The authors of this application predict that causal inference will
continue to be a major research imperative of Artificial
Intelligence. In many cases, we are interested in knowing the causal
risk factors of a chosen target. For examples, in medicine, we would
be interested in knowing the causal risk factors for breast cancer
metastasis; In biomedicine, we would be interested in knowing the
genetic risk factors that are directly responsible for Alzheimer’s
disease; in clinical practice, we would be interested in knowing the
treatment effect of a new drug on a particular disease. Note that
cause and causal effect (the treatment effect in my occasions) go
side by side. We have many existing machine learning methods that
can help learn associations/correlations/dependencies between some
factors and a chosen target, which is often called a class feature
in machine learning, but those factors are not causal and may not
have a causal influence on the chosen target. Learning the causal
factors of a chosen target from data is a state-of-the-art approach
for understanding the causal mechanisms, helping improve
personalized prediction of the chosen target and conduct
intervention efficiently and effectively.

Algorithm which uses propensity scores based off of all covariates in order to do matching, and then calculates the average treatment effect using these matches.

Algorithm using an iterative process to slowly work through all the data selecting the best match based on propensity scores calcualted via all covariates. May take a while due to the iterative nature. Best to use on small datasets.

Uses propensity scoring and inverse treatment weighting. For the unit that received treatment, the weight is the reciprocal of the propensity score. For the unit that did not receive treatment, the weight is the reciprocal of 1 minus the propensity score.

Uses the g-formula general equation to obtain the marginal causal effect of the specified treatment on the outcome. This method assumes the model is correctly specified.

FCI (Fast Causal Inference) is an algorithm designed to be find causal relationships between variables in directed acrylic graphs. This iteration outputs a png of the causal graph.

rFCI is an iteration of FCI, which has been adjusted for improvements in runtime. FCI is an algorithm with polynomial complexity, leading to excessive runtimes with larger quantities of nodes. rFCI speeds up this process, but as a result provides less information.

GES(Greedy Equivalance Search Architecture) is an algorithm with similar goals to FCI, however it begins with a completely empty graph instead of a full graph and adds edges as needed. The GES graph is known to have some issues finding confounding variables, however it is generally a better starting place for many datasets.

FGES is another iteration of GES, and shares a similar relationship to GES as rFCI does to FCI. FGES is a faster and more optimized version of GES that is ideal for sparser graphs.

MBIL is another causal learning algorithm designed to find causal learning factors in various datasets. Initially developed to find causal factors of breast cancer metastasis, MBIL has been shown to outperform with other causal learning methods on select datasets. The MBIL algorithm is the most customizable algorithm on this site, as it provides the user with numerous parameters to help the user gain more specificity. Like all the other algorithms, the MBIL method can take up to 1 minute to run on larger datasets.