Technical Analysis of “Chang'an” – The Beidou Star System for Reducing Content Uncertainty and Boosting Hit Potential

In this talk, senior Alibaba algorithm expert Cai Longjun (alias Muji) from Youku’s Content Intelligence team presents a technical case study of the hit TV series Chang’an Twelve Hours . The discussion focuses on how Youku leverages AI to increase the determinism of “blockbuster” content.

According to Youku’s Beidou Star metrics, the series achieved a daily index exceeding 1,000,000, roughly double the 500,000‑600,000 range of ordinary popular dramas.

The creation of a long‑form hit is fraught with uncertainty that permeates every stage—from script selection, casting, set and costume design, to shooting, post‑production, marketing, and distribution. Each step can significantly affect the final audience reception.

Three major sources of uncertainty are identified:

Delayed gratification and incomplete information: viewers receive only partial cues, leading to divergent interpretations.

Complex system engineering: many production roles and processes must be digitized, modeled, and quantified.

Content‑specific professional skills: technical choices in scene reconstruction, cinematography, costume, and composition must align with audience preferences and commercial value.

Youku addresses these challenges with a combination of NLP, computer vision, speech understanding, and knowledge graph (KG) technologies to perform “content outward deconstruction” and “core creation comprehension,” thereby extracting multi‑dimensional data from the content.

The Beidou Star system is described as a human‑machine hybrid brain that tackles uncertainty across the entire content lifecycle (acquisition, production, distribution, promotion, and playback). It provides a deterministic boost for hit generation.

Content creation understanding involves intelligent script analysis and mining. By converting traditional drama theory (e.g., two‑thousand‑year‑old theatrical values) into technical capabilities, Youku builds a “brain‑like” decision‑making layer that quantifies narrative elements.

Example statistics from Chang’an :

Over 120 characters appear in the script.

Character screen‑time distribution: Zhang Xiaojing 15%, Li Bi 10%, Tan Qi 5%, Long Bo 4%, Yao Ruyin 3%.

Zhang Xiaojing and Li Bi together account for >90% of character relationships.

Tan Qi contributes >80% of relationships, acting as a functional driver of plot.

Interaction analysis shows Zhang Xiaojing ↔ Tan Qi as the most frequent pair, followed by Li Bi ↔ Tan Qi/​Xu Bin. Compared with IP scripts, interactions between Zhang Xiaojing and Li Bi are reduced.

Emotion analysis reveals peaks in episodes 3 and 10 of the first 16 episodes, creating tension in the storyline.

Multimodal analysis of the first episode combines visual and audio cues (actor expressions, scenes, actions) to generate a predicted “viewer emotion curve,” which is later refined with real viewing data.

Predictive capability building focuses on quantifying uncertainty. A predictive model is constructed to assess the strength of uncertainty, improving decision‑making efficiency. Challenges include data scarcity/quality, model complexity, and cross‑domain generalization.

Solutions are organized into four layers:

Foundation layer – KG & domain knowledge, feature engineering, learning acceleration.

Data layer – synthetic minority oversampling (SMOTE), membership transformation, semi‑supervised learning.

Model layer – combination of DNN, Relation Net, and multi‑task learning (MTL) to reduce over‑fitting.

Uncertainty Learning – variational inference framework for predicting content uncertainty.

SMOTE (Synthetic Minority Oversampling Technique) is used to generate new minority samples, followed by membership transformation to evaluate similarity to real samples, yielding an approximate 5% efficiency gain.

Complex models are decomposed: DNN fits intricate factors, Relation Net handles external competition reasoning, and MTL integrates the components. Relation Net (2016) processes shape‑based objects via CNN, pairs them, combines with LSTM‑encoded questions, and outputs answer probabilities.

Uncertainty Learning leverages variational methods and Bayesian networks for probabilistic inference.

In the promotion stage, mining capabilities analyze minute‑by‑minute viewership, rewatch, and bullet‑screen data post‑release, aligning them with storyline segments to refine recommendation and promotion strategies.

The talk concludes with two future directions: (1) building a decision engine that fuses rule‑based logic, machine learning, Bayesian causal reasoning, and hybrid neural computation; (2) advancing quantitative psychology research to better understand audience preferences.

Overall, the presentation demonstrates how AI, big data, and uncertainty modeling can transform the highly unpredictable content industry into a more deterministic, data‑driven ecosystem.