[推薦] Twitter 推薦演算法總覽
twitter 在愚人節前夕開源了他的推薦演算法 source code,可以看到他們是如何運作的
Source
https://www.youtube.com/watch?v=a8wCny94-_U 12:30
https://github.com/twitter/the-algorithm
Summary
整體來說跟其他 social network recommendation engine 差不多
想像一個訊息漏斗,由廣到深可以分為三層
- source relevant tweets: 分為 in-network(things you follow) 和 out-of-network(推薦)
- 50% in network
- 15% social graph: 有興趣的主題或是有互動過的用戶 overlap,不會有突破同溫層的情況
- 35% embedding space: 主動關注過的主題或標籤組成中推薦
- tweets are ranked: 針對每個 tweet 都依序詢問特定問題,increasing scores
- does this user interact with the author a lot?
- has the user seen a lot of this author recently?
- how much interest does the user have in this type of content?
- is the author a twitter blue subscriber?
- is the author Elon Musk?
- filter out specific preference: 你的 twitter 首頁呈現的結果,decreasing scores
- blocked, muted, nsfw filter
- 避免看到太多同個 author 的 tweet
- 多媒體(text, image…)平均呈現
發佈時間
2023-4-4