Causal inference via neuroevolutionary selection
The technology is directed towards receiving training data regarding a set of observations. Each observation includes a feature set, a treatment, and an outcome. A first generation of machine learning models is trained, via the training data, to predict an outcome for a feature set of a given observation. A new generation of models is generated by selecting a subset of models from the trained first generation of models based on a fitness criteria of each model to generate an intermediate layer for use in predicting a treatment. An algorithm is applied to the selected subset of models to generate the new generation of models. Transformed training data is generated using the training data and a model of the new generation of models. The transformed training data includes, for each observation, a transformed feature set comprising a representation of the feature set in a latent space of the model.
1 . A non-transitory computer-readable medium storing executable instructions, which when executed by a processing device, cause the processing device to perform operations comprising:
receiving training data regarding a set of observations, each observation including a feature set, a treatment, and an outcome;
training, using the training data, a first generation of machine learning (ML) models to predict an outcome for a feature set of a given observation;
generating a new generation of ML models by:
selecting a subset of ML models from the first generation of ML models based on a fitness criteria of each ML model to generate an intermediate layer for use in predicting a treatment; and
applying an algorithm to the selected subset of ML models to generate the new generation of ML models;
generating, using a ML model of the new generation of ML models, a transformed training data from the training data, the transformed training data comprising, for each observation, a transformed feature set comprising a representation of the feature set in a latent space of the ML model, wherein the transformed feature set causes an accuracy of a causal effect estimation by one or more ML models trained based on the training data to improve by at least reducing a treatment-related bias; and
providing a trained ML model of the new generation of ML models capable of estimating a heterogeneous causal effect for a subject not included in the set of observations based on a representation of the subject in the latent space of the ML model.
2 . The non-transitory computer-readable medium of claim 1 , wherein training the first generation of ML models comprises:
for each ML model of the first generation of ML models, initializing a set of weights for a first transformation of the ML model based on stochastic sampling of one or more distributions of initial weights; and
for each ML model of the first generation of ML models, iteratively updating the set of weights for the first transformation based on the fitness criteria.
3 . The non-transitory computer-readable medium of claim 1 , wherein generating the new generation of ML models comprises:
selecting a fittest-model, based on the fitness criteria, from the first generation of ML models; and
generating the new generation of ML models to include the fittest-model from the first generation of ML models.
4 . The non-transitory computer-readable medium of claim 1 , wherein generating the new generation of ML models comprises:
for each possible pairing of two ML models from the selected subset of ML models, generating a descendent-model based on stochastically generating a set of genetic crossovers between the two ML models of the pair of ML models; and
generating the new generation of ML models to include the descendent-model of each possible pairing of two ML models from the selected subset of ML models.
5 . The non-transitory computer-readable medium of claim 4 , wherein a pair of two ML models includes a first ML model and a second ML model, and the set of genetic crossovers between the first ML model and the second ML model includes a stochastic shuffling of elements of a first transformation of the first ML model and elements of a first transformation of the second ML model to form a first transformation for a descendent-model of the pair of two ML models that generates the transformed feature set for each observation of the transformed training data.
6 . The non-transitory computer-readable medium of claim 1 , wherein each ML model of the new generation of ML models includes a separate first transformation that generates the transformed feature set for each observation of the transformed training data and a separate second transformation that generates a third feature set for each observation of the transformed training data, and generating the new generation of ML models comprises:
for each ML model of the new generation of ML models, initializing a set of weights for the second transformation of the ML model based on stochastic sampling of one or more distributions of initial weights; and
for each ML model of the new generation of ML models, iteratively updating the set of weights for the second transformation based on decreasing a value of an objective function that indicates an expected value for the third feature set being predictive of a treatment assignment for the observation.
7 . The non-transitory computer-readable medium of claim 6 , wherein generating the new generation of ML models further comprises:
for each ML model of the new generation of ML models, assigning a fitness metric to the ML model that is based on the fitness criteria, wherein the fitness metric scales with the decreased value of the objective function for the ML model; and
employing the fitness criteria to select a fittest-model of the new generation of ML models, wherein the selected fittest-model has been assigned a largest fitness metric of the new generation of ML models.
8 . The non-transitory computer-readable medium of claim 1 , wherein the operations further comprise:
determining a heterogeneous effect of the treatment based on the transformed training data and a causal ML model.
9 . The non-transitory computer-readable medium of claim 8 , wherein the causal ML model is a meta-learning model, the heterogeneous effect is a conditional average treatment effect, and the operations further comprise:
employing the transformed training data to train the meta-learning model.
10 . The non-transitory computer-readable medium of claim 1 , wherein the algorithm is a genetic algorithm.
11 . A method comprising:
receiving data regarding a set of observations, the data for each observation including a feature set, a treatment, and an outcome;
determining, using a machine learning (ML) model of a new generation of ML models, a transformed feature set from the feature set for each observation, the new generation of ML models generated by applying an algorithm to a subset of ML models selected from a first generation of ML models based on a fitness criteria of each ML model to generate an intermediate layer for use in predicting a treatment, the transformed feature set comprising, for each observation, a representation of the feature set in a latent space of the ML model;
generating training data comprising, for each observation, the transformed feature set for the observation, the treatment for the observation from the data, and the outcome for the observation from the data; and
training a causal model based on the training data, wherein the causal model is trained to generate a causal effect estimation associated with a subject not included in the set of observations by reducing a treatment-related bias based on the training data and a representation of the subject in a vector space associated with the transformed feature set.
12 . The method of claim 11 , wherein the algorithm is a genetic algorithm and the causal model is implemented by a meta-learner.
13 . The method of claim 11 , wherein training the causal model includes training the causal model to determine a heterogeneous effect of the treatment based on the transformed training data.
14 . The method of claim 11 , wherein each ML model of the new generation of ML models is implemented by a neural network-based architecture and the intermediate layer is an intermediate layer in the neural network-based architecture.
15 . The method of claim 11 , further comprising:
employing other training data to train the first generation of ML models to predict an outcome for a feature set of a given observation.
16 . The method of claim 15 , wherein training the first generation of ML models comprises:
for each ML model of the first generation of ML models, initializing a set of weights for a first transformation of the ML model based on stochastic sampling of one or more distributions of initial weights; and
for each ML model of the first generation of ML models, iteratively updating the set of weights for the first transformation based on the fitness criteria.
17 . The method of claim 15 , wherein generating the new generation of ML models comprises:
selecting a fittest-model, based on the fitness criteria, from the first generation of ML models; and
generating the new generation of ML models to include the fittest-model from the first generation of ML models.
18 . A system comprising:
a memory component; and
a processing device coupled to the memory component, the processing device to perform operations comprising:
receiving training data regarding a set of observations, each observation including a feature set, a treatment, and an outcome;
training, using the training data, a first generation of first machine learning (ML) models to predict an outcome for a feature set of a given observation;
for each first ML model from the first generation of first ML models, training a second ML model to predict a treatment using an intermediate layer of the first ML model as an input layer to the second ML model;
generating a new generation of first ML models by:
selecting a subset of first ML models from the trained first generation of first ML models based on a fitness criteria of each first ML model to generate an intermediate layer for use by a corresponding second ML model in predicting a treatment; and
using an algorithm to generate the new generation of first ML models using the selected subset of first ML models;
generating, using a first ML model of the new generation of ML models, a transformed training data from the training data, the transformed training data comprising, for each observation, a transformed feature set comprising a representation of the feature set in a latent space of the first ML model and the transformed training data is generated based on a neural network layer trained using the algorithm, wherein the algorithm is a genetic algorithm that evolves model parameters over successive generations; and
providing a trained ML model of the new generation of ML models capable of estimating a causal effect for a subject not included in the set of observations based on the representation of the subject in the latent space of the ML model.
19 . The system of claim 18 , wherein generating the new generation of ML models comprises:
for each possible pairing of two ML models from the selected subset of ML models, generating a descendent-model based on stochastically generating a set of genetic crossovers between the two ML models of the pair of ML models; and
generating the new generation of ML models to include the descendent-model of each possible pairing of two ML models from the selected subset of ML models.
20 . The system of claim 19 , wherein a pair of two ML models includes a first ML model and a second ML model, and the set of genetic crossovers between the first ML model and the second ML model includes a stochastic shuffling of elements of a first transformation of the first ML model and elements of a first transformation of the ML second model to form a first transformation for a descendent-model of the pair of two ML models that generates the transformed feature set for each observation of the transformed training data.