CIKM’24 | OptDst: Learning the Optimal Distribution for LTV Estimation

Title: OptDst: Learning Optimal Distribution for Customer Lifetime Value Prediction

Address: https://arxiv.org/pdf/2408.08585

Meeting: CIKM 2024

1. Introduction

alt text

2.1 Distributed Learning Module DLM

According to the overall loss function introduced in section 2.1, it is necessary to consider the ziln loss of SDN and the mask vector generated by DSM. The DLM module will update the SDN parameters, and the DSM will also update the selection strategy accordingly, making optimization difficult and suboptimal. In this section, the author proposes an alignment mechanism inspired by meta pseudo labels.

alt text

In OptDst, each SDN in the DLM module is trained using the data allocated to that sub distribution by DSM. However, during the optimization process, the two separate sets of parameters in OptDst will interfere with each other. Meanwhile, DSM lacks clear supervisory signals, making it difficult to align with the output of DLM. Relying solely on…