A Long-term Value Prediction Framework In Video Ranking
Researchers tackle billion-scale position bias and attribution ambiguity to predict long-term user engagement.
Deep Dive
Researchers from Alibaba (Taobao) published a new AI framework for video ranking that predicts long-term user value (LTV). It introduces three key modules: PDQ for position bias, a multi-dimensional attribution module for causal clarity, and cross-temporal author modeling. Deployed at billion-scale on Taobao, it improves long-term engagement metrics while maintaining efficient training and serving, showing stable gains in online A/B tests.
Why It Matters
This directly improves recommendation quality for billions of users, moving beyond short-term clicks to foster sustained platform engagement.