Automatic Discovery of High-Performance Features for Customer Lifetime Value Optimization via Low-Variance Random Projection

摘要

Techniques for automatic discovery of high-performance features for customer LTV optimization via low-variance random projection are described. In one or more implementations, a random projection matrix is generated that is usable to compress a dataset representing a plurality of features associated with one or more customers. Using a first subset of the plurality of features, a simulator is created to model customer behavior. In addition, a policy is trained to determine which advertisements to present to a new customer based on a second subset of the plurality of features. In implementations, the policy is trained by at least using the random projection matrix to compress the second subset of the plurality of features. Subsequently, a performance of the policy is evaluated using the simulator to determine a level of the performance of the policy. This process is repeated a number of times in order to evaluate several possible candidate transformations and compressions of the dataset, with the goal of autonomously discovering and identifying a new compressed set of high-performing features for use in LTV learning algorithms.

主权项

1. A computer-implemented method, comprising:
generating a random projection matrix that is usable to compress a dataset representing a plurality of features associated with one or more customers; creating a simulator to model customer behavior based on a first subset of the plurality of features; training a policy to determine which advertisements to present to a new customer based on a second subset of the plurality of features, the policy being trained by at least using the random projection matrix to compress the second subset of the plurality of features; and evaluating a performance of the policy using the simulator to determine a level of the performance of the policy.