New Framework Fixes Value Forecasting in Recommender Systems
PIT-SUN improves the accuracy of high-value metrics like LTV and GMV by solving a fundamental mathematical conflict in target transformation.
Predicting high-value metrics such as gross merchandise volume (GMV) and lifetime value (LTV) often fails because the data is heavy-tailed or zero-inflated. Standard mean squared error (MSE) leads to mean collapse and tail shrinkage in these environments. While transforming the target data can stabilize these gradients, nonlinear transforms typically lose expectation consistency, meaning the model cannot accurately recover the original value after inversion.
https://arxiv.org/abs/2607.08202
The PIT-SUN framework resolves this by using an empirical marginal table to define a bounded normal-score coordinate and an inverse-quantile lookup. Rather than directly inverting transformed predictions, it applies a multiplicative recovery process to estimate the original-space expectation. This approach allows the system to compress tails without sacrificing the mathematical consistency required for accurate value forecasting.
Industrial deployment and tests on large-scale datasets show robust improvements in calibration, ranking quality, and point accuracy. The framework achieves these gains with lightweight deployment overhead, making it a viable upgrade for production recommender systems targeting financial outcomes.