Parallelization techniques for variable selection and predictive models generation and its applications

A system (100) comprising:

• a memory (102) storing instructions;
• one or more communication interfaces (106);
• and
• one or more hardware processors (104) coupled to the memory (102) via the one or more communication interfaces (106), wherein the one or more hardware processors (104) are configured by the instructions to:
• define (i) an input data comprising a set of independent variables, a set of dependent variables, and a subset size, (ii) a set of control parameters comprising maximum allowed inter correlation coefficient of independent variables Rint, and a minimum or maximum objective function threshold (Rcri) pertaining to the set of independent variables, and (iii) a block size and a grid size pertaining to a graphics processing unit (GPU), the block size and grid size are indicative of number of execution threads to be executed in one or more blocks in a grid;
• rank, using a lexicographical technique, one or more subsets of independent variables from the set of independent variables;
• assign the one or more subsets ranked to corresponding one or more execution threads of the GPU;
• execute, in parallel, in each of the one or more execution threads specific to the one or more blocks in the grid:
• generate, using an unranking technique, an initial set of subsets from the one or more subsets assigned, and recursively generating, based on the initial set of subsets, a next set of subsets based on the one or more subsets assigned, wherein when the initial set of subsets is being generated, a comparison of an inter correlation coefficient (Ra) across independent variables in a subset of the initial set is performed;
• perform a comparison of (i) an inter correlation coefficient (Ra) of (a) an updated independent variable in a next subset of a subsequent set and (b) one or more independent variables in the same next subset and (ii) the maximum allowed inter correlation coefficient of independent variables Rint, reduces the comparison and memory read operation, wherein the updated independent variable in the next subset is obtained upon recursively generating one or more next subsets from the initial set of subsets;
• determine, based on the comparison, an optimal subset from each corresponding execution thread from each block to obtain one or more optimal subsets, each of the one or more optimal subsets comprises one or more independent variables, and
• synchronize, for the grid and the block, the one or more optimal subsets from each of the corresponding execution thread specific to each block to obtain a best subset from the synchronized one or more optimal subsets for variable selection, wherein values of the one or more optimal subsets are stored in a shared memory;
• generate, based on the comparison, one or more predictive models for the best subset;
• and
• calculate model correlation coefficient Rm based on the one or more generated predictive models and one or more dependent variables from the set of dependent variables, wherein the one or more dependent variables are stored in the shared memory, wherein the model correlation coefficient Rm comprise standard error, mean square error and variance, and wherein the processor implemented method eliminates Leave One Out (LOO) cross-validation performed after the one or more predictive models generated for obtaining the best subset.

One or more non-transitory machine readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors causes:

• defining, by the one or more hardware processors, (i) an input data comprising a set of independent variables, a set of dependent variables, and a subset size, (ii) a set of control parameters comprising maximum allowed inter correlation coefficient of independent variables Rint, and a minimum or maximum objective function threshold (Rcri) pertaining to the set of independent variables, and (iii) a block size and a grid size pertaining to a graphics processing unit (GPU), the block size and grid size are indicative of number of execution threads to be executed in one or more blocks in a grid;
• ranking, using a lexicographical technique, one or more subsets of independent variables from the set of independent variables;
• assigning the one or more subsets ranked to corresponding one or more execution threads of the GPU;
• executing, in parallel, in each of the one or more execution threads specific to the one or more blocks in the grid:
• generating, using an unranking technique, an initial set of subsets from the one or more subsets assigned, and recursively generating, based on the initial set of subsets, a next set of subsets based on the one or more subsets assigned, wherein when the initial set of subsets is being generated, a comparison of an inter correlation coefficient (Ra) across independent variables in a subset of the initial set is performed;
• performing a comparison of (i) an inter correlation coefficient (Ra) of (a) an updated independent variable in a next subset of a subsequent set and (b) one or more independent variables in the same next subset and (ii) the maximum allowed inter correlation coefficient of independent variables Rint, reduces the comparison and memory read operation, wherein the updated independent variable in the next subset is obtained upon recursively generating one or more next subsets from the initial set of subsets;
• determining, based on the comparison, an optimal subset from each corresponding execution thread from each block to obtain one or more optimal subsets, each of the one or more optimal subsets comprises one or more independent variables;
• and
• synchronizing, for the grid and the block, the one or more optimal subsets from each of the corresponding execution thread specific to each block to obtain a best subset from the synchronized one or more optimal subsets for variable selection, wherein values of the one or more optimal subsets are stored in a shared memory;
• generating, based on the comparison, one or more predictive models for the best subset;
• and
• calculating model correlation coefficient Rm based on the one or more generated predictive models and one or more dependent variables from the set of dependent variables, wherein the one or more dependent variables are stored in the shared memory, wherein the model correlation coefficient Rm comprise standard error, mean square error and variance, and wherein the processor implemented method eliminates Leave One Out (LOO) cross-validation performed after the one or more predictive models generated for obtaining the best subset.