Putt
Regular Member
- Jan 1, 2020
- 448
- 373
If you have any questions, feel free to ask and I'll make sure to find out the answer.
1. User Interest Vector Space Modelling
As we expected, TikTom categorizes both users and creators with 'tags' that are related to interests. I'm 99% sure that these line up with the 20 or so content tags that appear in the Creator Marketplace. Rather than using a 100% dynamic system to pick FYP content, the algorithm only searches within the scope of your 'tags'. This is how FYPs can be so different and it can be difficult to discover new types of content if you already 'trained' your FYP.
Equally, this will likely make niche shifts difficult as if you start uploading a different type of content, your reach will be pushed towards a bad audience. Something to be wary of if you're buying accounts. If you buy a sports account and start publishing gaming content, rather
2. Positive Reinforcement Weighting
The order of 'significance' for engagements (high to low) is Share, Comment, Like, Viewtime.
3. Device Data Capturing
4. Cross Domain Learning
The algorithm is set up to predict behaviour based on similar people. If you like 'A' and 99% of people who like 'A', like 'B', it will assume you also like 'B'. It also does this cross-domain. What this means is the algorithm will identify that people who like Shawn Mendes are more likely to enjoy football over baseball or whatever.
5. Pre-Publication Content Scoring
This tech was researched for the purpose of their ad platform but it is likely used or soon to be used by the main platforn. It describes machine learning algorithms that can identify the quality of a piece of content without having to show it to users to collect data. This could easily be one of the reasons behind 0 views posts. Imagine Google assigning ads a quality score instantly after creation of ad copy.
Other Stuff
More analysis + research to come...