Amazon AI Researchers Introduce MEMENTO: Neural Model for Estimating Individual Treatment Effects for Multiple Treatments

Estimating the effect of a binary intervention from historical observational data has been extensively investigated within the causal inference literature. However, the main goal of these investigations has been to estimate the typical impact of the intervention. It is possible that the variation in response to the intervention (treatment) between individuals and groups within a population will be lost in the calculation of the mean treatment effect (ATE). Therefore, it is crucial to calculate the treatment’s impact individually.

Most causal estimation methods have concentrated on a situation where the treatments are either on or off. Making predictions regarding causal effects when there are numerous mutually incompatible treatments is, by extension, a significant challenge in many fields.

Without having direct access to the process that gave rise to the treatments, observational data comprises past interventions, their effects, and maybe further context. For cases with more than two distinct values within a discrete and finite set of treatments, prediction is of special relevance. Confounding is an important part of inferring causal effects from observational data. A confounder of the treatment’s effect on its outcome is a variable that influences the therapy and its outcome. Controlling for such a confounder is the conventional practice if its impact can be quantified.

Most of the work on treatment impact estimation is devoted to binary treatments, which does not easily generalize to the case of multiple treatments. A new study by Amazon provides a method for estimating the individual-level effects of treatments when those effects are discontinuous and finite. 

The researchers provide expanded definitions for factual and counterfactual prediction mistakes in the presence of many treatments, extending this method to multi-treatment scenarios and determining an upper constraint on the total amount of factual and counterfactual losses. 

By minimizing an upper bound on the total of the factual and counterfactual losses, the method obtains consistent representations of the confounders across treatment types. 

To minimize this loss, the researchers suggest using the neural model. Since the suggested technique and system memorize treatments (or events) through the construction of representations, they refer to it as MEMENTO. As a framework, MEMENTO supplies both a loss function and a model that seeks to minimize that function. Therefore, any underlying modeling technique that can optimize for the loss can be used.

Amazon uses MEMENTO to determine a product’s MOQ or minimum order quantity. The team compares the results of competing algorithms on public and fake datasets to guarantee reproducibility. In some use cases, experiments on real and semi-synthetic datasets reveal that MEMENTO can outperform established approaches for multi-treatment situations by about 10%.

With the implementation of MEMENTO on the Minimum Order Quantity problem, the system entered full-scale production on March 21. According to an A/B experiment done on a new marketplace, the findings show MEMENTO has an impact of 4.7% reduction in shipping costs when applied to the issue of MOQ.

This methodology could be used in various contexts inside and outside of Amazon. In the future, the team hopes to expand its study to include continuous treatments and find ways to provide accurate estimations when key variables are missing.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'MEMENTO: Neural model for estimating individual treatment effects for multiple treatments'. All Credit For This Research Goes To Researchers on This Project. Check out the paper.

Please Don't Forget To Join Our ML Subreddit