- Joined
- Jul 6, 2019
- Messages
- 149
- Reaction score
- 14
- Points
- 25
I was approached to share insights on the TikTok algorithm and future plans. I'll use this thread to periodically share intriguing details.
Feel free to ask questions, and I'll do my best to find answers.
- User Interest Vector Space Modeling: TikTok categorizes users and creators with 'tags' related to interests. These likely align with the content tags in the Creator Marketplace. Instead of using a fully dynamic system for the FYP, the algorithm searches within your 'tags' scope. This explains the diverse FYPs and the challenge in discovering new content if your FYP is already 'trained.'
- Positive Reinforcement Weighting: Engagement significance order is Share, Comment, Like, Viewtime.
- Device Data Capture:
- Cross-Domain Learning: The algorithm predicts behavior based on similar users. If you like 'A' and most others who like 'A' also like 'B', it assumes you'll like 'B' too. This cross-domain capability means it can associate preferences across different interests, like correlating Shawn Mendes fans with a preference for football over baseball.
- Pre-Publication Content Scoring: Initially developed for their ad platform, this technology is likely utilized or soon to be utilized on the main platform. It employs machine learning to evaluate content quality without user interaction, potentially explaining posts with zero views. It's akin to Google assigning instant quality scores to ads upon creation of ad copy.
Further analysis and research are on the horizon...