朋友推荐

People You May Know: u2u推荐

1. requirements

场景/产品

  • Symmetrical Friendship: The system should recommend connections where both users are likely to accept the friendship.

目标

  • Network Growth: help users discover potential connections

  • User Retention: Improve user retention by enhancing the social experience

约束

  • Scale: the total number of users on the platform, and daily active users

  • Latency: Recommendations should be generated in real-time (e.g., < 200ms) to ensure a smooth user experience

  • average connections for one user

2. ML task & pipeline

  • 利用共同好友、位置、教育背景、工作经历判断可能认识

  • 输入用户信息，输出和该用户最相似的k个用户作为推荐

3. data

  • User Profiles: Demographics, interests, and preferences

  • Social Graph: Existing connections and interactions (e.g., mutual friends, common groups)

  • Behavioral Data: Past interactions (e.g., profile views, connection requests)

4. feature

User Features

  • Demographics: Age, location, education

  • Interests: Hobbies, favorite topics

  • Behavior: Frequency of interactions, types of interactions

Connection Features

  • Common Friends: Number of mutual friends. 共同好友个数

  • Shared Interests: Overlap in interests or groups

  • Interaction History: Past interactions between users (e.g., profile views, messages)

Context Features

  • Time of Day: Recommendations may vary based on the time of day

  • Device: Recommendations may differ based on the device used (e.g., mobile vs. desktop)

5. model

ranking

Collaborative Filtering: Recommend connections based on similar users' connections.

point-wise learning to rank

  • task 2 users as input, output the probability of forming a friend

graph based prediction

  • graph level prediction

    • predict if a chemical compound us an enzyme

  • node level prediction

    • predict if a specific user is a spammer

  • edge level prediction

    • predict if two users likely to connect

6. evaluation

offline

  • precision, recall, and AUC-ROC

online

  • A/B Testing: Continuously test new models and features to measure their impact on engagement and network growth

  • Feedback Loop: Use user feedback to retrain and improve models

7. deployment & serving

  • 部分 batch serving

8. monitoring & maintenance

reference

Previous信贷风控Next去重复性/版权检测

Last updated