe follows the given constrain. (df[DataFrame], variables[list], labels[list])->bool
   • format_function: Write a string that describes the constrain equation. (variables[list], labels[list])->str
   • encode_dataframe: Create the new custom columns in the dataframe. (df[DataFrame], variables[list], labels[list])->DataFrame
   • decode_dataframe: Restore the original columns in the dataframe and calculate the propagated errors. (df[DataFrame], variables[list], labels[list], errors[DataFrame])->DataFrame, DataFrame

Future additions

In the foreseeable future, some new constrains will be implemented, those are:

1. Variable sum
2. Constant and variable products
3. Conditionals

Theoretical foundation

All of the worked out mathematics for the developed constrains can be found at the calculations pdf