推薦系統的未來發展(三十三)

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"寫在前面：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"大家好，我是強哥，一個熱愛分享的技術狂。目前已有 12 年大數據與AI相關項目經驗， 10 年推薦系統研究及實踐經驗。平時喜歡讀書、暴走和寫作。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"業餘時間專注於輸出大數據、AI等相關文章，目前已經輸出了40萬字的推薦系統系列精品文章，今年 6 月底會出版「構建企業級推薦系統：算法、工程實現與案例分析」一書。如果這些文章能夠幫助你快速入門，實現職場升職加薪，我將不勝歡喜。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"想要獲得更多免費學習資料或內推信息，一定要看到文章最後喔。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"內推信息","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果你正在看相關的招聘信息，請加我微信：liuq4360，我這裏有很多內推資源等着你，歡迎投遞簡歷。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"免費學習資料","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果你想獲得更多免費的學習資料，請關注同名公衆號【數據與智能】，輸入“資料”即可！","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"學習交流羣","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果你想找到組織，和大家一起學習成長，交流經驗，也可以加入我們的學習成長羣。羣裏有老司機帶你飛，另有小哥哥、小姐姐等你來勾搭！加小姐姐微信：epsila，她會帶你入羣。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着科學技術的進步，移動互聯網快速發展，手機越來越便宜，擁有智能手機不再是一件遙不可及的事情，互聯網用戶規模已接近增長的頂點。攝像頭和信息處理軟件技術(各種濾鏡、剪輯等工具)的進步讓每一個人都可以輕鬆地生產高質量的內容，信息的產生以指數級增長，我們的生活中充斥着海量的信息。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在上述背景下，怎麼高效快速地獲取對自己有價值的信息對每個互聯網公民來說是愈發重要的事情，推薦系統的出現可以輕鬆地應對這一棘手的難題。推薦系統作爲一種高效的信息過濾工具，可以很好地部分解決用戶精準高效獲取信息的問題(搜索、導航等也是解決用戶獲取信息的手段)，並且也是非常重要甚至是不可或缺的一種手段(在人們需求不明確時，藉助推薦系統獲取信息是非常必要的，而每一個人都有不明確的需求)。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統作爲一項技術在國內的發展時間不長，從2012年頭條成立之初將推薦系統作爲核心產品功能到現在差不多有8年時間，在這8年中推薦系統的商業價值在國內逐步得到認可和肯定，大家都認可推薦系統在內容分發、用戶體驗、商業變現等方面的重大價值。推薦系統目前已經成爲toC互聯網產品的標配技術，任何一個toC產品要想很好地爲用戶提供一種被動高效獲取信息的工具，推薦系統是繞不過去的。在特定情況下人類需求的不確定性、信息的爆炸式增長這兩個條件讓推薦系統成爲一項長久而實用的技術，推薦系統不會曇花一現，它會伴隨着人類的發展而不斷髮展進化。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在前面的一系章節中，我們對推薦系統的算法、工程、評估、展示、交互、業務等方方面面都進行了深入的介紹。雖然推薦系統在國內的出現只有短短8年，但是在各個方面都取得了極大的進步，發展越來越快，各種新的方法、應用場景、產品形態層出不窮。未來推薦技術會朝哪些方向發展？推薦行業又有哪些變化？推薦系統的應用場景和價值體現又有什麼新的特點呢？這些問題都值得我們深入思考。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"針對上述問題，作者結合自己對推薦系統的理解和行業判斷，在本章中講講推薦系統的未來發展與變化。具體來說，作者會從政策及技術發展對推薦系統的影響、推薦系統的就業變化、推薦系統的應用場景及交互方式、推薦算法與工程架構、人與推薦系統的有效協同、推薦系統多維度價值體現等6個方面來講解推薦系統的未來發展和變化。本章爲讀者提供多角度來思考推薦系統的未來發展與變化，期望讀者可以更好地把握推薦系統未來發展的脈絡，對推薦系統的未來變化有更深入的瞭解。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.1 政策及技術發展對推薦系統的影響","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統的發展是與整個大環境和技術發展趨勢密不可分的，一定會受到國家政策層面和技術發展的影響。不過對推薦系統來說，作者認爲政策和技術的影響是都正向的。下面我們就從政策和技術兩個維度來進行分析。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.1.1 政策層面的影響","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着數據化、智能化等概念的興起，大數據與人工智能在科技發展中起着越來越重要的作用，大數據與人工智能得到了國家層面的重視。要想發展好大數據與人工智能，首先必須有相關人才。國內從2016年開始逐漸有一些高校開始開設大數據和人工智能專業甚至創辦大數據、人工智能學院，這類高校呈上漲趨勢，目前全國開設了大數據相關專業的高校超過282個(見參考文獻1)。在2019年全國已經有35所高校獲得人工智能學科建設資格(見參考文獻2)。除了國家政策層面的支持，這類專業也受到了市場的青睞，就業前景較好。這兩方面的原因，促使高校樂意不遺餘力地推進大數據與人工智能專業的建設。教育層面對大數據與人工智能的支持，爲依賴這些技術的業務和產品提供了源源不斷的人才儲備。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統本身就是人工智能中非常重要並且有極大業務價值的子領域，同時構建推薦算法模型也依賴於對大規模用戶行爲數據的處理，大數據技術也是推薦系統必備的技術。因此，推薦系統直接受益於國家在教育層面對大數據與人工智能的支持，未來有充足的人才來源。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上面提到的只是國家在教育層面的佈局，其實國家將大數據與人工智能提到了戰略的高度，希望通過大數據與人工智能來革新各個產業。政策層面的大力支持，媒體的大勢宣導，今日頭條的樣板示範作用，讓個性化推薦相關產品和業務得到更多投資人、公司管理層的重視，這也有利於推薦系統在更多產品和業務中落地。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.1.2 科技層面的影響","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"雲計算技術是近十來年非常火的技術，雲計算行業已經發展得越來越成熟，大公司早已佈局，並已成爲盈利源泉，是業務的第三增長極，國外的有亞馬遜的AWS、微軟的Azure等，國內有阿里雲、騰訊雲等。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"經過近十年的發展，雲計算基礎設施已經相對健全，未來會在SAAS服務和toB行業應用中大力發展，這其中就包括推薦SAAS服務。創業公司只需要利用雲平臺提供的各種SAAS服務就可以輕鬆搭建推薦系統各個模塊，大大降低了推薦系統的准入門檻。除了雲計算公司提供這類服務，toB的創業公司也在這方面有所佈局，也提供PAAS或者SAAS的推薦服務。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"構建一套完善、穩定、高效、低成本、靈活的推薦系統是一件非常困難的事情，涉及到數據、算法、工程、產品交互、業務指標等方方面面，只有對這些知識點有深入全面的瞭解，再結合公司的業務才能構建出具備商業價值的推薦系統。在這一背景下，創業公司一般可以選擇利用雲服務來構建推薦業務，這種方式投入低，無固定成本，是非常好的選擇。只有中、大規模公司或者將推薦作爲核心競爭力的公司纔會自建一套推薦算法業務體系。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2020年突如其來的新冠病毒疫情，給每個人的生活帶來了極大的影響，限制了每個人的線下活動，用戶將更多時間用在了線上。有很多研究認爲這些變化是持久的，不會隨着疫情的消失而消失。如果真是這樣，這也間接提升了推薦系統等互聯網服務面對的用戶規模，爲推薦系統的發展創造了新的機遇與挑戰。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上面這些變化，對推薦行業產生了深遠的影響，對推薦系統各方面都會帶來極大的改變。最直接受到影響的是推薦系統就業的變化，這就是我們下一節主要講述的內容。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.2 推薦系統的就業","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們在最後一章《推薦算法工程師的成長之道》中會講到推薦算法是一個非常好的職業選擇，主要表現在就業範圍廣(推薦、搜索、廣告技術一脈相承，技術體系極爲類似)、薪資高、有業務價值(讀者可以參考這一章第一節“爲什麼說推薦算法是好的職業選擇”)。本節我們就來講解在大環境和科技層面不斷髮展變化的情況下推薦系統就業的變化。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們在31.1節講到各類高校開設大數據與人工智能課程、成立相關院系，未來推薦相關的人才供給會更加充裕。大數據和人工智能是當下的熱點，而推薦系統是人工智能中非常重要並且有極大應用前景和商業價值的方向，人的從衆本性會導致對熱點盲目追隨崇拜，趨之若鶩，因此一定會有很多從其他方向轉崗到推薦算法領域的人才。雖然將來會有更多的企業提供推薦產品和服務，但作者個人判斷推薦方向的人才肯定會供過於求，相關職位競爭壓力極大。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"構建一套完善的推薦系統是一個非常複雜的系統工程，因此纔有這麼多雲計算公司和toB創業公司將推薦系統打造爲高效易用的SAAS或者PAAS服務，在不久的將來，很多公司不會自己從零開始搭建推薦算法團隊了，而是直接購買雲平臺或者toB公司的推薦服務。因此，推薦方向的工作形式和工作重點可能會有如下幾類變化。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.2.1 推薦算法商業策略師是新的職業方向","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着推薦系統相關的雲產品越來越成熟，創業公司會更傾向於直接購買推薦雲服務，快速搭建自己的推薦算法產品，而不是從零開始自己摸索。利用雲產品的好處是輕量、快速，讓公司將更多的精力放到核心業務上，輕裝上陣，快速發展業務。關於這一塊的詳細討論，讀者可以參考《從零開始構建企業級推薦系統》中27.2.3節的介紹。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"爲了更好地將雲推薦產品落地到企業中，對需要的人才技能及要求會有變化，這時不需要特別懂具體的算法實施和工程，更多的是希望瞭解各類算法的優缺點和應用場景，能將推薦算法跟本公司的業務結合起來，讓推薦算法更貼合本公司的業務情況，最終讓推薦算法產生業務價值。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這類人員需要了解推薦系統全流程，知道構建推薦系統可能遇到的困難，有全局把控能力，善於溝通，有對商業的敏銳嗅覺。這樣的人才作者稱爲","attrs":{}},{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"推薦算法商業策略師","attrs":{}},{"type":"text","text":"，他們的主要工作是怎麼基於推薦雲服務將推薦落地到本公司的業務中。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.2.2 在特定領域和場景下出現新的推薦形態","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着科技的發展，特別是智能硬件、5G通信技術、語音等新交互方式的發展，推薦系統的應用場景及交互方式會拓展到更多場景和領域。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在新的業務場景及新交互方式下，怎麼構建推薦業務及推薦算法是非常值得思考的一個問題，也是未來新的機會。讀者可以參見31.3節關於推薦場景及交互方式變化的介紹。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"由於是新的行業和場景，短期雲計算公司提供的推薦服務還很難覆蓋到這些行業和場景，因此，在這些新興的行業和場景中，是需要企業自己來實現相關的推薦服務的(當然雲計算公司的產品可以提供一定的補充作用)，這對於推薦算法從業人員來說也是新的機遇。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着新領域的逐步成熟，雲計算與toB服務公司也會湧入新賽道。提供推薦SAAS或者PAAS服務的雲計算公司或者toB創業公司也需要大量精通推薦算法和工程的專業人才，在這些新領域提供推薦解決方案。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.2.3 推薦從業者需要更加關注業務價值產出","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統本身就是一個比較偏業務和工程的方向，企業構建推薦系統的目的就是希望藉助推薦系統來獲得更多的商業價值(我們在前面的《推薦系統的商業價值》中已經對推薦系統的商業價值進行過深入介紹)。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在當前互聯網紅利見頂的情況下，原來那種通過融資燒錢發展用戶的粗放經營模式不再有市場。在當前競爭日益激烈的商業環境下，企業從創立第一天就應該考慮商業變現的事情，需要在創業早期階段就嘗試商業化，學習這方面的技能，積累相關經驗，這樣才更有可能生存下來。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統作爲一個非常有價值的變現工具，需要肩負起商業變現的責任，因此推薦從業人員需要更加關注推薦系統的業務價值產出，並儘量量化推薦系統的價值，建立價值產出的閉環體系。只有讓老闆看到推薦的價值，老闆纔會大力支持，最終推薦業務纔有更好的發展空間。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.2.4 推薦系統相關技術培訓市場更加火爆","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"前面提到很多新興toC互聯網行業都將推薦系統作爲核心功能，加上雲計算和toB創業公司對推薦工程師的大量需求，市面上對推薦算法人才的需求是比較旺盛的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦算法工程師的工資水漲船高，吸引很多其他方向的工程師轉行從事推薦算法相關工作，他們沒有推薦相關技能儲備及項目經驗，因此需要進行學習和培訓。另外雖然很多高校開設了人工智能方面的專業，但是大學所學課程跟企業對技能的要求還是有比較大的差距的，嚴重脫節。這些學生要想在激烈的競爭環境中找到推薦系統相關工作，可能需要提前找一個推薦系統相關的實習工作，積累經驗，或者參加相關職業技能培訓。這兩種情況都促使市面上出現很多從事推薦系統相關技能培訓的副業者及相關技術培訓創業公司。這些公司進行在線或者線下技能培訓，這也間接提供了錄製推薦課程或者培訓推薦技能的工作機會。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.3 推薦系統的應用場景及交互方式的多元化","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"目前的推薦系統主要應用於PC端和移動端，特別是在移動端，佔推薦系統產品的絕大多數。未來隨着智能化的發展，智能設備會出現在更多的場景中，這些場景中的應用當然也可能需要藉助推薦技術來分發信息。同時這些場景不同於移動端，在交互方式上會有變化，可以藉助語音、手勢等更多新的交互方式與用戶互動。下面我們就對3種可行的推薦系統應用場景進行說明。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.3.1 家庭場景","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"國內最早在2015年5月份樂視智能電視發佈，隨後某米、某鯨、某風、某爲、傳統5大電視廠商(某虹、某維、某L、某信、某佳)紛紛入局智能電視行業，國外電視廠商也強勢殺入中國智能電視市場。各類智能盒子(某米盒子、某貓魔盒等)種類繁多，五花八門，家庭互聯網進入智能時代。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"目前智能電視上唯一的殺手級應用就是看視頻。在智能電視上的操作主要是以遙控器爲主(雖然很多智能電視具備語音交互能力，但是目前還存在居多問題，導致交互能力有限)，操作相對手機來說更爲不便，因此個性化推薦的作用就凸顯出來，智能電視上是更適合做智能推薦的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"作者所在公司開發的家庭智能軟件產品某貓，作爲聚焦家庭智能終端(電視機、智能盒子)上的視頻應用，從2012年就開始構建個性化推薦系統，目前已有超過15種智能推薦產品形態。推薦系統在提升用戶體驗、創造商業價值等方面產生了巨大的價值。某奇藝、某訊視頻、某酷等互聯網視頻巨頭都已經佈局智能電視端，並且它們都提供了一定的智能推薦能力。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在智能電視或者智能盒子上構建推薦系統，由於交互方式和展示方式上的特殊性，以及面對的是非移動多人場景，跟移動端有很大的差別，且更有難度，這裏面有很多點是值得探索和挖掘的，比如怎麼更好地跟用戶交互、怎麼識別多人場景並提供精準推薦能力。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"家庭場景中另外一個不得不說的智能硬件是智能音箱。前幾年某馬遜的Echo在美國大熱，引爆了智能音箱市場，國內快速跟進，某AT、某米、某大訊飛等一衆企業紛紛佈局，上演了智能音箱大戰。國內目前每年有千萬級的銷售量，逐步成爲家庭中僅次於智能電視的現象級硬件產品。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"智能音箱以語音交互爲主(帶屏智能音箱也可以採用觸控的方式交互)，智能音箱上的應用目前種類非常多，以音樂、故事、知識、生活幫助(查天氣股票等)爲主，除了音樂這類音頻的產品可以非常自然地整合個性化推薦能力(是手機上的個性化電臺在智能音箱上可以自然延伸)，目前還沒有看到將智能推薦應用到智能音箱中的其他場景(比如購物等)。由於交互方式、展現方式的限制，在除音樂等音頻之外的應用上，怎麼整合智能推薦的精準推薦和信息分發能力，是一個比較難的事情，也是一個非常值得探索的方向。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.3.2 車載場景","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"車載場景是一個非常重要的場景，用戶規模巨大，同時也是一個比較特殊的場景。在車載場景下司機的主要注意力在開車，車載智能設備的交互方式一定以語音交互爲主，應用也會有所侷限，音樂、新聞等是主要的應用場景。推薦系統也會聚焦在音樂、新聞等信息流推薦上，其他的智能推薦應用場景很難挖掘。這跟智能音箱類似，這裏不再贅述。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.3.3 VR(虛擬現實)/AR(增強現實)/MR(混合現實)場景","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"虛擬現實/增強現實/混合現實等技術的發展，給人類提供了了解世界、獲取信息的一扇窗，目前這些智能設備還不夠成熟，基於設備之上的應用也相對少(主要是遊戲類、視頻類)。這類設備的交互方式以語音、手勢、觸控、頭部動作等爲主。這類設備上生態還極不成熟，內容也相對少，目前還不滿足做智能推薦的條件，但是是一個比較值得期待的方向。特別是當混合現實發展成熟時，就像谷歌眼鏡那樣，人可以在行走中獲取信息，並可以整合周圍的環境信息，推薦系統一定有很多新奇的玩法。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"除了上面提到的三類應用場景，其他場景如智能冰箱上做智能推薦也是可行的。智能冰箱可以記錄家居生活中食物的消費情況，瞭解家庭的飲食習慣。基於對家庭消費習慣的挖掘，進行精準的個性化食物推薦是非常可行的一種策略，這就跟電商直接聯繫起來了，非常有商業價值，值得期待和探索。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"應用場景的變化一定伴隨着交互方式的變化，在上述幾類場景中主流的交互方式都跟手機上的觸屏交互不一樣，因此應用場景對基於這些場景下的智能推薦的交互及展示方式是有極大影響的。對於推薦系統UI交互和視覺展示的未來發展，讀者可以參考《推薦系統的UI交互與視覺展示》中的24.5節“推薦系統UI交互和視覺展示的展望”，我們在這一節中也進行過深入的探討。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.4 推薦算法與工程架構的發展","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統中最重要、最核心、最有技術含量的一個模塊非推薦算法莫屬了。目前主流的、在工業界大量使用的推薦算法有基於內容的推薦算法、各類協同過濾算法等，這些傳統的推薦算法時至今日還在推薦系統中發揮着巨大的價值。隨着機器學習技術、大數據技術、雲計算及軟硬件的發展，會有更多新的學習範式應用於推薦系統中。除了算法層面的變化外，通信技術(特別是5G)的發展讓各類具備流暢體驗的實時推薦成爲可能，推薦系統在數據處理、工程架構等方面也會迎來新的發展與機會。下面我們就從算法和工程兩個角度來梳理推薦推薦系統的未來發展。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.4.1 推薦算法新的機會","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"最近幾年隨着深度學習技術的發展，深度學習在推薦系統中的應用也越來越廣泛。深度學習由於可以獲得比傳統算法更好的精準度、不需要做複雜的人工特徵工程而受到推薦算法工程師的追捧，深度學習技術逐漸成了推薦系統中的主流技術。作者在《深度學習在推薦系統中的應用》中已經對深度學習在推薦中的應用進行了比較全面的介紹，其中12.7節“深度學習推薦系統的未來發展”中對深度學習在算法模型維度、工程維度、應用場景維度、數據維度、產品呈現與交互維度等進行了探討，讀者應該還有印象。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統本質上是一個交互式學習引擎，它會根據用戶對推薦物品的反饋(是否瀏覽、點擊、購買等)來調整後續給該用戶的推薦結果，這個過程是一個互動的過程，用戶與推薦系統互動得越多越頻繁，推薦系統就越懂用戶，給用戶的推薦也會越精準。在機器學習領域有一種學習範式就是互動式學習的典範，這就是強化學習(參見下面圖1)。強化學習中智能體通過與環境互動(action)獲得環境的反饋(feedback)，基於反饋調整自己與環境的交互，形成新的交互方式與策略，最終通過多倫互動，智能體可以更好地從環境中學習，獲得更大的綜合回報。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/a4/a488bc0eb6e711b39f39dcaa5e171d43.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"圖1：強化學習範式","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果我們在強化學習範式下考慮推薦系統，推薦算法就是智能體，而使用推薦系統的人就是環境，推薦系統通過與人互動(推薦系統給人推薦標的物，而人對推薦的標的物進行行爲操作)更深入地瞭解人的行爲特點、興趣偏好。推薦系統從與人互動中不斷迭代，獲得更好的推薦效果。強化學習在推薦系統中的應用，工業界已經有一些成果，感興趣的讀者可以查看參考文獻3、4、5、6，分別是某條、某東、YouTube將強化學習應用於推薦中的案例。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對強化學習感興趣的讀者可以學習參考文獻7，這是強化學習之父Sutton寫的一本非常有影響力的書。隨着推薦系統越來越趨向於實時化，作者相信強化學習在推薦中的應用一定是未來非常值得探索的方向，也一定會產生極大的商業價值。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"機器學習中另外一個非常重要的學習範式是遷移學習，所謂遷移學習簡單來說是將從一個領域獲得的知識通過某種方式應用於另外一個領域(需要尋找到這兩個領域之間的某種關聯關係)。這種學習範式對人類來說是再平常不過的事情了，我們平常所說的舉一反三、觸類旁通等就是人類大腦的遷移學習能力。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"遷移學習在推薦系統中的應用目前有少量嘗試，讀者可以查看參考文獻8、9、10、11、12，這些都是遷移學習在推薦系統上的探索。我們在《嵌入方法在推薦系統中的應用》11.5.3節中講到盒馬利用遷移學習將淘寶用戶特徵遷移到盒馬中可以很好地解決用戶冷啓動問題，這算是遷移學習在工業級推薦系統中一個比較好的應用案例。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"目前很多平臺型的大公司都孵化出了很多產品，構建出了超大規模的產品矩陣，比如阿里系下的產品矩陣、頭條系下的產品矩陣等，在這些產品之間進行遷移學習是非常自然的事情。另外雲計算公司服務於非常多的同類型公司，這裏面就有非常多遷移學習可以落地的場景，雲計算公司從一個公司構建推薦算法服務的經驗和獲得的算法成果都可以遷移到另外一家同類型的公司中(當然需要考慮到信息安全和隱私，這在下面提到的聯邦學習框架下是可行的)。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"監督學習目前還是機器學習中最重要、應用最廣的學習範式，但是獲得大量標準樣本是非常費時費力費錢的，怎麼在沒有大量標註樣本的情況下學習是一個非常重要的問題(在醫學等領域標記樣本不足是很自然的事情)。遷移學習提供了一種可行的方案，另外一個可行的方式是","attrs":{}},{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"半監督學習","attrs":{}},{"type":"text","text":"(參見參考文獻13)，半監督學習利用標記樣本和無標記樣本來進行學習，可以很好解決標記樣本不足的問題。目前我們所獲得的數據中無標記數據量是非常巨大的，比如視頻、音頻、評論信息、標的物介紹文本等，這些信息在半監督學習範式下都可以使用。參考文獻14就是半監督學習在推薦系統上的嘗試。這方面的技術目前還很少看到在企業級推薦系統上的應用，但是一定是未來非常值得深入挖掘的一個方向。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"目前國內的產品一般都是通過霸王協議來獲取用戶數據，多多少少都是不太符合人性和道義的，隨着用戶隱私意識的增強和法律層面對隱私保護的重視，未來推薦系統可能更難獲得更多的用戶數據，這就要求推薦算法從更加保護用戶隱私的方向努力，在這個方向上","attrs":{}},{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"聯邦機器學習","attrs":{}},{"type":"text","text":"(見參考文獻15)就是一種非常好的學習範式。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"聯邦機器學習是一個機器學習框架，能有效幫助多個機構在滿足用戶隱私保護、數據安全和政府法規的要求下，進行數據使用和機器學習建模。聯邦機器學習在推薦系統中的應用已經在業界有比較好的嘗試了，未來肯定會是推薦系統發力的一個方向。關於用戶隱私和聯邦機器學習相關的知識點讀者可以參考《推薦系統的價值觀》30.3.3節中的介紹。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上面講到了在推薦算法上未來推薦系統可能的方向和變化，在數據處理及工程方面，推薦系統也會面對很多的調整、變化與發展，下面我們來簡單梳理一下。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.4.2 推薦系統工程層面的發展變化","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"現在主流的基於內容的推薦算法和協同過濾推薦算法都只利用了部分用戶及標的物數據進行模型訓練，還遠遠沒有將所有可用的信息綜合起來進行推薦。這一方面是數據量太大，二是數據處理複雜(特別是富媒體數據處理起來成本高、技術複雜)，三是更多的數據對推薦算法性能及可拓展性的要求更高。在不久的將來隨着特徵工程技術的進步、數據處理能力的增強、計算成本的降低以及算法自身的發展，獲取更多的數據進行更加複雜模型的訓練成爲可能，更多的數據和更復雜的模型也會讓最終的推薦效果更好。這種複雜的模型可以是更深層的深度學習模型還可以是各種模型的混合推薦。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着通信技術的發展，特別是5G技術的普及，信息傳輸的速度更快、傳輸費用更便宜，我們可以在極短的時間內獲得大量的數據，計算能力的增強和算法模型的發展讓處理數據更加快速及時，同時用戶也趨向於獲得及時快速的互動，在這些因素的影響下，推薦系統正變得越來越實時。目前大火的信息流推薦就是很好的體現。實時推薦不僅用戶體驗好，並且還具備更好的商業價值(實時推薦增加了信息分發的效率，讓單個推薦位有更大的週轉率，提升了單個位置的商業價值)，實時推薦是推薦系統未來最爲重要的發展方向之一。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"要想做好實時推薦，除了算法外，對工程架構、交互方式等都需要進行相應調整。在工程上需要採用流式處理技術(如Flink、Spark Streaming等)來進行特徵處理與模型訓練，這樣纔可以更好地響應用戶的實時操作。交互方式上也需要給用戶提供更加自然流暢的交互，目前在移動端的下拉刷新就是一種比較好的交互方式，當推薦場景拓展到家庭智能設備、車載設備、甚至虛擬設備上時，交互方式都需要進行重大革新。作者在《實時個性化推薦》22.8節“實時推薦系統的未來發展”中已經對實時推薦系統未來的發展方向進行了比較全面的介紹，這裏不再贅述。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"還有一個不得不提的點是特徵工程，這是任何機器學習算法必須要面對的問題，隨着富媒體信息在所有信息中所佔比重越來越大以及實時推薦對特徵處理時效性的要求，這個問題變得日益嚴峻。幸好深度學習等技術可以減少人工特徵工程的難度，另外自動化特徵工程在某種程度上也可以緩解這個問題。關於特徵工程未來發展及變化，我們已經在《推薦系統之數據與特徵工程》17.5節“推薦系統數據與特徵工程未來趨勢”中進行過詳細介紹，這裏不再細說。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"目前的推薦算法都是部署在雲端的，所有人共用一套推薦算法體系。隨着邊緣計算技術的發展，未來是極有可能在終端上部署比較複雜的模型的，到那時就可以爲每個用戶構建一個個性化的推薦算法模型，直接在終端給用戶生成推薦結果。這種部署方式有幾大優點：一是推薦會更加及時，可以給用戶更好的體驗；二是每個人擁有自己量身定製的算法，算法精準度也會更高；三是信息直接在終端進行處理，也更加安全可靠。《實時個性化推薦》22.8節“實時推薦系統的未來發展”中對這個情況進行了比較細緻的描述，讀者可以參考。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"總之，隨着機器學習算法自身的發展，隨着硬件、信息處理技術、信息傳輸技術的發展，未來的推薦系統在算法實現方案、工程架構等方面都會產生極大的變化，會出現更多的可能性。這些都是值得我們去期待、去思考、去探索的方向。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.5 在與推薦系統協作上凸顯人的價值","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"雖然以深度學習爲代表的第三次人工智能浪潮給學術界和產業界帶來了極大的利好，機器學習在很多方面的能力達到甚至超越了人類。但在涉及到創造和情感方面，在可預見的未來機器是無法取代人的。而爲用戶提供有價值的信息和情感聯繫是好的、具備人文關懷的推薦系統必須要具備的能力，這就要求人和機器有效協同，這也是未來很長一段時間推薦系統的常態。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從數據過濾、特徵選擇、模型調整、結果乾預、展示優化、效果調控等推薦系統的各個維度人工都可以發揮極大的價值。《推薦系統的人工調控策略》中對人工怎麼調控推薦系統，人工在推薦系統中的定位和作用等進行了比較深入的介紹。加入了人工因素的推薦系統更有情感也更加可控。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"目前人工在推薦系統中所起的調控作用還比較粗暴，更多可能是對結果層面的干預，未來人工怎麼跟推薦系統更好地協同，怎樣在推薦系統中發揮人的創造力和情感力量是非常值得思考和探索的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"31.6 關注推薦系統多維價值體現","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統作爲一種獲取商業價值的工具，已經被過度商業化了。在用戶體驗上雖有所考慮、有所收斂，但是做得還不夠。作者在《推薦系統的價值觀》中對推薦系統應該從哪些維度來體現價值進行了比較深入的介紹，特別是人文關懷、生態健康發展和弘揚社會正向價值觀這3個方向上，給出了自己的思考，這也是當前推薦系統價值體現中非常缺失的部分。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着科技的發展，特別是雲計算將很多技術能力變成像水電煤一樣可以方便獲取的資源，大家在技術能力上的差異會越來越小，這時能夠讓你脫穎而出的可能是你的產品能不能打動用戶、能不能跟用戶產生共情。推薦系統作爲一個跟用戶強交互的產品，也是滿足這種趨勢變化的。因此，未來能夠做好推薦系統的企業一定是能夠定義好推薦系統價值的企業，不光要考慮商業價值，更應該考慮用戶體驗、情感鏈接和人文關懷。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"總結","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本章基於作者在推薦領域多年的實踐經驗和深入思考，從多個維度對推薦系統的未來發展進行了總結和介紹。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"國家層面對大數據與人工智能技術的大力支持，有利於推薦行業獲得更多的專業人才，同時競爭也明顯加劇。雲計算等技術的發展讓構建推薦系統就像購買商品一樣方便，創業公司可以更輕量、更便捷、低成本地在產品中整合推薦能力。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"政策層面的支持、技術的發展，對推薦行業就業也會產生深遠影響。企業更需要推薦算法的商業策略師更好將推薦算法落地到產品中，而推薦從業人員需要關注推薦系統的業務價值產出。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"物聯網、通信技術、硬件技術的發展，拓展了推薦系統的應用場景，推薦系統會應用到更多的領域，包括家庭場景、車載場景、虛擬現實場景等。在這些新場景中推薦系統與人的交互方式會發生極大的變化，語音交互、手勢交互等新交互方式成爲可能。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在推薦算法上，最近幾年深度學習已經在推薦上取得了非常好的效果，未來新的推薦範式，如強化學習、遷移學習、半監督學習、聯邦機器學習等都會在推薦系統中獲得規模化使用。技術的進步讓推薦系統利用更多的富媒體數據訓練模型成爲可能，推薦系統也會更加實時化、個性化，甚至可能每個用戶都會擁有一套量身定製的個性化推薦引擎。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推薦系統不僅要獲得商業價值，在用戶體驗、人文關懷、生態繁榮、弘揚正向價值觀等維度也需要有所突破，這些多維度的價值會越來越重要，會成爲推薦系統的核心競爭力。在這些價值發揮中人的作用就凸顯出來，未來很長一段時間，人與機器是協同發展的。推薦系統只有更多地注入人的情感和靈魂，才能做得更好，整個推薦行業纔會更加欣欣向榮。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"參考文獻","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"1. [開設大數據專業的高校] ","attrs":{}},{"type":"link","attrs":{"href":"https://jingyan.baidu.com/article/3d69c55125fd54f0cf02d718.html","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://jingyan.baidu.com/article/3d69c55125fd54f0cf02d718.html","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2. [開設人工智能專業的高校] ","attrs":{}},{"type":"link","attrs":{"href":"https://www.sohu.com/a/321455764_383037","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://www.sohu.com/a/321455764_383037","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"3. [2019 今日頭條] Deep Reinforcement Learning for Online Advertising in Recommender Systems ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"4. [2018 京東] Deep Reinforcement Learning for List-wise Recommendations ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"5. [2018 京東] Recommendations with Negative Feedback via Pairwise Deep Reinforcement Learning ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"6. [2018 YouTube] Top-K Off-Policy Correction for a REINFORCE Recommender System ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"7. [圖書：強化學習 第二版] ","attrs":{}},{"type":"link","attrs":{"href":"http://product.dangdang.com/27926613.html","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"http://product.dangdang.com/27926613.html","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"8. Social-behavior Transfer Learning for Recommendation Systems","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"9. [2012] Transfer Learning in Collaborative Filtering with Uncertain Ratings","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"10. [2012] Selective Transfer Learning for Cross Domain Recommendation","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"11. [2016] Transferring User Interests Across Websites with Unstructured Text for Cold-Start Recommendation","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"12. [2015] A Multi-View Deep Learning Approach for Cross Domain User Modeling in Recommendation Systems","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"13. [半監督學習] ","attrs":{}},{"type":"link","attrs":{"href":"https://baike.baidu.com/item/%25E5%258D%258A%25E7%259B%2591%25E7%259D%25A3%25E5%25AD%25A6%25E4%25B9%25A0/9075473?fr=aladdin","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://baike.baidu.com/item/%E5%8D%8A%E7%9B%91%E7%9D%A3%E5%AD%A6%E4%B9%A0/9075473?fr=aladdin","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"14. Bridging Collaborative Filtering and Semi-Supervised Learning: A Neural Approach for POI Recommendation","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"15. [聯邦機器學習]","attrs":{}},{"type":"link","attrs":{"href":"https://baike.baidu.com/item/%25E8%2581%2594%25E9%2582%25A6%25E6%259C%25BA%25E5%2599%25A8%25E5%25AD%25A6%25E4%25B9%25A0/23618046?fr=aladdin","title":null,"type":null},"content":[{"type":"text","marks":[{"type":"underline","attrs":{}}],"text":"https://baike.baidu.com/item/%E8%81%94%E9%82%A6%E6%9C%BA%E5%99%A8%E5%AD%A6%E4%B9%A0/23618046?fr=aladdin","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]}]}