How Probability-Driven Adaptive Streaming Won the ACM MM 2022 Grand Challenge

Background

At ACM Multimedia 2022 in Lisbon, the Kuaishou audio‑video technology team won the Grand Challenge: Short Video Streaming with their paper “PDAS: Probability‑Driven Adaptive Streaming for Short Video”. Over a hundred teams entered, and more than 20 reached the finals.

The challenge evaluated algorithms on Quality of Experience (QoE) and bandwidth waste, selecting the solution with the highest QoE and lowest waste.

Challenges in Short‑Video Multi‑Bitrate Streaming

Balancing pre‑load data volume to avoid QoE degradation while limiting bandwidth waste.

Determining pre‑load order to maximize bandwidth utilization without causing playback stalls.

Selecting pre‑load bitrate that accounts for user behavior, avoiding high‑bitrate chunks that are never watched.

PDAS Algorithm Design

Statistical User‑Behavior Probability Model : Quantifies user retention at different timestamps to model QoE and waste.

User‑Behavior‑Aware Pre‑load Model : Jointly decides maximum buffer length based on user behavior and network conditions, stopping downloads when appropriate.

Experience‑and‑Cost‑Aware Multi‑Bitrate Strategy : Optimizes bitrate decisions by modeling short‑video QoE, accounting for quality changes, user sliding behavior, and stall duration.

The algorithm computes a maximum buffer that incorporates user retention probability, playback order, and bandwidth, reducing waste for low‑retention videos and improving QoE for high‑retention ones.

Experimental Evaluation

Using the ACM MM 2022 open‑source simulation platform, PDAS outperformed two baseline algorithms and showed advantages over two ablation variants.

Compared to baselines, PDAS achieved up to 22.8% QoE improvement and up to 22.34% bandwidth cost reduction.

Conclusion

By integrating a statistical user‑behavior model, a behavior‑aware pre‑load mechanism, and an experience‑and‑cost‑aware bitrate strategy, PDAS secured the first place in all three evaluation rounds of the Grand Challenge, demonstrating strong generalization and practical applicability.