Modeling consumer activity

An apparatus comprising:

• a plurality of processor computational resources;
• a memory storing code executable by the plurality of processor computational resources to perform:
• generating potential model types;
• dividing activity data into a training data set, a test data set, and a validation data set;
• training weights for the potential model types with the training set data;
• selecting a model type that minimizes a cost function wherein the weights are trained against the validation data set and the cost function is evaluated against the test data set;
• dividing the activity data into batches based on time interval;
• calculating a cost learning weight vector for predicting costs for a consumer application by a consumer as a function of consumer characteristics using the activity data, selected model type, and the algorithmic parameters, wherein the learning weight vectors are calculated by minimizing the cost function J where
• J
• =
• 1
• 2
• ⁢
• m
• ⁢
• ∑
• i
• =
• 1
• m
• ⁢
• (
• h
• i
• -
• y
• i
• )
• 2
• +
• DS
• ⁢
• ∑
• j
• =
• 1
• n
• ⁢
• (
• θ
• j
• -
• Θ
• j
• )
• 2
• ,
• m is a number of activity data samples in a current batch, n is a number of features, θj is the learning weight vector, Θj is the previous learning weight vector, hi is a learning hypothesis for the learning weight vector, yi is an activity data instance, D is a constant, and S is a sampling ratio;
• calculating an income learning weight vector for predicting income for a consumer application from a consumer as a function of the consumer characteristics using the cost function J;
• generating a current lifetime value group correlation model for a plurality of consumers using the income learning weight vector and activity data in a current batch that has not been incorporated into the learning hypothesis for the income learning weight vector as inputs instead of the activity data from previous batches, wherein the current lifetime value group correlation model is used in place of activity data;
• generating a current cost per install group correlation model for installations of the consumer application for the plurality of consumers using the cost learning weight vector and activity data in the current batch that has not been incorporated into the learning hypothesis for the cost learning weight vector as inputs instead of the activity data from previous batches, wherein the current cost per install group correlation model is used in place of activity data;
• excluding mutually exclusive previous lifetime value group correlation models;
• subtracting a previous income learning weight vector from the lifetime value group correlation model to penalize divergent outputs from learned outputs;
• excluding mutually exclusive previous cost per install group correlation models;
• subtracting a previous cost learning weight vector from the cost per install group correlation model to penalize divergent outputs from the learned outputs;
• determining a row target of rows of the activity data processed per time interval as a function of a ratio of a number of unprocessed rows of the activity data and a rate of receiving the rows of the activity data per the time interval;
• allocating the plurality of processor computational resources to achieve the row target;
• and
• predicting a return on investment for each consumer based on the current and previous lifetime value group correlation models and the current and previous cost per install group correlation models.

A program product comprising a non-transitory computer readable storage medium storing code executable by a plurality of processor computational resources to perform:

• generating potential model types;
• dividing activity data into a training data set, a test data set, and a validation data set;
• training weights for the potential model types with the training set data;
• selecting a model type that minimizes a cost function wherein the weights are trained against the validation data set and the cost function is evaluated against the test data set;
• dividing the activity data into batches based on time interval;
• calculating a cost learning weight vector for predicting costs for a consumer application by a consumer as a function of consumer characteristics using the activity data, selected model type, and algorithmic parameters, wherein the learning weight vectors are calculated by minimizing the cost function J where
• J
• =
• 1
• 2
• ⁢
• m
• ⁢
• ∑
• i
• =
• 1
• m
• ⁢
• (
• h
• i
• -
• y
• i
• )
• 2
• +
• DS
• ⁢
• ∑
• j
• =
• 1
• n
• ⁢
• (
• θ
• j
• -
• Θ
• j
• )
• 2
• ,
• m is a number of activity data samples in a current batch, n is a number of features, θj is the learning weight vector, Θj is the previous learning weight vector, hi is a learning hypothesis for the learning weight vector, yi is an activity data instance, D is a constant, and S is a sampling ratio;
• calculating an income learning weight vector for predicting income for a consumer application from a consumer as a function of the consumer characteristics using the cost function J;
• generating a current lifetime value group correlation model for a plurality of consumers using the income learning weight vector and activity data in a current batch that has not been incorporated into the learning hypothesis for the income learning weight vector as inputs instead of the activity data from previous batches, wherein the current lifetime value group correlation model is used in place of activity data;
• generating a current cost per install group correlation model for installations of the consumer application for the plurality of consumers using the cost learning weight vector and activity data in the current batch that has not been incorporated into the learning hypothesis for the cost learning weight vector as inputs instead of the activity data from previous batches, wherein the current cost per install group correlation model is used in place of activity data;
• excluding mutually exclusive previous lifetime value group correlation models;
• subtracting a previous income learning weight vector from the lifetime value group correlation model to penalize divergent outputs from learned outputs;
• excluding mutually exclusive previous cost per install group correlation models;
• subtracting a previous cost learning weight vector from the cost per install group correlation model to penalize divergent outputs from the learned outputs;
• determining a row target of rows of the activity data processed per time interval as a function of a ratio of a number of unprocessed rows of the activity data and a rate of receiving the rows of the activity data per the time interval;
• allocating the plurality of processor computational resources to achieve the row target;
• and
• predicting a return on investment for each consumer based on the current and previous lifetime value group correlation models and the current and previous cost per install group correlation models.