EXTRACTING RELEVANT FEATURES FROM ELECTRONIC MARKETING DATA FOR TRAINING ANALYTICAL MODELS

摘要

In certain embodiments, an analytical application accesses electronic marketing data that is automatically generated by interactions with marketing communications. The analytical application represents the set of electronic marketing data as a data matrix, in which columns of the matrix correspond to features of the data set. The analytical application selects a constraint for a singular value decomposition of the initial matrix and performs the singular value decomposition with the constraint. The constrained singular value decomposition derives, from the initial matrix, a matrix of singular vectors having a threshold number of rows with non-zero coefficients. The analytical application identifies certain columns from the initial matrix that correspond to the rows of the derived matrix with the non-zero coefficients and selects the features corresponding to those columns. The analytical application trains the analytical model using the selected features.

主权项

1. A method for efficiently using computing resources for training automated analytical models by identifying relevant features of data obtained from an online marketing campaign, the method comprising:
accessing, by a processor, a data set comprising features of electronic marketing data, wherein the features have values that are automatically generated by a plurality of user interactions with electronic marketing communications; selecting, by the processor, a subset of the features so that less than all of the data set is used to train an analytical model, wherein selecting the subset of features comprises:
identifying a size of the subset of the features,organizing the data set into an initial matrix, wherein columns of the initial matrix corresponds to respective features of the data set,selecting, based on the identified size of the subset, a constraint for a singular value decomposition of the initial matrix, wherein using the constraint in the singular value decomposition results in a matrix of singular vectors that is derived from the initial matrix to have fewer than a threshold number of rows with non-zero coefficients,deriving the matrix of singular vectors by performing the singular value decomposition with the constraint, andselecting the subset of features from the data set based on identifying a subset of columns from the initial matrix corresponding to the rows of the derived matrix with the non-zero coefficients; and training, by the processor, the analytical model using data from the data set corresponding to the selected subset of features.