A Long-term Value Prediction Framework In Video Ranking

Accurately modeling long-term value (LTV) at the ranking stage of short-video recommendation remains challenging. While delayed feedback and extended engagement have been explored, fine-grained attribution and robust position normalization at billion-scale are still underdeveloped. We propose a practical ranking-stage LTV framework addressing three challenges: position bias, attribution ambiguity, and temporal limitations. (1) Position bias: We introduce a Position-aware Debias Quantile (PDQ) module that normalizes engagement via quantile-based distributions, enabling position-robust LTV estimation without architectural changes. (2) Attribution ambiguity: We propose a multi-dimensional attribution module that learns continuous attribution strengths across contextual, behavioral, and content signals, replacing static rules to capture nuanced inter-video influence. A customized hybrid loss with explicit noise filtering improves causal clarity. (3) Temporal limitations: We present a cross-temporal author modeling module that builds censoring-aware, day-level LTV targets to capture creator-driven re-engagement over longer horizons; the design is extensible to other dimensions (e.g., topics, styles). Offline studies and online A/B tests show significant improvements in LTV metrics and stable trade-offs with short-term objectives. Implemented as task augmentation within an existing ranking model, the framework supports efficient training and serving, and has been deployed at billion-scale in Taobao’s production system, delivering sustained engagement gains while remaining compatible with industrial constraints.

💡 Research Summary

The paper tackles the practical problem of predicting long‑term value (LTV) for short‑video items at the ranking stage of a large‑scale recommender system. While most prior work focuses on immediate metrics (CTR, CVR, watch time) or on re‑ranking where listwise context can be modeled, the authors argue that ranking‑stage LTV is essential yet under‑explored due to billions of candidates and strict latency constraints. They identify three core challenges: (1) position bias – videos shown later in a feed tend to receive higher engagement simply because they are exposed to more active users; (2) attribution ambiguity – naïve accumulation of downstream watch time mixes causal influence with unrelated content; (3) limited temporal scope – session‑level signals ignore cross‑day re‑engagement driven by creators or topics.

To address these, the authors propose a three‑component framework that can be added as a task augmentation to an existing ranking model without architectural overhaul.

1. Position‑aware Debias Quantile (PDQ)
   PDQ treats slide‑time (the cumulative watch time of a video and its successors up to a horizon Q) as a target for quantile regression rather than raw regression. Training data are first partitioned into M “page groups” based on exposure position (e.g., page number in an infinite scroll). Within each group, isofrequency quantiles of slide‑time are computed using the inverse CDF. Each sample receives a quantile label y, which is a normalized value between 0 and 1. A simple MSE loss is then minimized on the predicted quantile. The method also handles zero‑slide cases by dynamically determining a starting quantile index S_k for each group, ensuring that censored low‑engagement samples do not distort the distribution. By moving the debiasing step into label generation, PDQ preserves positional information without adding extra model parameters or inference latency. Experiments show a 12 % reduction in slide‑time prediction error compared with baseline position‑aware baselines.

2. Multi‑dimensional Attribution Module
   The authors replace the naïve cumulative watch‑time metric with a weighted sum: S_j = Σ_{i>j} c_{ji}·t_i, where c_{ji}∈