Graph Network Data in Trend‑Driven Video Production on Kuaishou: Concepts, Applications, and Frequency Illusion

Speaker : Li Jianwei, Data Analysis Expert at Kuaishou (edited by Wang Liuyue, Shanghai University of International Business and Economics; produced by DataFunTalk).

Overview : The rapid production and sharing of short videos on Kuaishou generates massive high‑heat content that spreads virally; graph structures naturally model social relationships, making graph network analysis essential for understanding and influencing this diffusion.

01 Graph Network Basic Concepts Overview

1. Essential Elements : Nodes represent users or entities and carry both intrinsic attributes (e.g., user profile, activity) and network‑derived attributes (e.g., number of connections). Edges represent relationships and may have weight (strength) and direction (information flow).

In practice, multiple network types (small‑world, scale‑free, random) may coexist as sub‑networks.

2. Key Nodes : Central nodes maintain network stability. Their definition varies by structure (e.g., high degree, betweenness). Common centrality metrics are used to identify them.

3. Basic Classifications : Directed graphs model one‑way follows; undirected graphs model mutual follows; weighted graphs capture interaction intensity.

4. Graph Network Applications :

Graph Neural Networks – used for social influence prediction, electronic health records, etc.

Graph‑based Causal Inference – detailed later.

Community Detection – algorithms partition messy graphs into sub‑communities linked by central nodes.

02 Current Use of Network Data at Kuaishou

1. Community Formation : Individual users are points; their “broad social” ties create a social graph that can be segmented into distinct communities.

2. From Disorder to Order : Using community detection algorithms such as Louvain, chaotic networks become ordered, revealing key nodes that bridge communities; different strategies can be applied to each community.

3. Offline‑to‑Online Migration : Companies aim to migrate offline friendships to the platform and use them for cold‑start recommendations.

Different user groups belong to different verticals; key nodes (mediators) can connect communities, enabling cross‑community influence (e.g., encouraging two streamers to co‑broadcast).

03 Practice of Network Data in Trend‑Driven Shooting

1. Wind and Trend : A trend must be easy to imitate, low‑cost to produce, entertaining, and mass‑appealing. New works can achieve cold‑start without platform push.

2. Trend Carriers : Hashtags, music, magical stickers, templates.

3. Key Nodes in the Network :

Four interpretable central node types are defined:

Super‑Spread Node : Highest impact, influences many first‑degree users, leading to strong, deep propagation.

Central Spread Node : Directly influences first‑degree users with strong intensity but shallow depth.

Seed Spread Node : Low direct influence but connects to crucial nodes (super or central), yielding low intensity but deep reach.

Low‑Efficiency Node : Minor impact, usually casual content creators.

4. Strategies for Key Nodes : Targeted production incentives (e.g., trend videos) to “precisely guide” influential users.

Specific recall and ranking weight adjustments.

Honor‑based incentive task system.

Directed manual operations.

5. Tool‑Enabled Business : Beyond algorithmic distribution, operational interventions (hot‑topic mining, promotion, ranking) boost community participation (likes, comments, shares, searches).

04 Using Key Nodes to Create Frequency Illusion

1. What Is Frequency Illusion? Also known as the Baader‑Meinhof phenomenon, it is a cognitive bias where after noticing something, one perceives it occurring more often, leading to an overestimation of its frequency.

2. Everyday Examples : Information asymmetry causes people to rely on personal perception, creating biased estimates of event probabilities.

3. Impact and Value : Repeated exposure to specific content increases the likelihood of users creating similar videos, amplifying trend propagation.

Graph‑based causal inference can test whether the illusion truly drives production or is confounded by variables such as user profile, behavior, holidays, etc.

When performing causal inference, selecting confounders X is crucial; features include user profile, consumption behavior, temporal trends, and production achievements.

Using a Double‑Machine‑Learning (DML) model, the study finds that frequency illusion significantly boosts user production across regions.

4. How to Induce Frequency Illusion : Based on conditional probability P(Follow‑shoot|Consume X) = P(Follow‑shoot ∧ Consume X) / P(Consume X), the following tactics are recommended:

Weight content from super‑spread nodes.

Remove dispersion strategies for specific themes.

Pre‑heat operations to increase user impression.

Highlight hot material in product UI.

Thank you for listening.