mdmpy package¶
Submodules¶
mdmpy.mdm module¶
This module contains the main class, MDM, in this package.
Within the class are methods related to solving for the Maximum Likelihood Estimate (MLE) of coefficients. Two methods are used to initialise a model using pyomo, which can then be passed onto a solver to be solved. Another way to solve for the coefficients is to use a gradient ascent.
- class mdmpy.mdm.MDM(input_dataframe: ~pandas.core.frame.DataFrame, ch_id: int, num_choices: int, list_of_coefs: ~numpy.ndarray[~typing.Any, ~numpy.dtype[~numpy.float64]], input_cdf: ~typing.Callable[[...], float] = <function exp_cdf>, input_pdf: ~typing.Callable[[...], float] = <function exp_pdf>)¶
Bases:
objectThe main class of the package. This models the Marginal Distribution Models (MDM)
- add_conv(conv_min: float = 0)¶
This function restricts the argument of the CDF and PDF such that they are above a set limit, commonly zero. This restricts the domain to a region whereby the 1-CDF is convex.
- grad_desc(initial_beta: ndarray[Any, dtype[float64]], max_steps: int = 50, grad_mult=1, eps: float = 1e-07, verbosity=0)¶
Starts a gradient-descent based method using the CDF and PDF. Requires a starting beta iterate.
TODO: to add f_arg_min which will be pass onto the gradient calculators and use grad_lambda_beta to move towards a convex region, which is above f_arg_min
TODO: Add some sort of linesearch method. That is, a method that uses a direction and tries varies stepsizes to check what happens to the loglikelihood with those stepsizes and uses some rules or heuristics to decide which stepsize to choose
- ll(input_beta: ndarray[Any, dtype[float64]], corr_lambs=None) float¶
This function gets the log-likelihood using the current beta. If the corresponding lambdas for each individual are given, then it will use those, rather than re-computing them, which saves computations
- model_init(heteroscedastic=False, alpha_to_be_fixed=0, alpha_cons_limits: tuple[float, float] = (0, 1), usc_ascs=False, asc_fixed_to_zero=None, model_seed: int | None = None, v=0)¶
This method initializes the pyomo model as an instance attribute m. Values can later be taken directly from m.
The various keyword arguments of this method are used to customise the resulting model. They include alternative-specific constants (ASCs) which act as y-intercepts depending on which choice was made.
Another keyword argument is involved in deciding whether the model will handle heteroscedasticity.
mdmpy.util module¶
This module contains utility functions which are called by the main class.
- mdmpy.util.default_bisect_func(input_cdf: Callable[[...], float], input_beta, input_x, lambda_: float) float¶
This is the default bisection function. The last input will be varied during the bisection search using partial from functools.
- mdmpy.util.exp_cdf(x, lambda_: float = 1)¶
Exponential Cumulative Distribution Function (CDF), with default parameter 1
- mdmpy.util.exp_pdf(x, lambda_: float = 1)¶
Exponential Probability Density Function (PDF), with default parameter 1
- mdmpy.util.find_corresponding_lambda(input_cdf: ~typing.Callable[[...], float], input_beta: ~numpy.ndarray[~typing.Any, ~numpy.dtype[~numpy.float64]], input_x: ~numpy.ndarray[~typing.Any, ~numpy.dtype[~numpy.float64]], bisect_func=<function default_bisect_func>, lamb_const: float = 50000, max_lamb_retries: int = 1000, lamb_coef: float = 1.4) float¶
This function is called to find lambda given the model, input beta and input x_i. It does this by first using a large number which may be too big and causes overflows, but then reduces the positive and negative parts separately until the valid gives a valid output. Then, with a positive output and a negative output, a bisection search for the root is performed.
Module contents¶
This package revolves around the implementation of Marginal Distribution Models, which is implemented as a class MDM