Feature generation and model selection for generalized linear models

A computing system, comprising:

• a branch construction logic configured to:
• identify a dataset that stores values for a target attribute and input attributes identifying candidate features, where a candidate feature comprises a combination of one or more input attributes;
• compute respective inclusion scores for respective candidate features in the branch, based, at least in part on a likelihood that the candidate feature will be selected for inclusion in a generalized linear model that combines selected candidate features to predict a value of the target attribute;
• and
• construct a branch of candidate features ordered according to inclusion score;
• and
• a streamwise feature selection logic configure to construct the generalized linear model by performing acceptance testing on candidate generalized linear models that include a next candidate feature in the branch, where candidate generalized linear models are compared to a last accepted generalized linear model and a candidate generalized linear model is accepted when one or more acceptance criteria are met.

A non-transitory computer-readable medium storing computer-executable instructions that when executed by a computer cause the computer to perform a method, the method comprising:

• identifying a dataset that stores values for one or more input attributes and a target attribute;
• until one or more model termination criteria are met, constructing a set of branches, where each branch comprises candidate features ordered according to respective inclusion score, where a candidate features comprises a combination of one or more input attributes, further where respective inclusion scores estimate a likelihood that respective candidate features will be selected for inclusion in a generalized linear model that comprises a combination of features that predict the target attribute;
• until one or more branch termination criteria are met for each branch in the set of branches:
• constructing a candidate generalized linear model that includes a next candidate feature in the branch;
• performing acceptance testing on the candidate generalized linear model such that when the candidate generalized linear model meets acceptance criteria, the candidate generalized linear model is accepted;
• when one or more reorder criteria are met, computing updated inclusion scores for remaining candidate features in the branch based, at least in part, on a correlation between respective candidate features and a residual error of the last accepted generalized linear model;
• re-ordering the remaining candidate features in the branch according to updated inclusion scores;
• and
• when branch termination criteria are met for the branch, accessing a next branch in the set of branches;
• and
• providing the last accepted generalized linear model as an output.