Data Analysis Computer System and Method For Conversion Of Predictive Models To Equivalent Ones

摘要

The present invention addresses two ubiquitous and pressing problems of modern data analytics technology. Many modern pattern recognition technologies produce models with excellent predictivity but (a) they are “black boxes”, that is they are opaque to the user; (b) they are too large, and/or expensive to execute in less powerful computing platforms. The invention “opens up” a black box model by converting it to a compact and understandable model that is functionally equivalent. The invention also converts a predictive model into a functionally equivalent model into a form that can be implemented and deployed more easily or efficiently in practice. The benefits include: model understandability and defensibility of modeling. A particularly interesting application is that of understanding the decision making of humans, comparison of the behavior of a human or computerized decision process against another and use to enhance education and guideline compliance/adherence detection and improvement. The invention can be applied to practically any field where predictive modeling (classification and regression) is desired because it relies on extremely broad distributional assumptions that are valid in numerous fields.

主权项

1. A computer-implemented method and system for converting predictive models to equivalent models comprising the following steps:
a. learning a model M1 from dataset D; b. generating a model B1 of the distribution of input variables in D; c. generating input patterns from B1 or D using statistical sampling; d. creating new data D1 that comprises of the generated inputs followed by the corresponding M1 model-estimated outputs; e. deriving all or multiple Markov Boundaries (MB1, . . . , MBn) of the response variable by application of appropriately instantiated TIE* method on D1; f. learning from each Markov Boundary (i.e., MBi), an equivalent representation DTi that is easier to understand by humans or is easier to implement in practical application settings; g. verifying and fine tuning each equivalent model to capture the outputs of M1 within acceptable accuracy e; h. keeping only the Markov Boundaries that satisfy the condition of step g; and i. outputting the catalogue of all validated models DTi comprising the complete final set of equivalent explanations of the function contained in model M1.