X,也就是 Twitter 推薦算法公開了,地址在這:
https://github.com/xai-org/x-algorithm
下面,是我整理的要點(diǎn):
1. 被人拉黑、靜音、舉報(bào),你的帖子會(huì)被壓
代碼里 P(block_author)、P(mute_author)、P(report)、P(not_interested) 四個(gè)負(fù)向預(yù)測,權(quán)重為負(fù),預(yù)測值越高得分越低
2. 連著發(fā)好幾條,后面的曝光會(huì)變少
Author Diversity Scorer 的作用是「Attenuate repeated author scores to ensure feed diversity」,同一作者連續(xù)出現(xiàn)會(huì)被衰減
3. 別人發(fā)了爆款,不會(huì)擠掉你的曝光
Candidate Isolation 設(shè)計(jì):「candidates cannot attend to each other—only to the user context」,每條帖子獨(dú)立打分
4. 用戶在你帖子上停多久,算法會(huì)預(yù)測
P(dwell) 是獨(dú)立預(yù)測項(xiàng),停留時(shí)間直接影響得分
5. 視頻只預(yù)測「會(huì)不會(huì)點(diǎn)開看」,不管看完沒
預(yù)測項(xiàng)是 P(video_view),不是完播率,和抖音邏輯不同
6. 以前那些手工調(diào)的規(guī)則,全刪了
README 寫的:「We have eliminated every single hand-engineered feature and most heuristics」,全靠 Grok 從互動(dòng)歷史學(xué)
7. 你看過的帖子不會(huì)再出現(xiàn)
過濾器里有 PreviouslySeenPostsFilter 和 PreviouslyServedPostsFilter
8. 不是你關(guān)注的人發(fā)的內(nèi)容,有單獨(dú)的調(diào)分邏輯
OON Scorer:「Adjust scores for out-of-network content」,具體怎么調(diào)沒公開
![]()
特別聲明:以上內(nèi)容(如有圖片或視頻亦包括在內(nèi))為自媒體平臺“網(wǎng)易號”用戶上傳并發(fā)布,本平臺僅提供信息存儲服務(wù)。
Notice: The content above (including the pictures and videos if any) is uploaded and posted by a user of NetEase Hao, which is a social media platform and only provides information storage services.