Non-converting publisher attribution weighting and analytics server and method

A method of a non-converting publisher attribution weighting and analytics server, comprising:

• determining ‘P’ number of publishing channels for advertisements in a marketing campaign for a set of purchasable items using a processor and a memory communicatively coupled with the processor,
• wherein a marketing effectiveness of the P publishing channels in generating converted users each with at least one of a desirable action and a purchase from the set of purchasable items in the marketing campaign are to be analyzed using the processor and the memory based on a set of marketing data from a data collection server in a cloud, and
• wherein the set of marketing data comprises ‘N’ number of data points of marketing data collected in the marketing campaign;
• determining that a particular data point among the N data points is associated with a user, the particular data point being a set comprising:
• a conversion value based on at least one of the desirable action, a first related action related to the desirable action, a second related action related to the purchase, an amount of the purchase, an amount of gained units of the purchase and a profit of the purchase,
• wherein the conversion value is non-negative, and
• wherein the conversion value is zero if the user is a non-converted user without the at least one of the desirable action and the purchase from the set of purchasable items, and
• at least one of ‘R’ number of types of marketing data associated with the user for each of the P publishing channels such that ‘T’ number of numerical quantities associated with the user are generated;
• constructing a ‘K’-th order attribution model where K is an integer greater than one, with the K-th order attribution model comprising:
• M number of independent variables comprising X_1, X_2, . . . , and X_M, and
• a dependent variable Y_1 to be estimated by a linear combination of more than one terms of the independent variables, comprising:
• a zero-th order constant term with a zero-th order model parameter,
• M number of first order linear terms each comprising one of the M independent variables weighted by one of M number of first order model parameters, and
• (M)(M−1)/2 number of second order cross terms each comprising a first function of two of the M independent variables weighted by one of (M)(M−1)/2 number of second order model parameters;
• associating the M independent variables of the K-th order attribution model with a set of M functions of the T numerical quantities associated with the user associated with the particular data point,
• wherein each of the set of M functions is at least one of a particular numerical quantity associated with a particular type of marketing data and an aggregation of the T numerical quantities associated with the particular type of marketing data;
• associating the dependent variable Y_1 of the K-th order attribution model with a second function of the conversion value associated with the user;
• determining a first observation matrix X1 of size N×M based on the set of marketing data and the K-th order attribution model:
• wherein N rows of the first observation matrix X1 to correspond to the N data points of the marketing data collected in the marketing campaign, and
• wherein M columns of the first observation matrix X1 to correspond to the set of M functions associated with the M independent variables of the K-th order attribution model applied to the T numerical quantities associated with data points of the set of marketing data;
• determining a first conversion vector Y 1 of size N×1, with N elements of the first conversion vector Y 1 being the second function of conversion values of the N data points of the marketing data collected in the marketing campaign,
• wherein all positive conversion values corresponding to a set of converted users associated with the N data points are not less than a positive threshold ‘TH1’;
• performing a first regression analysis based on the first observation matrix X1 and the first conversion vector Y 1 to find a first set of optimal model parameters of the K-th order attribution model that best estimate the dependent variable Y_1 by the linear combination of the more than one terms of the M independent variables with respect to a first goodness-of-fit measure,
• wherein the first regression analysis comprises an application of a gradient descent algorithm which is a minimization algorithm and a factorization machine algorithm to obtain a factorized matrix V of size M×L, where L is an integer not greater than M;
• performing a first refining step and a second refining step to identify a set of insignificant second order cross terms of the K-th order attribution model,
• wherein the first refining step to comprise, for a second order cross term with two independent variables X_i and X_j:
• computing a matrix Q=V*(V̂T) of size M×M which is the matrix multiplication of the factorized matrix V and its matrix transpose, and
• classifying the second order cross term with the two independent variables X_i and X_j as unimportant if (ij)-th element of matrix Q is less than a ‘TH3’ percentile in i-th row of the matrix Q in terms of absolute magnitude,
• wherein the second refining step to comprise, for the second order cross term with two independent variables X_i and X_j:
• classifying the second order cross term with the two independent variables X_i and X_j as unimportant if percentage of the converted users among the data points in which both X_i and X_j are collected is smaller than a threshold ‘TH4’, and
• wherein the second order cross term being classified as unimportant by at least one of the first refining step and the second refining step is identified as insignificant;
• constructing a second K-th order attribution model which is the K-th order attribution model minus the set of insignificant second order cross terms identified in the first refining step and the second refining step;
• construct a modified second K-th order attribution model by replacing the first function of two of the M independent variables in the second K-th order attribution model by a third function of two of the M independent variables;
• associating the M independent variables of the modified second K-th order attribution model with a second set of M functions of the T numerical quantities associated with the user;
• associating the dependent variable Y_1 of the modified second K-th order attribution model with a fourth function of the conversion value associated with the user;
• determining a second observation matrix X2 of size N×(M2) based on the set of marketing data, the modified second K-th order attribution model, and the second set of M functions:
• wherein M2 is not less than M,
• wherein M2 is not greater than M+(M)(M−1)/2,
• wherein M columns of the second observation matrix X2 to correspond to the M columns of the first observation matrix X1 with the set of M functions replaced by the second set of M functions, and
• wherein (M2−M) columns of the second observation matrix X2 to each correspond to the third function of two through K quantities in the second order cross terms of the modified second K-th order attribution model, each of the two quantities being one of the second set of M functions applied to the T numerical quantities associated with data points of the set of marketing data;
• determining a second conversion vector Y2 of size N×1 based on the marketing data:
• wherein the second conversion vector Y2 is the first conversion vector Y1 with the second function replaced by the fourth function;
• performing a second regression analysis based on the second observation matrix X2 and the second conversion vector Y2 to find a second set of optimal model parameters of the modified second K-th order attribution model that best estimate the dependent variable Y_1 by the linear combination of the more than one terms of the M independent variables with respect to a second goodness-of-fit measure;
• computing attribution scores associated with the P publishing channels based on the K-th order attribution model, the modified second K-th order attribution model, the first set of optimal model parameters and the second set of optimal model parameters;
• receiving a request through a network for the attribution scores associated with the P publishing channels from a marketer client communicatively connected to the processor and the memory through the network;
• and
• communicating the attribution scores associated with the P publishing channels to the marketer client through the network.

A method of a non-converting publisher attribution weighting and analytics server, comprising:

• determining ‘P’ number of publishing channels for advertisements in a marketing campaign for a set of purchasable items using a processor and a memory communicatively coupled with the processor,
• wherein a marketing effectiveness of the P publishing channels in generating converted users each with at least one of a desirable action and a purchase from the set of purchasable items in the marketing campaign are to be analyzed using the processor and the memory based on a set of marketing data from a data collection server in a cloud, and
• wherein the set of marketing data comprises ‘N’ number of data points of marketing data collected in the marketing campaign from time T1 to time T2;
• determining that a particular data point among the N data points is associated with a user, the particular data point being a set comprising:
• a conversion value based on at least one of the desirable action, a first related action related to the desirable action, a second related action related to the purchase, an amount of the purchase, an amount of gained units of the purchase and a profit of the purchase,
• wherein the conversion value is non-negative, and
• wherein the conversion value is zero if the user is a non-converted user without the at least one of the desirable action and the purchase from the set of purchasable items, and
• at least one of ‘R’ number of types of marketing data associated with the user for each of the P publishing channels such that ‘T’ number of numerical quantities associated with the user are generated;
• constructing a ‘K’-th order attribution model where K is an integer greater than one, with the K-th order attribution model comprising:
• M number of independent variables comprising X_1, X_2, . . . , and X_M, and
• a dependent variable Y_1 to be estimated by a linear combination of more than one terms of the independent variables, comprising:
• a zero-th order constant term with a zero-th order model parameter,
• M number of first order linear terms each comprising one of the M independent variables weighted by one of M number of first order model parameters, and
• (M)(M−1)/2 number of second order cross terms each comprising a first function of two of the M independent variables weighted by one of (M)(M−1)/2 number of second order model parameters,
• wherein the first function of the two independent variables is the multiplication of the two independent variables;
• associating the M independent variables of the K-th order attribution model with a set of M functions of the T numerical quantities associated with the user associated with the particular data point,
• wherein each of the M functions is at least one of a particular numerical quantity associated with a particular type of marketing data and a collection of numerical quantities associated with the particular type of marketing data to provide a flexibility to internally perform aggregations and other calculations;
• associating the dependent variable Y_1 of the K-th order attribution model with a second function of the conversion value associated with the user,
• wherein the second function of the conversion value is the conversion value;
• determining a first observation matrix X1 of size N×M based on the marketing data and the K-th order attribution model:
• wherein N rows of the first observation matrix X1 to correspond to the N data points of the marketing data collected in the marketing campaign, and
• wherein M columns of the first observation matrix X1 to correspond to the set of M functions associated with the M independent variables of the K-th order attribution model applied to the T numerical quantities associated with the N data points of the set of marketing data;
• determining a first conversion vector Y 1 of size N×1, with N elements of the first conversion vector Y 1 being conversion values of the N data points of the marketing data collected in the marketing campaign,
• wherein all positive conversion values corresponding to a set of converted users associated with the N data points are not less than a positive threshold ‘TH1’;
• performing a first regression analysis based on the first observation matrix X1 and the first conversion vector Y 1 to find a first set of optimal model parameters of the K-th order attribution model that best estimate the dependent variable Y_1 by the linear combination of the more than one terms of the independent variables with respect to a first goodness-of-fit measure,
• wherein the first regression analysis comprises the application of a gradient descent algorithm which is a minimization algorithm and a factorization machine algorithm to obtain a factorized matrix V of size M×L, where L is an integer not greater than M;
• performing at least one refining step to identify a set of insignificant second order cross terms of the K-th order attribution model,
• wherein the at least one refining step to comprise:
• computing a matrix Q=V*(V̂T) of size M×M which is the matrix multiplication of the factorized matrix V and its matrix transpose,
• compute a row mean rowMean_Q and a row standard deviation rowSD_Q for each row of matrix Q, and
• identifying all elements in the row of matrix Q that are less than a row-adaptive threshold ‘TH3’ as insignificant, where TH3=rowMean_Q+C*rowSD_Q is the row mean plus C times row standard deviation, wherein C is greater than 0.0 and less than <3, and
• wherein the at least one refining step to further comprise, for each of the second order cross term with two independent variables:
• identifying all of ‘A’ number of data points in which the two independent variables are collected,
• determining ‘B’ number of converted users among the A data points in which the two independent variables are collected, and
• identifying the second order cross term with the two independent variables as insignificant if the fraction B/A is less than a threshold ‘TH4’;
• constructing a second K-th order attribution model which is the K-th order attribution model minus the set of insignificant second order cross terms identified in the first refining step and the second refining step;
• construct a modified second K-th order attribution model by replacing the first function of two of the M independent variables in the second K-th order attribution model by a third function of two of the M independent variables;
• associating the M independent variables of the modified second K-th order attribution model with a second set of M functions of the T numerical quantities associated with the user;
• associating the dependent variable Y_1 of the modified second K-th order attribution model with a fourth function of the conversion value associated with the user;
• determining a second observation matrix X2 of size N×(M2) based on the set of marketing data, the modified second K-th order attribution model, and the second set of M functions:
• wherein M2 is not less than M,
• wherein M columns of the second observation matrix X2 to correspond to the M columns of the first observation matrix X1 with the set of M functions replaced by the second set of M functions, and
• wherein (M2−M) columns of the second observation matrix X2 to each correspond to the third function of two quantities in the second order cross terms of the modified second K-th order attribution model, each of the two quantities being one of the second set of M functions applied to the T numerical quantities associated with data points of the set of marketing data;
• determining a second conversion vector Y2 of size N×1 based on the marketing data:
• wherein the second conversion vector Y2 is the first conversion vector Y1 with the second function replaced by the fourth function;
• performing a second regression analysis based on the second observation matrix X2 and the second conversion vector Y2 to find a second set of optimal model parameters of the modified second K-th order attribution model that best estimate the dependent variable Y_1 by the linear combination of the more than one terms of the independent variables with respect to a second goodness-of-fit measure;
• computing attribution scores associated with the P publishing channels based on the K-th order attribution model, the modified second K-th order attribution model, the first set of optimal model parameters and the second set of optimal model parameters;
• receiving a request through a network for the attribution scores associated with the P publishing channels from a marketer client communicatively connected to the processor and the memory through the network;
• and
• communicating the attribution scores associated with the P publishing channels to the marketer client through the network.