Mozilla怒噴當前推薦系統技術：算法“陳舊弱智”，效果非常糟糕！

{"type":"doc","content":[{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Mozilla噴當前視頻平臺引領者所使用的推薦系統技術：使用的算法“陳舊弱智”，效果非常“糟糕”，堪稱“恐怖秀”。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"根據Mozilla本週三發佈的調查研究結果表明，大部分飽受用戶們吐槽的YouTube視頻推薦內容都出自該網站陳舊的AI算法之手。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"該調查研究從去年9月開始啓動，總共涉及到37380名YouTube觀衆。根據 Mozilla 的報告，這是同類研究中規模最大的一次，而且顯示出來的結果只是“冰山的一角”，其中每項發現都值得進一步跟蹤並做出深刻剖析。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Mozilla 敦促 YouTube對內容審覈與推薦模型予以透明化公開，並建議給用戶提供退出個性化推薦的選項。但 YouTube 每季度從廣告中獲得的收入高達 60 億美元，實現提供退出“個性化推薦”選項不太可能。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"這套推薦系統已經用了十幾年，但還存在哪些問題？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"對比成立於2015年的快手，2016年上線的抖音， 創建於 2005 年的YouTube算是推薦系統技術的早期引路人。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"YouTube成立沒多久，網站上的視頻數量就迅猛增長，成爲全球最大的視頻網站。截止 2008 年，整個 YouTube 視頻量已突破四千五百萬，每分鐘上傳視頻量 7 小時。截止 2014 年，每分鐘上傳視頻量超過 100 小時。2019 年，月度活躍用戶達 19 億。如此龐大的視頻量，使得用戶難以搜索到其感興趣的視頻。YouTube 的成功最終得益於推薦系統，同時它也是實時大規模推薦系統技術的探路者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"雖然多年來一直被用戶吐槽視頻推薦效果，但YouTube在該研究方向上卻處於業界前沿。幾篇已經發表的論文顯示，2008年YouTube研究並使用了基於用戶-視頻圖的隨機遍歷算法；2010年，算法升級爲基於物品的協同過濾算法；2013年將推薦問題轉換成多分類問題，並解決從神經網絡最後的衆多輸出節點中找出最大概率的輸出節點。此舉也爲2016年將推薦核心算法升級爲"},{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/article\/2016\/09\/YouTube-Recommendation-neural-ne","title":null,"type":null},"content":[{"type":"text","text":"深度學習算法"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"打下了基礎。這幾篇論文《Video Suggestion and Discovery for YouTube》、《The YouTube Video Recommendation System》、《Label Partitioning For Sublinear Ranking》、《"},{"type":"link","attrs":{"href":"https:\/\/mp.weixin.qq.com\/s?__biz=MzU1NDA4NjU2MA==&mid=2247494669&idx=2&sn=b1ca666f647373b0de4be5da388e53bc&chksm=fbea55c2cc9ddcd4f909dd2ea65102d9e0637857dac813d35f71229d0c897a435077c987418f&scene=27#wechat_redirect","title":null,"type":null},"content":[{"type":"text","text":"Deep Neural Networks for YouTube Recommendations"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"》和《Recommending what video to watch next: A multitask ranking system》都是推薦系統的典範之作。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"從去年開始，來自190個國家的總計37380名YouTube觀衆自願參加了這項由Mozilla牽頭開展的衆包研究；在2020年7月至2021年5月期間，Mozilla共收到3362份關於不感興趣視頻的提交報告。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"根據本週三發佈的調查結果，“YouTube推薦算法自身只是問題的縮影，由此可以想見商業算法正在給民衆的生活蒙上一層不透明、不確定的陰影。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" “YouTube的算法每天向用戶提供約7億小時的視頻觀看時長，但公衆對其底層運作方式可謂知之甚少。我們甚至找不到任何官方支持的研究方法。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"作爲火狐瀏覽器的開發商，Mozilla公司開發出一款名爲RegretsReporter的瀏覽器擴展供YouTube用戶們下載。在安裝之後，該擴展程序會記錄網民在YouTube上的觀看活動、記錄所觀看視頻的詳細信息，並允許用戶輕鬆標記出自己覺得根本不感興趣的內容。通過將數據彙集起來並加以分析，Mozilla希望深入研究YouTube推薦引擎的行爲模式與實際效果。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"這項研究的結果有幾項亮點："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"志願參與調查的用戶們也有多種不同的抗拒理由，有些視頻與政治陰謀論有關、有些是與COVID-19疫苗相關的虛假信息、也有一些是對熱門大片《玩具總動員》的拙劣模仿。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Mozilla研究人員發現，在志願參與調查的用戶們提交的全部不感興趣視頻中，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"有71%來自YouTube平臺的whiz-bang AI推薦算法"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"他們估計，在被舉報的視頻中，甚至"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"有12.2%的內容有違YouTube自己提出的視頻管理方針及政策"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"——換句話說，這些視頻壓根不應該出現在YouTube網站上，但推薦算法居然還將其廣泛傳閱。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"研究還發現，推薦的視頻被志願者舉報的可能性比他們自己搜索到的視頻高 40%。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在 Mozilla 志願者對視頻進行負面反饋後，只有43.6% 的推薦與志願者之前觀看的視頻完全無關。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"YouTube推薦算法在非英語國家的表現似乎更差。其中巴西、德國與法國的推薦質量最差，美國和英國則分別排名第八位與第十六位。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"必須承認的是，當前軟件並不夠完美"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/mp.weixin.qq.com\/s?__biz=MzU1NDA4NjU2MA==&mid=2247488344&idx=1&sn=a9b32b10c1cfc7a1b4a61823aeaaa806&chksm=fbe9aa97cc9e238127e8a0a6ee8035eb33d46b114aec7cb67baa21206ea89f3a8d0ea5a8fec1&scene=27#wechat_redirect","title":null,"type":null},"content":[{"type":"text","text":"多年來"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，YouTube 的視頻推薦算法一直被指責通過向公衆投放經過放大的仇恨言論、政治極端主義、虛假垃圾信息，助長了社會弊病，以此謀取數十億人的眼球，從而增加廣告收入。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"雖然 YouTube 的母公司谷歌偶爾會對圍繞算法爆發出來的反對意見做出迴應：宣佈一些政策調整，以及限制或清除奇怪的仇恨賬戶，但不確定YouTube什麼時候會重啓這些誘導用戶點擊不健康視頻的規則。根據 Mozilla 的研究，YouTube 的人工智能仍然表現得如此糟糕，這也表明谷歌在用膚淺的改革主張模糊這方面的批評。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"谷歌公司一位發言人在聲明中表示，“我們這套推薦系統的目標，是幫助觀衆快速找到自己喜愛的內容。這套系統光是在主頁上的單日推薦量就超過2億條視頻。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" “我們使用超過800億條信息爲推薦系統提供指引，包括觀衆對感興趣內容的調查回覆。我們一直致力於改善YouTube平臺的觀看體驗；單在過去一年，我們就推出了30多項不同調整，希望減少有害內容的推薦比例。伴隨這項舉措，用戶以推薦方式接觸到極端視頻內容的機率已經遠低於1%。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"據報道，YouTube 最終刪除了近 200 個志願者在這次調查中反饋過的視頻。這些視頻在被刪除之前總共有 1.6 億次觀看。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"YouTube多年來一直在努力改善推薦系統，並不斷調整以提高效果表現。但必須承認，這款自動化軟件仍然不夠完美——特別是還在將有違內容管理政策的視頻推薦給用戶。Mozilla認爲，造成這種結果的核心原因，在於YouTube一直對所使用的自家的推薦算法底層邏輯三緘其口。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"報告指出，“我們認爲，此次研究揭露出的總是還只是冰山一角；其中每項發現都值得進一步跟蹤並做出深刻剖析。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" “我們還意識到，如果不加以干預並對YouTube算法開展更嚴格的審查，那麼相關問題將繼續失控蔓延，最終給整個互聯網社區產生愈發惡劣的影響。儘管YouTube方面宣稱已經在一部分問題上取得了進展，但研究人員幾乎無法驗證這些說法，也極難對YouTube推薦算法進行真正有意義的研究。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Mozilla公司認爲，YouTube應該發佈關於其推薦系統工作原理的數據，並對內容審覈與推薦模型予以透明化公開。只有這樣，研究人員才能真正以獨立方式對這款AI軟件開展審計。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"參考鏈接："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/www.theregister.com\/2021\/07\/08\/youtubes_mozilla_algorithm\/","title":null,"type":null},"content":[{"type":"text","text":"https:\/\/www.theregister.com\/2021\/07\/08\/youtubes_mozilla_algorithm\/"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]}]}