How LingXi Revolutionizes User‑Level QoE with Scalable Adaptive Video Streaming

1. Background: From Traditional QoS to Personalized QoE

Personalized QoE optimization for video streaming has attracted extensive research, yet existing methods struggle in large‑scale production due to intrusive explicit user ratings, bandwidth‑based interventions that degrade experience, discontinuous optimizations, and poor scalability.

2. System Design: LingXi Overview

The LingXi system is designed to overcome these limitations and provide a deployable, sustainable, non‑intrusive user‑level QoE optimization framework. Table 1 (illustrated below) contrasts LingXi with prior approaches.

2.1 System‑level QoS Optimization Bottlenecks

Large‑scale online A/B tests comparing three QoS‑oriented algorithms (baseline, video‑quality‑first, stall‑reduction‑first) showed no statistically significant improvement in total watch time, demonstrating that improving system‑level QoS alone no longer translates into real user‑experience gains.

2.2 Identifying Key QoE Factors: Focus on Stalls

Analysis of millions of playback traces revealed that stall duration has an impact magnitude up to 10⁻¹, far exceeding video quality (10⁻³) and smoothness (10⁻²). Consequently, stalls are the dominant negative factor for QoE.

2.3 Personalized Optimization Space: User‑level "Thousand‑Faces"

Users exhibit significant, stable, and dynamic differences in stall tolerance. This provides a solid theoretical basis for user‑level personalization.

3. Algorithm Design: Core Components of LingXi

LingXi is not a new ABR algorithm; it is a modular framework compatible with any existing ABR. Its three core components are:

Online Bayesian Optimization (OBO) : Treats the unknown relationship between ABR parameters and user QoE as a black‑box and continuously searches for per‑user optimal parameters using Gaussian‑process surrogate models and acquisition‑function maximization.

Monte Carlo Sampling : For each candidate parameter, simulates multiple virtual playback sessions based on a bandwidth model derived from the user’s history, evaluates exit probabilities with a predictor, and aggregates results to estimate long‑term QoE impact.

Hybrid Exit‑Rate Predictor : A neural network that predicts user exit probability when a stall occurs, using short‑term playback state (bitrate, throughput, stall length) and long‑term user state (historical stall intervals). For non‑stall scenarios, a statistical model based on aggregate logs is used.

4. Experimental Evaluation: Large‑scale A/B Tests

A 10‑day A/B experiment on the Kuaishou platform compared LingXi with a highly optimized baseline ABR. Results show simultaneous improvements in total watch time (QoE), average video bitrate (QoS), and total stall duration (QoS). Notably, low‑bandwidth users (<2000 kbps) experienced a ~15 % reduction in stall time, confirming the system’s ability to adapt parameters dynamically based on user sensitivity.

5. Conclusion

LingXi marks a paradigm shift from static, system‑level QoS targets to thousands of dynamic, user‑specific QoE objectives. By integrating online Bayesian optimization, Monte Carlo simulation, and a hybrid exit‑rate predictor, the system delivers measurable QoE and QoS gains at massive scale, especially for weak‑network users, and validates the feasibility of true "thousand‑faces" personalization in adaptive video streaming.