2017年視頻推薦中用戶興趣建模、識別的挑戰和解法.pdf

1、視頻推薦搜索中的用戶興趣優酷 搜索、推薦、內容智能負責人 數據智能部總監 李玉Agenda優酷視頻個性化搜索推薦簡介 視頻個性化搜索推薦中的用戶興趣表達的挑戰 當前工業界常見方法的問題探討 我們的嘗試的方法優酷個性化服務簡介個性化服務在優酷Data一多半的視頻播放通過個性化搜索推薦技術分發 對于CTR、人均播放量、人均時長、留存率等均有顯著提升 幫助用戶發現好內容，幫助高質量內容觸達精準受眾6億+視頻5億+用戶Algo視頻推薦中用戶興趣表達的挑戰視頻推薦的用戶興趣表達的挑戰技術挑戰：劇、綜、影、漫：用戶選擇成本高，用戶追的劇、綜藝少，推薦成功率低 用戶目的性強，發現、瀏覽、逛的心智低 長節目可

2、選擇空間有限 頭部節目用戶行為稀疏，大量用戶每月只觀看3個以下節目，對比：短視頻信息流場景：通過數百個觀看行為推薦30個 優酷頭部節目：通過3、4個觀看行為推薦30個 數據噪聲多、分布驅熱、highly biased，常用推薦算法模型描述能力不足視頻推薦的用戶興趣表達的挑戰 cont.技術挑戰：視頻內容興趣復雜，感性、微妙、亞文化細分多樣，對于符合興趣大方向的驚喜度（serendipity）與多樣性要求更高，對比：電商：興趣明確：想買4K電視、牛仔褲、連衣裙；高度結構化，類目體系清晰 視頻：興趣感性、微妙：喜歡香港武俠片但是討厭成龍;喜歡日本動漫，今敏等、但討厭宮崎駿；興趣會進化、發展、細分，

3、如：相聲：郭德綱 小岳岳-方清平；或者-王玥波評書；或者-侯寶林 劉寶瑞 馬三立 傳統 科幻迷：從淺度：看星戰、地心引力-中度：星際穿越-深度：銀翼殺手、降臨、三體；微妙的亞文化：二次元、游戲、直播；文藝青年；腐、柜；追劇族、韓劇迷、恐怖片迷 興趣體現的是用戶的個人認同 興趣多維度正交，如：只看”大制作”、美劇質感 不喜歡重復，期待驚喜（serendipity）識別、表達用戶興趣的重要性Retargeting（看了又看）:推薦用戶有過交互的內容（看了又看）成功率高，長期價值低 局部提升非全局提升（搶其他渠道流量）成功率高因此ctr高 容易陷入局部最優 熱點推薦 推薦近期熱點 容易陷入局部最優

4、個性化興趣推薦 推薦符合每個用戶興趣的內容 成功率低因此ctr偏低 更具長期價值 短期收益可能小，但容易長期收斂 推薦命中成功率：retargeting 熱點 個性化發現 推薦命中（不命中）價值：個性化發現 推薦熱點 retargeting個性化內容推薦較少模型興趣預測不準確興趣命中少正樣本不足當前工業界常見方法的問題探討個性化推薦工業界常用方法流程：召回、排序 特征：統計特征 用戶畫像：DEMO、用戶對于標簽的frequency、recency 高維組合特征 Item based similarity(i2i)Common Algo Framework(對應的優酷的方法)DataMatchF

5、eatureRankRRankFTRL,DNN,XGBoost,FFMEnsembleRerankFFeatureItem/User/User2Item StatisticsUser Profile:(Demo,Interest profile,search profile,view history)Item tags,categories,topicsitem/tag/topic relevance scoresMMatchItem Based CF,DNN CFSlim CFTag to Item,User2user2ItemStar2ItemPopularity,TrendingDDat

6、aETLoffline/streaming常用方法對于表達用戶視頻興趣的問題Demo（年齡、性別、地域），設備類型、城市.問題：用戶的內容興趣與以上信息相關性不大 問題：三線城市50歲男性可能和一線城市30歲女性的觀看習慣一致 基于內容標簽的用戶畫像 人工內容標簽：恐怖片、動作片、搞笑、香港片、韓國片 Topic Modeling標簽：LDA提取視頻標題、描述的主題（內容數據噪聲大）基于統計的方法（frequency、recency）建立用戶標簽 問題：人工標簽主觀性大、噪聲大 問題：人工標簽粒度容易過于寬泛 問題：topic modeling標簽噪聲大、數據稀疏 問題：往往基于統計的方法，很

7、難精準描述用戶的興趣 問題：容易受到驅熱的影響常用方法對于表達用戶興趣的問題 cont.高維組合特征 通過組合以上各種特征，產生更豐富的信息 問題：容易受到噪聲影響 問題：計算量過大 Item based similarity(i2i)CF similarity SVD+/MF Slim DNN 簡單高效Problem of I2IItem based CF是學術和工業界都最有效的方法之一 Item based方法比User based方法更有效。主要因為user 維度行為更稀疏，噪聲更大。Item的維度積累歷史行為更多，variance更小。問題1:由于基于item維度的全局統計，每個用戶觀

8、看item的不同原因信息被平均掉。對于一個視頻，有的用戶因為熱度觀看，有的用戶因為主題的類型觀看，有的用戶因為主演、導演觀看。問題2:不同用戶群體的不同喜好在全局Item similarity的計算過程中被平滑掉。問題3:對于長尾item行為數據過于稀疏 問題4:粒度太細，數據稀疏，擴展能力弱 問題5:驅熱、哈利波特現象介紹我們的一些嘗試基礎用戶畫像做法用戶 興趣畫像用戶觀看行為內容標簽內容的標簽、類目體系演員、導演等Metadata內容針對每個標簽、類目的興趣強度分興趣畫像用戶對于各類標簽觀看的FrequencyRecency用戶觀看行為問題：基于統計，無法區分驅熱、類型、明星等信息 粒度過

9、于粗User Interest Latent VectorEnd2End 黑盒模型由于噪聲與概率分布假設的問題并非全局收斂，需縮小搜索空間 拆解為多個更容易的子問題 機器學習解一個End2End大問題 拆解為若干個更容易的小問題 傳統End2End方法易受數據稀疏與噪聲影響：End2End模型：觀看歷史節目推薦，易受噪聲影響 拆解為子問題預測模型：觀看歷史寬泛興趣分類Latent Vector節目推薦，對于噪聲更魯邦 寬泛興趣Latent vector人工構建類目體系+審核，降噪?LatentVector?用戶興趣的建模的work-CTRCollaborative Topic Modeling

10、 for Recommending Scientific Articles 用戶興趣的建模的work-CTPFContent-based recommendations with Poisson factorization A Practical Algorithm for Solving the Incoherence Problem of Topic Models In Industrial Applications 用戶興趣的建模的work-CTPF with popularity,stars tags and queries實現性能優化，scalable to internet sca

11、le 基于parameter server架構的分布式實現 EM不是全局收斂。針對每個topic進行人工審核，再作為初始值進行迭代。擴展到文本+標簽+meta+流行度 基于興趣向量的個性化I2I similarity長期興趣與短期興趣的平衡Phased GRU RecNetBased on:SESSION-BASED RECOMMENDATIONS WITH RECURRENT NEURAL NETWORKS-ICLR2016Listwise Loss:BPR/TOP1 Loss捕捉用戶興趣中的時域規律：長期短期平衡 有一些短期興趣滿足后，多樣性需求會變強過一段時間需求又會周期性的出現長期興趣

12、與短期興趣的平衡Phased GRU RecNet cont.GRU:默認的假設是等距采樣:Published as a conference paper at ICLR 20162.2DEEPLEARNING INRECOMMENDERSOne of the first related methods in the neural networks literature where the use of Restricted Boltz-mann Machines(RBM)for Collaborative Filtering(Salakhutdinov et al.,2007).In thi

13、s work anRBM is used to model user-item interaction and perform recommendations.This model has beenshown to be one of the best performing Collaborative Filtering models.Deep Models have been usedtoextractfeaturesfromunstructuredcontentsuchasmusicorimagesthatarethenusedtogetherwithmore conventional c

14、ollaborative filtering models.In Van den Oord et al.(2013)a convolutional deepnetwork is used to extract feature from music files that are then used in a factor model.More recentlyWang et al.(2015)introduced a more generic approach whereby a deep network is used to extractgeneric content-features fr

15、om any types of items,these features are then incorporated in a standardcollaborative filtering model to enhance the recommendation performance.This approach seems tobe particularly useful in settings where there is not sufficient user-item interaction information.3RECOMMENDATIONS WITHRNNSRecurrent

16、Neural Networks have been devised to model variable-length sequence data.The maindifference between RNNs and conventional feedforward deep models is the existence of an internalhidden state in the units that compose the network.Standard RNNs update their hidden state h usingthe following update func

17、tion:ht=g(Wxt+Uht?1)(1)Where g is a smooth and bounded function such as a logistic sigmoid function xtis the input ofthe unit at time t.An RNN outputs a probability distribution over the next element of the sequence,given its current state ht.A Gated Recurrent Unit(GRU)(Cho et al.,2014)is a more ela

18、borate model of an RNN unit thataims at dealing with the vanishing gradient problem.GRU gates essentially learn when and by howmuch to update the hidden state of the unit.The activation of the GRU is a linear interpolationbetween the previous activation and the candidate activationht:ht=(1?zt)ht?1+z

19、tht(2)where the update gate is given by:zt=?(Wzxt+Uzht?1)(3)while the candidate activation functionhtis computed in a similar manner:ht=tanh(Wxt+U(rt?ht?1)(4)and finaly the reset gate rtis given by:rt=?(Wrxt+Urht?1)(5)3.1CUSTOMIZING THEGRUMODELWe used the GRU-based RNN in our models for session-base

20、d recommendations.The input of thenetwork is the actual state of the session while the output is the item of the next event in the session.The state of the session can either be the item of the actual event or the events in the session sofar.In the former case 1-of-N encoding is used,i.e.the input v

21、ectors length equals to the numberof items and only the coordinate corresponding to the active item is one,the others are zeros.Thelatter setting uses a weighted sum of these representations,in which events are discounted if theyhave occurred earlier.For the stake of stability,the input vector is th

22、en normalized.We expect thisto help because it reinforces the memory effect:the reinforcement of very local ordering constraintswhich are not well captured by the longer memory of RNN.We also experimented with adding anadditional embedding layer,but the 1-of-N encoding always performed better.The co

23、re of the network is the GRU layer(s)and additional feedforward layers can be added betweenthe last layer and the output.The output is the predicted preference of the items,i.e.the likelihoodof being the next in the session for each item.When multiple GRU layers are used,the hiddenstate of the previ

24、ous layer is the input of the next one.The input can also be optionally connected3Published as a conference paper at ICLR 20162.2DEEPLEARNING INRECOMMENDERSOne of the first related methods in the neural networks literature where the use of Restricted Boltz-mann Machines(RBM)for Collaborative Filteri

25、ng(Salakhutdinov et al.,2007).In this work anRBM is used to model user-item interaction and perform recommendations.This model has beenshown to be one of the best performing Collaborative Filtering models.Deep Models have been usedtoextractfeaturesfromunstructuredcontentsuchasmusicorimagesthatarethe

26、nusedtogetherwithmore conventional collaborative filtering models.In Van den Oord et al.(2013)a convolutional deepnetwork is used to extract feature from music files that are then used in a factor model.More recentlyWang et al.(2015)introduced a more generic approach whereby a deep network is used t

27、o extractgeneric content-features from any types of items,these features are then incorporated in a standardcollaborative filtering model to enhance the recommendation performance.This approach seems tobe particularly useful in settings where there is not sufficient user-item interaction information

28、.3RECOMMENDATIONS WITHRNNSRecurrent Neural Networks have been devised to model variable-length sequence data.The maindifference between RNNs and conventional feedforward deep models is the existence of an internalhidden state in the units that compose the network.Standard RNNs update their hidden st

29、ate h usingthe following update function:ht=g(Wxt+Uht?1)(1)Where g is a smooth and bounded function such as a logistic sigmoid function xtis the input ofthe unit at time t.An RNN outputs a probability distribution over the next element of the sequence,given its current state ht.A Gated Recurrent Uni

30、t(GRU)(Cho et al.,2014)is a more elaborate model of an RNN unit thataims at dealing with the vanishing gradient problem.GRU gates essentially learn when and by howmuch to update the hidden state of the unit.The activation of the GRU is a linear interpolationbetween the previous activation and the ca

31、ndidate activationht:ht=(1?zt)ht?1+ztht(2)where the update gate is given by:zt=?(Wzxt+Uzht?1)(3)while the candidate activation functionhtis computed in a similar manner:ht=tanh(Wxt+U(rt?ht?1)(4)and finaly the reset gate rtis given by:rt=?(Wrxt+Urht?1)(5)3.1CUSTOMIZING THEGRUMODELWe used the GRU-base

32、d RNN in our models for session-based recommendations.The input of thenetwork is the actual state of the session while the output is the item of the next event in the session.The state of the session can either be the item of the actual event or the events in the session sofar.In the former case 1-o

33、f-N encoding is used,i.e.the input vectors length equals to the numberof items and only the coordinate corresponding to the active item is one,the others are zeros.Thelatter setting uses a weighted sum of these representations,in which events are discounted if theyhave occurred earlier.For the stake

34、 of stability,the input vector is then normalized.We expect thisto help because it reinforces the memory effect:the reinforcement of very local ordering constraintswhich are not well captured by the longer memory of RNN.We also experimented with adding anadditional embedding layer,but the 1-of-N enc

35、oding always performed better.The core of the network is the GRU layer(s)and additional feedforward layers can be added betweenthe last layer and the output.The output is the predicted preference of the items,i.e.the likelihoodof being the next in the session for each item.When multiple GRU layers a

36、re used,the hiddenstate of the previous layer is the input of the next one.The input can also be optionally connected3Published as a conference paper at ICLR 20162.2DEEPLEARNING INRECOMMENDERSOne of the first related methods in the neural networks literature where the use of Restricted Boltz-mann Ma

37、chines(RBM)for Collaborative Filtering(Salakhutdinov et al.,2007).In this work anRBM is used to model user-item interaction and perform recommendations.This model has beenshown to be one of the best performing Collaborative Filtering models.Deep Models have been usedtoextractfeaturesfromunstructured

38、contentsuchasmusicorimagesthatarethenusedtogetherwithmore conventional collaborative filtering models.In Van den Oord et al.(2013)a convolutional deepnetwork is used to extract feature from music files that are then used in a factor model.More recentlyWang et al.(2015)introduced a more generic appro

39、ach whereby a deep network is used to extractgeneric content-features from any types of items,these features are then incorporated in a standardcollaborative filtering model to enhance the recommendation performance.This approach seems tobe particularly useful in settings where there is not sufficie

40、nt user-item interaction information.3RECOMMENDATIONS WITHRNNSRecurrent Neural Networks have been devised to model variable-length sequence data.The maindifference between RNNs and conventional feedforward deep models is the existence of an internalhidden state in the units that compose the network.

41、Standard RNNs update their hidden state h usingthe following update function:ht=g(Wxt+Uht?1)(1)Where g is a smooth and bounded function such as a logistic sigmoid function xtis the input ofthe unit at time t.An RNN outputs a probability distribution over the next element of the sequence,given its cu

42、rrent state ht.A Gated Recurrent Unit(GRU)(Cho et al.,2014)is a more elaborate model of an RNN unit thataims at dealing with the vanishing gradient problem.GRU gates essentially learn when and by howmuch to update the hidden state of the unit.The activation of the GRU is a linear interpolationbetwee

43、n the previous activation and the candidate activationht:ht=(1?zt)ht?1+ztht(2)where the update gate is given by:zt=?(Wzxt+Uzht?1)(3)while the candidate activation functionhtis computed in a similar manner:ht=tanh(Wxt+U(rt?ht?1)(4)and finaly the reset gate rtis given by:rt=?(Wrxt+Urht?1)(5)3.1CUSTOMI

44、ZING THEGRUMODELWe used the GRU-based RNN in our models for session-based recommendations.The input of thenetwork is the actual state of the session while the output is the item of the next event in the session.The state of the session can either be the item of the actual event or the events in the

45、session sofar.In the former case 1-of-N encoding is used,i.e.the input vectors length equals to the numberof items and only the coordinate corresponding to the active item is one,the others are zeros.Thelatter setting uses a weighted sum of these representations,in which events are discounted if the

46、yhave occurred earlier.For the stake of stability,the input vector is then normalized.We expect thisto help because it reinforces the memory effect:the reinforcement of very local ordering constraintswhich are not well captured by the longer memory of RNN.We also experimented with adding anadditiona

47、l embedding layer,but the 1-of-N encoding always performed better.The core of the network is the GRU layer(s)and additional feedforward layers can be added betweenthe last layer and the output.The output is the predicted preference of the items,i.e.the likelihoodof being the next in the session for

48、each item.When multiple GRU layers are used,the hiddenstate of the previous layer is the input of the next one.The input can also be optionally connected3Published as a conference paper at ICLR 20162.2DEEPLEARNING INRECOMMENDERSOne of the first related methods in the neural networks literature where

49、 the use of Restricted Boltz-mann Machines(RBM)for Collaborative Filtering(Salakhutdinov et al.,2007).In this work anRBM is used to model user-item interaction and perform recommendations.This model has beenshown to be one of the best performing Collaborative Filtering models.Deep Models have been u

50、sedtoextractfeaturesfromunstructuredcontentsuchasmusicorimagesthatarethenusedtogetherwithmore conventional collaborative filtering models.In Van den Oord et al.(2013)a convolutional deepnetwork is used to extract feature from music files that are then used in a factor model.More recentlyWang et al.(

51、2015)introduced a more generic approach whereby a deep network is used to extractgeneric content-features from any types of items,these features are then incorporated in a standardcollaborative filtering model to enhance the recommendation performance.This approach seems tobe particularly useful in

52、settings where there is not sufficient user-item interaction information.3RECOMMENDATIONS WITHRNNSRecurrent Neural Networks have been devised to model variable-length sequence data.The maindifference between RNNs and conventional feedforward deep models is the existence of an internalhidden state in

53、 the units that compose the network.Standard RNNs update their hidden state h usingthe following update function:ht=g(Wxt+Uht?1)(1)Where g is a smooth and bounded function such as a logistic sigmoid function xtis the input ofthe unit at time t.An RNN outputs a probability distribution over the next

54、element of the sequence,given its current state ht.A Gated Recurrent Unit(GRU)(Cho et al.,2014)is a more elaborate model of an RNN unit thataims at dealing with the vanishing gradient problem.GRU gates essentially learn when and by howmuch to update the hidden state of the unit.The activation of the

55、 GRU is a linear interpolationbetween the previous activation and the candidate activationht:ht=(1?zt)ht?1+ztht(2)where the update gate is given by:zt=?(Wzxt+Uzht?1)(3)while the candidate activation functionhtis computed in a similar manner:ht=tanh(Wxt+U(rt?ht?1)(4)and finaly the reset gate rtis giv

56、en by:rt=?(Wrxt+Urht?1)(5)3.1CUSTOMIZING THEGRUMODELWe used the GRU-based RNN in our models for session-based recommendations.The input of thenetwork is the actual state of the session while the output is the item of the next event in the session.The state of the session can either be the item of th

57、e actual event or the events in the session sofar.In the former case 1-of-N encoding is used,i.e.the input vectors length equals to the numberof items and only the coordinate corresponding to the active item is one,the others are zeros.Thelatter setting uses a weighted sum of these representations,i

58、n which events are discounted if theyhave occurred earlier.For the stake of stability,the input vector is then normalized.We expect thisto help because it reinforces the memory effect:the reinforcement of very local ordering constraintswhich are not well captured by the longer memory of RNN.We also

59、experimented with adding anadditional embedding layer,but the 1-of-N encoding always performed better.The core of the network is the GRU layer(s)and additional feedforward layers can be added betweenthe last layer and the output.The output is the predicted preference of the items,i.e.the likelihoodo

60、f being the next in the session for each item.When multiple GRU layers are used,the hiddenstate of the previous layer is the input of the next one.The input can also be optionally connected3Published as a conference paper at ICLR 20162.2DEEPLEARNING INRECOMMENDERSOne of the first related methods in

61、the neural networks literature where the use of Restricted Boltz-mann Machines(RBM)for Collaborative Filtering(Salakhutdinov et al.,2007).In this work anRBM is used to model user-item interaction and perform recommendations.This model has beenshown to be one of the best performing Collaborative Filt

62、ering models.Deep Models have been usedtoextractfeaturesfromunstructuredcontentsuchasmusicorimagesthatarethenusedtogetherwithmore conventional collaborative filtering models.In Van den Oord et al.(2013)a convolutional deepnetwork is used to extract feature from music files that are then used in a fa

63、ctor model.More recentlyWang et al.(2015)introduced a more generic approach whereby a deep network is used to extractgeneric content-features from any types of items,these features are then incorporated in a standardcollaborative filtering model to enhance the recommendation performance.This approac

64、h seems tobe particularly useful in settings where there is not sufficient user-item interaction information.3RECOMMENDATIONS WITHRNNSRecurrent Neural Networks have been devised to model variable-length sequence data.The maindifference between RNNs and conventional feedforward deep models is the exi

65、stence of an internalhidden state in the units that compose the network.Standard RNNs update their hidden state h usingthe following update function:ht=g(Wxt+Uht?1)(1)Where g is a smooth and bounded function such as a logistic sigmoid function xtis the input ofthe unit at time t.An RNN outputs a pro

66、bability distribution over the next element of the sequence,given its current state ht.A Gated Recurrent Unit(GRU)(Cho et al.,2014)is a more elaborate model of an RNN unit thataims at dealing with the vanishing gradient problem.GRU gates essentially learn when and by howmuch to update the hidden sta

67、te of the unit.The activation of the GRU is a linear interpolationbetween the previous activation and the candidate activationht:ht=(1?zt)ht?1+ztht(2)where the update gate is given by:zt=?(Wzxt+Uzht?1)(3)while the candidate activation functionhtis computed in a similar manner:ht=tanh(Wxt+U(rt?ht?1)(

68、4)and finaly the reset gate rtis given by:rt=?(Wrxt+Urht?1)(5)3.1CUSTOMIZING THEGRUMODELWe used the GRU-based RNN in our models for session-based recommendations.The input of thenetwork is the actual state of the session while the output is the item of the next event in the session.The state of the

69、session can either be the item of the actual event or the events in the session sofar.In the former case 1-of-N encoding is used,i.e.the input vectors length equals to the numberof items and only the coordinate corresponding to the active item is one,the others are zeros.Thelatter setting uses a wei

70、ghted sum of these representations,in which events are discounted if theyhave occurred earlier.For the stake of stability,the input vector is then normalized.We expect thisto help because it reinforces the memory effect:the reinforcement of very local ordering constraintswhich are not well captured

71、by the longer memory of RNN.We also experimented with adding anadditional embedding layer,but the 1-of-N encoding always performed better.The core of the network is the GRU layer(s)and additional feedforward layers can be added betweenthe last layer and the output.The output is the predicted prefere

72、nce of the items,i.e.the likelihoodof being the next in the session for each item.When multiple GRU layers are used,the hiddenstate of the previous layer is the input of the next one.The input can also be optionally connected3reset gateupdate gate長期興趣與短期興趣的平衡Phased GRU RecNet cont.用戶session實際情況是有的se

73、ssion一天100個行為，有的session一個月只有一個行為 Phased GRU,引入time gate k，根據采樣間隔控制變量的更新(同時增加一定程度的采樣間隔):Based on:Phased LSTM:Accelerating Recurrent Network Training for Long or Event-based Sequences 基于傳染病模型的有限行為用戶興趣預測大量用戶行為非常稀疏，每月觀看量不超過3次 用戶群體的興趣演變遵循類似傳染病傳播的機制 預測：?基于Nystrom CUR的explorationNxN的I2I矩陣有很多元素很稀疏，explore收集

74、數據需要很多流量，代價很高 Nystrom CUR:可以用c個landmark item來代表整個I2I相似度矩陣 通過statistical leverage score選擇c個item 重點explore對于c個item有過觀看的用戶nnc基于HIN圖、聚類等方法的興趣識別利用用戶與節目的播放記錄構建二部圖，每個節點的標簽按相似度傳播給相鄰節點，在節點傳播的每一步，每個節點按照相鄰節點的標簽來更新自己的標簽。與該節點相似度越大，其相鄰節點對其標注的影響權值也越大。當絕大多數節點的標簽不再更新時，整個網絡按照標簽就形成了各自所屬的社區。算法思想權重設定Item節點的權重為該節目觀看人數的倒數

75、 User節點的權重為該用戶觀看節目數量的倒數 U-I連邊的權重為該用戶對該節目的觀看完成率 U-I連邊的權重加入隨機因子效果評估Item在類簇中的掛載成功率為100%僅有單個Item掛載的類簇占99.48%，最多一個類簇內包含32個節目類簇內包含的用戶個數的分布直方圖如右所示，其中最大的類簇包含用戶45313個將全部用戶劃分為35830個類簇典型CASE序號節ID節名稱1323580汽車城之建筑隊2323577汽車城之車特洛伊3318953和迷你卡車學習4323581汽車城之湯姆的油漆店5323573汽車城之超級變形卡車6323571汽車城之拖車湯姆Hierarchical View Fee

76、dback Aggregation算法模型能力有限，End2End模型精準capture個性化特征能力有限 最優解在非常高緯空間中，由于噪聲與模型收斂能力問題，需人工輔助降低搜索空間維度 使用交叉特征的統計值，效果好于使用離散交叉裸id特征 結合業務理解，輔助模型更好capture個性化特征 結合統計量的variance進行噪聲過濾 交叉統計：更好capture不同用戶群體對于不同視頻類型的興趣，如：愛看韓劇的人群對于臺灣偶像劇的人均vv；愛看日本恐怖片的人群對于美國恐怖片的人均vv；20歲一線城市女性看游戲人均vvUser Age Gender Geo Video Tag Popularit

77、y Category Source Exclusive Purchased User Interest Category Topic Tag Match Type Relevance Popularity Trending Context Time of day Day of week Location User id 個性化排序在優酷視頻搜索稀疏全連接域內信息的二次編碼concat全局全連接個性化排序在優酷視頻搜索-特征域劃分及編碼query user video id域 統計域 用戶觀看序列 標簽興趣 文本 超高維的稀疏編碼來表征獨立個體 利用神經網絡來擬合個體共性 視頻表達是基礎 按特征的重要度和關聯性分域 億級參數 挑戰：特征維度高 模型存儲空間大，離線訓練計算時間成本高，在線實現資源占用高，前向網絡計算不能滿足RT要求 特征分域 隨機編碼 掛靠編碼 抽樣技術We Are Hiring ly136216alibaba-T hanks