This is an accompanying repository for a paper, in it you will find code and documentation, representing the ideas discussed in the paper, as well as extra (animated) visualisations.
The paper focusses on a number of requirements that must be defined for any causal simulation. These requirements will form a guideline throughout the documentation of this code.
A core idea behind the discussed papers is one of base functions. Two base functions are present in the current code base:
D is the dimension count, f is the frequency and G (displacement vector) will govern the average causal effect in the simulation.
Here, K (the coefficient vector) implies a crucial difference with the Sine_Base as it implies that a Polynomial_Base must be created for every cause. This choice is motivated by the fact that one can control the difference between two coefficient_vector's while in a Sine_Base, this difference is already controlled by the displacement_vector.
Drift is controlled by the top level class Simulation(C, D, base_functions, drift_rate, drift_over_time, sudden_drift, drift_moments, std) through its parameters drift_rate, drift_over_time, sudden_drift and drift_moments where the first three are necessary when random drift is required. These first three parameters can be omitted when the drift_moments are chosen before the simulation is running. These moments are moments in function of a t representing an experiment number. When the drift moments are to be chosen randomly by the Simulation, drift_rate describes the amount that is drifted, drift_over_time governs how often or how gradual the drift is taking place and sudden_drift is False when the drift_rate is linearly distributed over drift_over_time and True when (on average) the complete drift_rate happens once in drift_over_time.
Drift for the Sine_Base is illustrated below.
When every a cause C is chosen, one can call .chose_cause(C) and an effect is returned. When a response is required that does not include drift or noise the method can be called as .chose_cause(C, drift=False, n=False).
In case of uplift modeling, one can consult the actual (ground truth) uplift for a specific x by calling .get_sim_uplift(x) on a Simulation. Other intricacies of the simulation can be acquired upon extension.
As an extra parameter to Simulation, std governs the standard deviation of the Gaussian noise over the returned effect of the simulation. Extensions towards other distributions for noise must be added manually.