How LingXi Achieves User‑Level QoE Optimization in Large‑Scale Adaptive Video Streaming

Paper Overview

Fastly and Tsinghua University’s Sun Lifeng team published a paper titled Towards User‑level QoE: Large‑scale Personalized Optimization of Adaptive Video Streaming , accepted to ACM SIGCOMM 2025. The work presents LingXi, the first system that can be deployed at massive scale to provide user‑specific adaptive video streaming optimization.

Background and Motivation

Traditional adaptive streaming focuses on system‑level QoS metrics (bitrate, smoothness) but these improvements often do not translate into better user experience. Large‑scale A/B tests showed that higher QoS did not increase total watch time. Analysis of millions of playback traces identified video stalls (rebuffering) as the dominant factor affecting user exit rates, with an impact order of magnitude larger than video quality or smoothness.

Key Findings

Conclusion 1: Stalls are the primary negative QoE factor; personalized modeling must prioritize stall response.

Conclusion 2: Users exhibit stable, significant individual differences in stall tolerance, providing a solid basis for per‑user optimization.

LingXi System Architecture

LingXi consists of three tightly coupled components that operate on each user independently while remaining compatible with any existing ABR algorithm.

Online Bayesian Optimization (OBO): For each user, a Gaussian‑process surrogate model is built from historical “parameter‑experience” pairs. An acquisition function balances exploration and exploitation to continuously update the optimal ABR parameters (e.g., stall‑penalty weight).

Monte‑Carlo Sampling: Given a candidate parameter set, the system simulates many virtual playback sessions using a bandwidth model derived from the user’s past network conditions. An exit‑rate predictor evaluates each segment, and the average exit probability estimates long‑term QoE impact.

Hybrid Exit‑Rate Predictor: A neural network (see Figure 5) predicts user exit probability when a stall occurs, using short‑term playback state (bitrate, throughput, stall duration) and long‑term user history (previous stall intervals). For non‑stall scenarios, a statistical model based on aggregate logs provides predictions.

The modular design allows LingXi to be inserted into existing transmission pipelines without interfering with playback, using only natural viewing behavior as implicit feedback.

Experimental Validation

A 10‑day, large‑scale A/B test on the Kuaishou platform compared LingXi against a highly tuned baseline ABR algorithm. Results (Figure 6) showed statistically significant improvements in:

Total watch time (QoE)

Average video bitrate (QoS)

Total stall duration (QoS)

Low‑bandwidth users (<2000 kbps) experienced up to a 15 % reduction in stall duration, confirming the system’s ability to adapt parameters based on individual stall sensitivity.

Further analysis (Figure 8) revealed a strong negative correlation between a user’s measured stall‑exit rate and the aggressiveness of the ABR parameters assigned by LingXi, demonstrating effective “one‑size‑fits‑all” personalization.

Conclusion

LingXi represents a paradigm shift from static, system‑level optimization to millions of dynamic, user‑specific goals. By integrating online Bayesian optimization, Monte‑Carlo simulation, and a hybrid exit‑rate predictor, the system delivers measurable QoE and QoS gains at production scale, especially for users on weak networks. The work establishes a practical pathway for large‑scale, personalized video streaming optimization.