How MaFengWo Quantifies User Content Contribution with a Three‑Factor Model

Introduction

In the era of personalized experiences, vertical and refined operations are a key competitive edge. A complete user portrait helps companies discover each user’s behavior traits, potential abilities, and interests, enabling targeted services.

MaFengWo, with massive travel‑experience data, explores how to mine basic user characteristics, travel‑theme preferences, and latent interests from UGC, allowing precise user tagging and linking high‑quality content, products, and services.

Why Mining User Content Contribution Matters

Encouraging users to share original content and learn from each other is core to MaFengWo’s growth. Contributions include travel guides, diaries, Q&A, and reviews, forming an interactive ecosystem that helps users complete travel‑decision loops.

Identifying travelers with rich solo‑trip experience and strong content‑creation ability informs strategies for content growth and user activation.

Relying solely on user levels to assess influence is problematic because levels only increase, do not decay, and cannot capture content‑output willingness.

User core output cannot be quantified : simple actions like check‑ins or comments gradually raise levels.

Level fixation after upgrade : inactive users may still appear influential.

Inability to sense content‑output intent : high‑level users may not show interest in specific topics.

Applying the Content Contribution Tag

The contribution ability tag is used in many MaFengWo products, such as targeted Q&A invitations and KOL discovery.

Travel Q&A Invitation

Active users who produce popular content are invited to answer specific travel questions, ensuring quick and accurate responses.

MaFengWo KOL Discovery

By leveraging the contribution tag, MaFengWo identifies active, professional travelers who generate high‑quality content, promoting them online and offline (e.g., “MaFengWo Guides”).

User Content Contribution Model

The model combines three dimensions: user activity, popularity, and sharing willingness.

User Content Contribution Ability = Output Willingness + Activity + Popularity

1. User Activity Model

Based on the classic RFM model, MaFengWo adapts the three factors:

A (Activity) : measured by e^(-αt), where t is days since last visit; α is set so that after one year the decay reaches 0.0001 (α≈0.0189).

F (Frequency) : frequency of content contributions (guides, Q&A, notes, etc.) within a period.

E (Engagement) : type‑based weight of the latest contribution (travel diary = 5, Q&A = 2.5, note = 3).

2. User Popularity

Popularity is derived from likes, comments, collections, and shares across different UGC types. MaFengWo calculates a composite score W using weighted sums of content‑type scores (α:β:χ ≈ 1:1.05:0.98, approximated to 1).

Logistic regression assigns weights to features such as likes, favorites, comments, and shares, producing a “Travel” score for diaries and analogous scores for Q&A and notes. Example weights (rounded):

Diary: w1=0.1, w2=0.5, w3=0.2, w4=0.4

Q&A: w1=0.2, w2=0.9, w3=0.3, w4=0.6

Note: w1=0.1, w2=0.5, w3=0.3, w4=0.6

3. User Sharing Willingness

Sharing willingness is modeled using the contribution tag and PageRank. The tag represents user interests; PageRank evaluates the relationship between topics and users, increasing the recommendation score for content that matches a user’s high‑willingness tags.

Formulas:

D = user’s content writing willingness

d_i = willingness for a specific content type (e.g., diary)

T_i = proportion of a specific content type in the user’s total shared content

C_i = number of high‑quality articles of a specific type contributed by the user

N = damping factor (default 0.85)

Conclusion

The contribution model addresses three gaps in traditional level‑based systems: lack of behavior‑type distinction, absence of time decay, and insufficient interest mining. It has been fully deployed in MaFengWo’s products.

Future work includes adding comment‑based dimensions, enriching influence metrics with content‑quality scores, and further refining the user tag ecosystem.