Method, system, and apparatus for casual discovery and variable selection for classification
AssigneeVanderbilt University
InventorsConstantin F. Aliferis, Ioannis Tsamardinos
Life of this patent
Estimating the term…
Abstract
A method of determining a local causal neighborhood of a target variable from a data set can include identifying variables of the data set as candidates of the local causal neighborhood using statistical characteristics, and including the identified variables within a candidate set. False positive variables can be removed from the candidate set according to further statistical characteristics applied to each variable of the candidate set. The remaining variables of the candidate set can be identified as the local causal neighborhood of the target variable.
Claims53
A computerized method of determining a local causal neighborhood of a target variable from a data set, said method comprising:
- (a) adding to a candidate set each variable in the data set having a potential direct causal relationship with the target variable based upon a predefined criteria of association with the target variable, such that no variable of the data set potentially having a direct causal relationship is excluded, the candidate set being defined so as to initially be an empty set prior to the adding step;
- and
- (b) removing from the candidate set each variable determined to be independent of the target variable after the addition of at least one other variable to the candidate set, wherein the remaining variables of the candidate set define the local causal neighborhood of the target variable, wherein the target variable measures the outcome of a predetermined physical phenomenon, and wherein the variables of the candidate set identify associated phenomena that cause or are caused by the predetermined physical phenomenon.
A computer system comprising hardware and associated software for determining a local causal neighborhood of a target variable from a data set, said system comprising:
- means for adding to a candidate set each variable in the data set having a potential direct causal relationship with the target variable based upon a predefined criteria of association with the target variable, such that no variable of the data set potentially having a direct causal relationship is excluded, the candidate set being defined so as to initially be an empty set prior to the adding step;
- and
- means for removing form the candidate set each variable determined to be independent of the target variable after the addition of at least one other variable to the candidate set, wherein the remaining variables of the candidate set define the local causal neighborhood of the target variable, wherein the target variable measures the outcome of a predetermined physical phenomenon, and wherein the variables of the candidate set identify associated phenomena that cause or are caused by the predetermined physical phenomenon.
A machine readable storage, having stored thereon a computer program having a plurality of code sections executable by a machine for causing the machine to perform the steps of:
- (a) adding to a candidate set each variable in the data set having a potential direct causal relationship with the target variable based upon a predefined criteria of association with the target variable, such that no variable of the data set potentially having a direct causal relationship is excluded, the candidate set being defined so as to initially be an empty set prior to the adding step;
- and
- (b) removing from the candidate set each variable determined to be independent of the target variable after the addition of at least one other variable to the candidate set, wherein the remaining variables of the candidate set define the local causal neighborhood of the target variable, wherein the target variable measures the outcome of a predetermined physical phenomenon, and wherein the variables of the candidate set identify associated phenomena that cause or are caused by the predetermined physical phenomenon.
A computerized method for determining a local causal neighborhood of a target variable from a data set, the method comprising:
- during a first phase, for an initially empty set defining a candidate set and for every variable X contained in the data set,
- (a) finding a minimum associated set of X, s(X), which is subset of the candidate set and that minimizes a predetermined measure of association between X and the target variable T given the minimum associated set s(X);
- (b) determining a variable F, the variable F contained in the data set but not included in the candidate set and not being the target variable T, which maximizes an association between the variable F and the target variable given the minimum associated set s(F), and adding the variable F to the candidate set if and only if F is not statistically independent of the target variable T;
- (c) iteratively repeating steps (a) and (b) until after a last iteration the candidate set is equal to the candidate set after a next-to-last iteration;
- and
- during a second phase, for every variable Y contained in the candidate set,
- (d) determining whether there exists a subset t of the candidate set such that the variable Y is statistically independent of the target T given the subset t, and if so, then removing Y from the candidate set;
- and
- (e) identifying the remaining variables of the candidate set as the local causal neighborhood of the target variable T;
- wherein the remaining variables of the candidate set define the local causal neighborhood of the target variable, wherein the target variable measures the outcome of a predetermined physical phenomenon, and wherein the variables of the candidate set identify associated phenomena that cause or are caused by the predetermined physical phenomenon.
A computerized method for determining a local causal neighborhood of a target variable from a data set, the method comprising:
- (a) identifying a variable V contained within the data set that maximizes an association with the target variable T based upon a predefined measure of association and adding the identified variable to a candidate set defined to be an initially empty set;
- (b) identifying a variable X within the candidate set for which there exists a subset s of the candidate set such that the variable X is independent of the target set given the subset s, removing the variable X, and marking the variable X;
- and
- (c) repeating steps (a) and (b) for each unmarked variable until the candidate set contains no variable for which there exists a subset of the candidate set such that the a variable is independent of the target set given any subset of the candidate set;
- wherein the remaining variables of the candidate set define the local causal neighborhood of the target variable, wherein the target variable measures the outcome of a predetermined physical phenomenon, and wherein the variables of the candidate set identify associated phenomena that cause or are caused by the predetermined physical phenomenon.
Dependencies are inferred from the claim text.