Insight into Retention and Churn: Building User Behavior Prediction and Recommendation Optimization with Douyin API

2025-05-13

In today’s fast-evolving short video ecosystem, maintaining user engagement and stickiness is a critical challenge for content creators, MCNs, and brand marketers alike. Douyin users swipe through dozens of videos daily, but will they remain interested in a specific content type, creator, or brand over time? Can we anticipate changes in their preferences? Which users are silently slipping away? And how can we use algorithms and personalized recommendations to bring them back?

This article explains how to build a user behavior prediction and recommendation optimization system using data from the LuckData Douyin API. We'll explore churn modeling strategies, recommendation system design, key performance metrics, and implementation examples, empowering you to transition from a simple content publisher to a data-driven content operator.

1. What Is User Churn and Why Does It Matter?

On Douyin, user churn doesn't necessarily mean account deletion — it often refers to users who have stopped following, stopped interacting with, or stopped actively consuming specific content.

Common Signs of Churn:

Declining engagement rates on brand videos (likes, comments, shares dropping)
Creator follower growth stagnates or turns negative
Significant drops in video completion rates and watch time
Users stop watching certain topic categories

Why Churn Prediction Is Crucial:

✅ More Precise Content Distribution: Identify users at risk of churning and push targeted content to re-engage them
✅ Smarter Creator Management: Intervene during peak churn periods with retention strategies
✅ Higher Marketing Efficiency: Re-engage low-activity users through personalized recommendations to improve ROI

2. Acquiring User Behavior Data: Entry via LuckData Douyin API

To build behavior profiles and predictive models, we need to collect the following key data types:

1. Viewing Behavior (Indirect)

You can infer changes in user interest by analyzing time series data such as views, like rates, and completion rates:

import requests
import datetime
headers = {'X-Luckdata-Api-Key': 'your_api_key'}
item_id = '7451571619450883355'
response = requests.get(
f'https://luckdata.io/api/douyin-API/get_pa29?type=trends&item_id={item_id}',
headers=headers
)
trend_data = response.json()['data']

This allows you to track video performance over the past 7–30 days and assess if interest is waning.

2. Creator Follower Trends

Monitoring a creator’s follower change trends enables churn prediction at the creator level:

response = requests.get(
f'https://luckdata.io/api/douyin-API/get_pa29?type=author&item_id={item_id}',
headers=headers
)
author_info = response.json()['data']['author']
fans_trend = author_info.get('fans_trend')  # Daily follower gain/loss data

Such data can support models that forecast follower loss or evaluate content quality regressions.

3. Building a User Churn Prediction Model

1. Data Preparation (Feature Engineering)

Using "Video × Date" as the unit of analysis, extract features such as:

View growth rate over the first 1–7 days post-publishing
Engagement rate (likes / views)
Ratio of negative sentiment in comments (via NLP analysis)
Posting frequency changes
Follower growth stagnation
Prolonged inactivity (no recent content)

2. Labeling Churn

Define churn based on future interaction behavior:

if current_7day_like_rate < 0.6 * avg_30day_like_rate:
label = 1  # Churned
else:
label = 0  # Retained

Custom thresholds can be set for different account types (e.g., educational vs. entertainment).

3. Model Selection

Random Forest / XGBoost: Well-suited for mid-sized datasets with strong interpretability
LSTM / Transformer: Ideal for time series modeling of user behavior
Logistic Regression: Fast, simple, and effective for early prototyping

Recommended libraries: scikit-learn, lightgbm, PyTorch.

4. Recommendation Optimization: From “Who to Recommend To” to “What to Recommend”

Once potential churn users are identified, the next question is: What content can win them back?

Recommended Strategies:

✅ Personalized Recommendations: Analyze the user’s historical top-engaged tags (e.g., beauty, vlog, knowledge) and suggest similar content
✅ Trending Content Push: Show popular rising videos to drifting users to rekindle interest
✅ Affinity Network Suggestions: If a user followed Creator A and A shares a large audience overlap with Creator B, recommend B’s content

Content Tag Extraction Example:

response = requests.get(
f'https://luckdata.io/api/douyin-API/get_pa29?type=items&item_id={item_id}',
headers=headers
)
tags = response.json()['data']['tags']  # Topic/hashtag information

You can build a user-interest vector from content tags and use vector similarity for matching recommendations.

5. Visualization: Displaying Prediction and Recommendation Results

BI dashboards or custom web interfaces can present insights clearly and effectively:

✅ User Churn Trend Graphs: Track daily increase in potential churn users
✅ Top Churning Creators: Rank creators by follower loss rate
✅ Recommendation Click-Through Rates: Measure re-engagement success
✅ Tag-Based Retention Heatmaps: Visualize the effectiveness of each topic in retaining users

Suggested Visual Elements:

Line Charts: Churn trends + model accuracy over time
Funnel Charts: Recommendation → Click → Completion funnel
Heatmaps: Tags vs. Retention performance

6. Future Directions for Expansion

✅ User Segmentation for Targeted Ops: Classify users as “active,” “silent,” and “on-the-edge” and design tailored content for each
✅ Automated Content Push Systems: Integrate with CRM and API pipelines to deliver personalized content to low-activity users
✅ Emotion-Driven Content Strategy: Automatically downrank or adjust content with high negative feedback in comments

7. Conclusion: Drive Retention and Growth with Douyin Data

With the Douyin API provided by LuckData, you can go beyond content performance analysis and tap into real-time insights about user interactions, follower dynamics, and content trends. This enables you to build a fully data-driven system for user insight and recommendation optimization.

By evolving from content creation to behavior modeling, you’ll master not just the art of content, but the science of user retention.

A truly effective content operator doesn’t just chase view counts — they master the rhythm behind long-term user engagement.