Amazon Researchers Designed A New Machine Learning Algorithm Based On Entropy Balancing That Learns Weights To Directly Maximize Causal Inference Accuracy Using End-To-End Optimization

A causal effect means that a certain thing is happening based on something that has already occurred. In business, the causal effect of a treatment is very important, for example, changing the font of a page based on the amount of time spent by a user. Treatments can either be binary or can be continuous.

Confounding factors: it’s the third variable while examining a cause and effect relationship.

Usually, there exist confounding factors that influence the treatment as well as response relationship and causal estimation accounts for them. Handling of confounders when treatments are binary is well studied, while it’s not the same for treatments that are continuous. Amazon researchers have put forward a new method to estimate the effects of treatments that are continuous.

Propensity scores

Continuous treatments induce infinitely many possible outcomes per unit. In this case, there’s something called the response curve, which is a mapping between the continuous input to continuous output. If there is a confounder, it becomes difficult to determine the causal relationship between them. The standard way to account for this is by propensity score weighting.

However, propensity scores can be large sometimes, which causes an imbalance and leads to uncertain inference. Entropy balancing is used which selects weights such that the difference between them is min (or the entropy is maximum).

End-to-end balancing

The new algorithm is based on entropy balancing, and it learns the weights to maximize the causal inference by end-to-end optimization. The paper demonstrates the consistency of the approach. They have also studied the impact of misspecification in the synthetic data generation process. i.e, the model always converges to the good estimation of the real response function irrespective of any inaccurate initial selection of random response function.

This new method, when compared to the best-performing predecessor, has improved its mean square error by 27% when the relationship between the response variable and intervention is linear and by 38% when it is nonlinear.

In summary, The research team has developed a new algorithm to estimate the effects of continuously varied treatment. This method uses end-to-end ML, entropy balancing, and propensity score weighting.

This Article is written as a summary article by Marktechpost Staff based on the research paper 'End-to-End Balancing for Causal Continuous Treatment-Effect Estimation'. All Credit For This Research Goes To Researchers on This Project. Checkout the paper and reference article.

Please Don't Forget To Join Our ML Subreddit