圖14所示：圖1.3用戶長期偏好Fig1.3Longtermuserpreference說明：圖中表示用戶在時刻的簽到興趣點ID的嵌入向量表示，|P|用戶訪問過的興趣點總和。2.用戶的短期和長期偏好建模用戶的短期偏好和當前偏好使用兩個不同LSTM模型進行分別建模，用戶短期偏好和當前偏好的輸入都包括簽到興趣點和簽到時間信息，主要區(qū)別在于輸入的數(shù)據(jù)時期不同。用戶短期偏好由用戶最近2個月內的簽到數(shù)據(jù)組成，輸入數(shù)據(jù)包括用戶簽到興趣點和簽到時間信息。簽到時間通過轉化進一步劃分為簽到星期和一天中的簽到時刻，簽到星期Week分為7天，分別用戶對應的one-hot編碼表示，例如周一表示為[1,0,0,0,0,0,0]。同理，簽到時刻Hour分為24小時，也用不同的one-hot編碼表示。用戶短期偏好模型的LSTM的時刻的簽到信息輸入可以表示為當前時刻簽到興趣點向量，簽到星期向量和簽到時刻向量的拼接得到，表示為[]。用戶的目前的偏好類似于用戶短期偏好，本文選用用戶最后一次簽到的6個小時內的簽到興趣點記錄作為當前偏好模型的輸入。圖1.SEQ圖3.\*ARABIC5用戶短期偏好建模圖Fig1.SEQFig3.\*ARABIC5Usershorttermpreferencemodelingdiagram啟發(fā)自ST-LSTM模型，融合對用戶具有一定的影響周期偏好對于下一個興趣點建模。在ST-LSTM中，ST-LSTM嘗試將用戶的簽到記錄按天劃分，通過獲取用戶歷史每天的簽到興趣點向量的平均作為用戶當天的信息表示。不用于它的模型，本文使用用戶最近的2個月內的簽到，并通過融入周期偏好因素建模ST-LSTPM，用以表示用戶歷史的簽到信息，這樣做的好處是避免了由于按天分割導致無法捕獲天于天之間的連續(xù)性，同時簽到時間信息和周期偏好信息也能輕易的整合。1.用戶的周期偏好建模本文利用分層注意力機制通過根據(jù)用戶的短期偏好輸出和用戶的當前偏好輸出得到用戶的周期偏好向量。用戶的周期偏好計算如下：(SEQ____________\*ARABIC1.7)(SEQ____________\*ARABIC1.8)(SEQ____________\*ARABIC1.9)其中表是短期偏好模型的每一時刻的輸出，是經(jīng)過一層全連接神經(jīng)網(wǎng)絡后的向量表示。在文獻ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"DOI":"10.18653/v1/n16-1174","ISBN":"9781941643914","abstract":"Weproposeanovelattentionmodelthatcanaccuratelyattendstotargetobjectsofvariousscalesandshapesinimages.Themodelistrainedtograduallysuppressirrelevantregionsinaninputimageviaaprogressiveattentiveprocessovermultiplelayersofaconvolutionalneuralnetwork.Theattentiveprocessineachlayerdetermineswhethertopassorblockfeaturesatcertainspatiallocationsforuseinthesubsequentlayers.Theproposedprogressiveattentionmechanismworkswellespeciallywhencombinedwithhardattention.Wefurtheremploylocalcontextstoincorporateneighborhoodfeaturesofeachlocationandestimateabetterattentionprobabilitymap.Theexperimentsonsyntheticandrealdatasetsshowthattheproposedattentionnetworksoutperformtraditionalattentionmethodsinvisualattributepredictiontasks.","author":[{"dropping-particle":"","family":"Yang","given":"Zichao","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Yang","given":"Diyi","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Dyer","given":"Chris","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"He","given":"Xiaodong","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Smola","given":"Alex","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Hovy","given":"Eduard","non-dropping-particle":"","parse-names":false,"suffix":""}],"container-title":"2016ConferenceoftheNorthAmericanChapteroftheAssociationforComputationalLinguistics:HumanLanguageTechnologies,NAACLHLT2016-ProceedingsoftheConference","id":"ITEM-1","issued":{"date-parts":[["2016"]]},"page":"1480-1489","title":"HierarchicalAttentionNetworksforDocumentClassification","type":"article-journal"},"uris":["/documents/?uuid=d089f061-5616-4c1e-89c4-bb1fb128f367"]}],"mendeley":{"formattedCitation":"^[57]","plainTextFormattedCitation":"[57]","previouslyFormattedCitation":"^[56]"},"properties":{"noteIndex":0},"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[57]中，向量被隨機初始化，然而在本文的模型中，使用當前偏好模型的最后一個時刻輸出向量表示。表示用戶最終的周期偏好。1.4模型訓練和預測得到用戶行為的時空軌跡、用戶歷史偏好向量、目前偏好向量和周期偏好向量后，本文以上向量拼接結果將作為用戶最終的偏好表示。注意本文沒有用到短期偏好，本文認為用戶的周期偏好已經(jīng)從短期偏好中提取了有效的信息。進一步，本文引用一層簡單的神經(jīng)網(wǎng)絡整合以上四種偏好之間復雜的影響，同時改變最終偏好的興趣向量維度與興趣點嵌入向量保持一致。受矩陣分解的啟發(fā)，本文使用最終偏好向量與興趣點向量的內積形式表示用戶訪問每個興趣點的分數(shù)。用戶訪問興趣點的分數(shù)被