[{"data":1,"prerenderedAt":201},["ShallowReactive",2],{"content:\u002F2025\u002F12\u002Frecemmend-algorithm":3,"surround:\u002F2025\u002F12\u002Frecemmend-algorithm":190},{"id":4,"title":5,"body":6,"categories":167,"date":169,"description":170,"draft":171,"extension":172,"image":173,"meta":174,"navigation":176,"path":177,"permalink":173,"published":173,"readingTime":178,"recommend":173,"references":173,"seo":183,"sitemap":184,"stem":185,"tags":186,"type":188,"updated":169,"__hash__":189},"content\u002Fposts\u002F2025\u002F12\u002Frecemmend-algorithm.md","短视频的推荐算法",{"type":7,"value":8,"toc":151},"minimark",[9,13,17,20,23,26,29,32,38,41,46,49,52,55,58,61,64,67,70,73,76,86,89,92,95,98,101,124,127,135,138],[10,11,12],"p",{},"推荐算法的目的是让用户在海量数据中定位到自己想要的数据,针对于短视频来说,就是在快速找到用户感兴趣的视频,市面上常见的推荐算法分为以下几种",[14,15,16],"h2",{"id":16},"基于流行度的的算法",[10,18,19],{},"根据pv,uv,点赞量,收藏量,分享量等等指标把最火的内容无差别推送给用户",[10,21,22],{},"特点: 比较简单粗暴,针对于大众,如果想要对不同用户做个性化推荐的话,可以根据用户群体分类下流量排行进行分类",[10,24,25],{},"比如含有美食标签和含有运动标签推荐的视频是不一样的",[14,27,28],{"id":28},"根据用户行为数据的算法",[10,30,31],{},"基于用户的历史数据,会出现冷启动问题(首次没有历史数据)",[33,34,35],"blockquote",{},[10,36,37],{},"基于协同过滤的推荐机制是现今应用最为广泛的推荐机制",[10,39,40],{},"主要细分为3个算法",[42,43,45],"h3",{"id":44},"基于用户的协同过滤算法","基于用户的协同过滤算法:",[10,47,48],{},"原理是通过比对多个用户与当前用户的行为,如果相似度越高,则会向当前的用户推荐相似度高的用户且没有被当前用户浏览的数据",[10,50,51],{},"比如用户A,B,C,D都喜欢王者荣耀视频,都喜欢刷张大仙的视频,点赞然后也喜欢在其视频下评论,计算这些人群评价得分(通过相似用户点赞浏览关注评论得出),把他们“认为”评分最高,而且A没看过的视频推给A",[42,53,54],{"id":54},"基于项目的协同过滤算法",[10,56,57],{},"原理是通过比对多个用户与当前用户的行为,如果相似度越高,则会向当前的用户推荐相似度高的用户的历史数据",[10,59,60],{},"比如A,B,C,D都是电脑设备的忠实脑残粉,A喜欢Mac,B喜欢华硕,C喜欢Mac,系统会像A推荐C历史购买过的商品",[42,62,63],{"id":63},"基于模型的协同过滤算法",[10,65,66],{},"机器学习与数据挖掘暂时用不到",[14,68,69],{"id":69},"基于内容的算法",[10,71,72],{},"根据用户过去喜好的内容和项目本身的一些特征",[10,74,75],{},"比如A过去痴迷于看NBA篮球比赛,系统也会时不时向A推送一些NBA类的视频",[77,78,79,83],"ul",{},[80,81,82],"li",{},"平台自动打标签(闲鱼)",[80,84,85],{},"大众自己打标签(掘金)",[10,87,88],{},"优点是不那么依赖大众推荐热门视频,跟用户历史数据有关",[10,90,91],{},"缺点是标签的相关性难以保证,需要用户配合",[14,93,94],{"id":94},"基于关联规则的推荐",[10,96,97],{},"适用于电商系统",[14,99,100],{"id":100},"评估推荐算法的效果",[10,102,103,104,108,109,108,112,115,116,119,120,123],{},"当推荐算法完成后，怎样来评估这个算法的效果? ",[105,106,107],"code",{"code":107},"CTR(点击率)","、",[105,110,111],{"code":111},"CVR(转化率)",[105,113,114],{"code":114},"停留时间","等都是很直观的数据。在完成算法后，可以通过线下计算算法的",[105,117,118],{"code":118},"RMSE(均方根误差)","或者线上进行",[105,121,122],{"code":122},"ABTest","来对比效果。",[14,125,126],{"id":126},"业界短视频方面的推荐系统是怎么做的",[10,128,129],{},[130,131,132],"a",{"href":132,"rel":133},"https:\u002F\u002Fwww.sohu.com\u002Fa\u002F899390218_122326373",[134],"nofollow",[10,136,137],{},"参考文章",[77,139,140,146],{},[80,141,142],{},[130,143,144],{"href":144,"rel":145},"https:\u002F\u002Fpdai.tech\u002Fmd\u002Falgorithm\u002Falg-domain-suggest.html",[134],[80,147,148],{},[130,149,132],{"href":132,"rel":150},[134],{"title":152,"searchDepth":153,"depth":153,"links":154},"",4,[155,157,163,164,165,166],{"id":16,"depth":156,"text":16},2,{"id":28,"depth":156,"text":28,"children":158},[159,161,162],{"id":44,"depth":160,"text":45},3,{"id":54,"depth":160,"text":54},{"id":63,"depth":160,"text":63},{"id":69,"depth":156,"text":69},{"id":94,"depth":156,"text":94},{"id":100,"depth":156,"text":100},{"id":126,"depth":156,"text":126},[168],"技术观察","2025-12-02","介绍短视频推荐系统中的用户画像、内容特征和召回排序思路，并结合业务需求分析实现方向。",false,"md",null,{"slots":175},{},true,"\u002F2025\u002F12\u002Frecemmend-algorithm",{"text":179,"minutes":180,"time":181,"words":182},"4 min read",3.785,227100,757,{"title":5,"description":170},{"loc":177},"posts\u002F2025\u002F12\u002Frecemmend-algorithm",[187],"algorithm","tech","QBe488GUC3F6GTU8TY9lFe_DMop82nwWcTo1Ucsr-Gs",[191,196],{"title":192,"path":193,"stem":194,"date":195,"type":188,"children":-1},"Webhook 工作原理与 Go 实现","\u002F2025\u002F11\u002Fwhat-is-webhook","posts\u002F2025\u002F11\u002Fwhat-is-webhook","2025-11-18",{"title":197,"path":198,"stem":199,"date":200,"type":188,"children":-1},"gRPC 工作原理与通信流程","\u002F2025\u002F12\u002Fgrpc-internals","posts\u002F2025\u002F12\u002Fgrpc-internals","2025-12-09",1786555054780]