算法如何促成億級別扶持曝光視頻？愛奇藝Budget Pacing智能扶植系統實踐

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"視頻的保量和扶植對於視頻平臺而言是一種重要的分發策略。不同於普通的推薦算法，出於商業、畫風等角度的考量，視頻扶植往往要保證指定的視頻物料有一定的曝光量，並儘可能地降低對推薦系統效果的負面影響。在本文中，我們主要介紹愛奇藝自研的基於改進的Budget Pacing算法的智能扶植系統是如何在保證對用戶消費指標影響可控的前提下，達到最大量級和最高效的扶持效果。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"背景"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於愛奇藝等在線視頻服務平臺來說，除了普通的推薦算法之外，出於各種考量，往往要求對於指定的新熱視頻增加曝光。"}]},{"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":"在具體的應用場景下，愛奇藝和隨刻短視頻流中，超級粉絲、電商視頻以及各類劇集綜藝的運營物料視頻等等，均需要使用扶植保量保證播放量的最大化。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/11\/11ecb4fcae09f12b5f8263007387232e.png","alt":"圖片","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":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/4a\/4a863bc0af7a12e88a0e013684f10ce7.png","alt":"圖片","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":"center","origin":null},"content":[{"type":"text","text":"圖一 feed流中的扶植保量視頻"}]},{"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":"feed流的曝光資源是有限的，內容間存在競爭，對扶植視頻增加曝光必然會擠壓推薦算法產出視頻的曝光資源，對推薦系統產生負面影響。"}]},{"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":"另一方面扶植視頻間也有差異，有的視頻保量額度消耗過快，甚至在最初的幾分鐘內就完成保量目標，早早退出競爭，無法觸達更匹配的用戶，降低消費指標；而另一些視頻又得不到有效分發。這就是典型的扶植保量問題。"}]},{"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":"因此我們設計了生態智能扶植系統，直接進行視頻級別的扶植保量，針對不同視頻的保量目標和真實的消費情況進行定製化的扶植，且所有保量視頻共同競爭恆定的展示資源，降低對消費的影響。另外採用類似計算廣告領域的Budget Pacing機制，將保量展示目標按照實際流量曲線分配在保量週期內，且根據具體消費情況動態調整保量程度，使視頻在保量週期內均勻分發，完成保量目標。"}]},{"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":"另外生態智能扶植系統使用提升視頻在精排中的位置的方式爲視頻提權，保證其扶植保量效果的確定性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/80\/80a4cbbbd82c4534ad1b8bede374d4f1.png","alt":"圖片","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":"center","origin":null},"content":[{"type":"text","text":"圖2:生態智能扶植系統流程結構圖"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"問題分析"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們認爲，視頻的扶植保量和計算廣告領域的預算控制問題有相似的地方：預算控制的核心問題是如何平穩地消費掉廣告主的預算，並幫助廣告主優化轉化效果。而視頻的扶植保量的核心問題是如何平穩的消耗掉視頻的目標曝光量，並保證將對整體的消費影響儘可能降低。所以視頻的扶植保量要完成以下兩個目標："}]},{"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","marks":[{"type":"strong"}],"text":"（1）勻速分發"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"通過視頻目標曝光量、當前曝光量和曝光曲線來控制視頻的投放速度；"}]},{"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","marks":[{"type":"strong"}],"text":"（2）提升視頻消費指標"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"保證扶植視頻對整體消費影響儘可能低。"}]},{"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":"計算廣告的預算控制方案目前分爲兩大類：Probabilistic throttling 和 Bid modification，Probabilistic throttling 通過一個概率控制廣告參與競價的頻率從而達到控制預算消費速度的目的，而 Bid modification 則通過修改競價的方式來控制預算消費速度。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/ea\/ea88c1b696027312fc3596105511438b.png","alt":"圖片","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":"center","origin":null},"content":[{"type":"text","text":"圖4(a)：Probabilistic throttling"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/d5\/d5c88dfe1e5c4ba836040a56ef651468.png","alt":"圖片","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":"center","origin":null},"content":[{"type":"text","text":"圖4(b):Bid modification"}]},{"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":"對於視頻扶植保量場景而言，視頻的目標曝光量即爲“預算”，因此每次曝光的“競價”是恆定的數值，即爲曝光次數，不能通過Bid modification來修改“競價”，因此我們採用Probabilistic throttling作爲基礎框架設計智能扶植系統。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"框架設計"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"扶植保量系統首先要保證扶植保量視頻的分發趨勢和大盤整體曝光趨勢保持一致。我們以5分鐘爲單位，將一天拆分爲288個時間片，根據歷史數據統計出一天內大盤整體曝光趨勢。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/c8\/c8b2fa873bb42913f9618e915353b3be.png","alt":"圖片","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":"center","origin":null},"content":[{"type":"text","text":"圖4(a)：時間片-流量曲線"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/b1\/b1225c5f3cfd2c4d347cbcdba9267628.png","alt":"圖片","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":"center","origin":null},"content":[{"type":"text","text":"圖片4(b)：時間片-流量累計曲線"}]},{"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":"值得注意的是，在原始的計算廣告的Probabilistic throttling方案中，廣告物料享有單獨的展示資源，因此廣告物料之間只需通過概率進行競爭，勝者則可以佔有專門爲廣告預留的展示位。"}]},{"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":"但在視頻扶植保量的場景中，不僅有扶植保量視頻之間競爭保量機會，扶植保量視頻也會和普通視頻競爭，因此除了要使用概率之外，我們又引入了精排位置提權機制，對於需要扶植保量的視頻，根據目標曝光量，直接提升視頻在精排中的排序位置，使得扶植保量視頻在與普通視頻的競爭中更佔優勢，從而達到擴大曝光的目的。"}]},{"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":"因此我們需要根據歷史數據統計出精排中排序位置和曝光量的關係曲線。這裏需要注意的是，由於不同時間片的曝光效能不同（如晚高峯的曝光效能大於凌晨），所以每一個時間片都有一個獨立的位置-曝光曲線。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/b7\/b70cb968dde43a54f8051f457a6ba1e8.png","alt":"圖片","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":"center","origin":null},"content":[{"type":"text","text":"圖5:位置-曝光曲線"}]},{"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":"則有具體算法如下："}]},{"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":"對任意扶植保量視頻  ，記其全天預期保量曝光爲  ，將一天拆分爲時間片集合  ，對於時間片  ，  表示截止時間片  結束時的"},{"type":"text","marks":[{"type":"strong"}],"text":"累積預期曝光"},{"type":"text","text":"，則根據圖四(b)中描述的累計曝光曲線，可將  拆分爲各個時間片的累計曝光集合  ，其中  。同時對於時間片  ，亦有  表示截止時間片  結束時的"},{"type":"text","marks":[{"type":"strong"}],"text":"累積真實曝光"},{"type":"text","text":"。"}]},{"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":"對下一個即將來到的時間片  ，該時間片內的"},{"type":"text","marks":[{"type":"strong"}],"text":"預期曝光"},{"type":"text","text":"  ，根據圖五可得滿足  曝光要求的最大的位置。同樣的，我們也可求得上一個剛結束的時間片  ，該時間片內的"},{"type":"text","marks":[{"type":"strong"}],"text":"真實曝光"},{"type":"text","text":"  ，根據圖五亦可得對應  曝光的最大的位置，設爲對於時間片  ，扶植保量視頻達到其預期曝光所需要的達到的精排位置的均值，而對於時間片  ，扶植保量視頻實際達到的精排位置的均值，因此在時間片，我們需要 在每一次精排結果中提升的位置。"}]},{"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":"其中，保證 σ 表示對超量分發的視頻不進行降權。"}]},{"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":"此外，由於圖五中精排位置與曝光量的對應關係是離散的，  所對應的曝光量（設爲  ）一般情況下會大於  的要求，因此爲了防止扶植保量的超發，約定扶植保量視頻  在時間片  參與競爭的概率爲："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"最後推薦引擎便可在任意時間片  對任意視頻  ，根據 σ 和  進行精排位置調整，從而達到扶植保量的效果。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"線上效果"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"日均扶持量："},{"type":"text","text":"生態智能扶植系統單日扶持曝光達到億級別。"}]},{"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","marks":[{"type":"strong"}],"text":"曝光完成率："},{"type":"text","text":"舊版保量曝光完成率（實際曝光\/曝光目標>80%）低於5%，生態智能扶植系統保量完成度大幅提升，隨刻端達到65-70%，基線端達到50-60%，保量完成率提升20倍左右。"}]},{"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","marks":[{"type":"strong"}],"text":"對系統影響："},{"type":"text","text":"實現扶持保量視頻分發智能控速，將對系統的人均播放時長影響從3.5%降低到2.5%，人均展示視頻數的影響從2%降低到0.15%。"}]},{"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","marks":[{"type":"strong"}],"text":"運營業務成果："},{"type":"text","text":"助力《奇葩說》等項目以及百科、開箱等垂類優質內容分發。保證了對優質視頻和優質創作者的曝光，對提高優質作者的平臺粘性有重大意義。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"總結與展望"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"以上是我們近期在生態扶植保量領域的一些工作，實踐證明基於Budget Pacing的生態智能扶植系統確實能夠對視頻的扶植保量任務產生很大的助益。後續我們還會從以下幾個角度進行優化："}]},{"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":"（1）秩序優化扶植系統和相關扶植召回，保證扶植完成度的基礎上進一步降低對整體消費的影響；"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（2）當前的扶植系統類似於計算廣告的競價消費方式，無法保證視頻的目標曝光量一定完成，考慮增加擔保式保量投放機制，保證視頻曝光的數量。"}]},{"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","marks":[{"type":"strong"}],"text":"參考文獻："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"1. Agarwal D, Ghosh S, Wei K, et al. Budget pacing for targeted online advertisements at linkedin[C]\/\/Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. 2014: 1613-1619."}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2.Xu J, Lee K, Li W, et al. Smart pacing for effective online ad campaign optimization[C]\/\/Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2015: 2217-2226."}]},{"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":"本文轉載自：愛奇藝技術產品團隊（ID：iQIYI-TP）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文鏈接："},{"type":"link","attrs":{"href":"https:\/\/mp.weixin.qq.com\/s\/Jf_lBpjuOemINzyNNYzLDQ","title":"xxx","type":null},"content":[{"type":"text","text":"算法如何促成億級別扶持曝光視頻？愛奇藝Budget Pacing智能扶植系統實踐"}]}]}]}