為了使本領域技術人員更好地理解本說明書實施例中的技術方案，下面將結合本說明書實施例中的圖式，對本說明書實施例中的技術方案進行詳細地描述，顯然，所描述的實施例僅僅是本說明書的一部分實施例，而不是全部的實施例。基於本說明書中的實施例，本領域普通技術人員所獲得的所有其他實施例，都應當屬保護的範圍。
本說明書實施例提供一種針對新場景的預測模型訓練方法，參見圖1所示，該方法可以包括以下步驟：
S101，獲得待遷移模型的集合；
由於新場景與舊場景的差異，在舊場景部署使用的各個模型，其中部分可能不適用於新場景，而部分可能適用於新場景，可以遷移至新場景，待遷移模型即為在舊場景部署使用、且可遷移至新場景的模型。
本說明書不限定獲得待遷移模型的集合的具體方式。
在本說明書實施例中，可以比較舊場景中各個模型輸入的特徵向量、與新場景中訓練樣本可提取的特徵向量，從而確定舊場景中各個模型是否可以遷移至新場景。具體地，首先獲得第一特徵集合，該集合中包括：預先確定的新場景訓練樣本可提取的若干特徵向量；接著針對在舊場景部署使用的任一模型：獲得第二特徵集合，該集合中包括：該模型所輸入的若干特徵向量；在該模型符合預設遷移規則的情況下，將該模型確定為待遷移模型；所述預設遷移規則包括：第一特徵集合與第二特徵集合的交集中包括的特徵向量滿足預設遷移條件。
上述預設遷移條件具體可以透過多種形式，從多種角度比較第一特徵集合與第二特徵集合。
例如，預設遷移條件可以為：交集中包括的特徵向量的數量不小於預設臨界值，即透過比較第一特徵集合與第二特徵集合的交集中特徵向量的數量，確定該模型是否可以遷移至新場景。如果交集中特徵向量數量較少，則該模型在新場景中表現較差的機率較高，因而可以認為該模型不可以被遷移至新場景；反之，則認為該模型可以被遷移至新場景。
又例如，新場景中的某些特徵向量對於模型訓練較為重要，則可以在衡量舊場景中模型是否適合遷移到新場景時，重點考慮是否包括這些特徵向量，因此，預設遷移條件可以為：根據交集中包括的各特徵向量的預設權重計算的加權分數不小於預設臨界值。對於模型訓練較為重要的特徵向量可以預設較高的權重，且越重要預設權重越高。從而，如果交集中包括的重要特徵向量較高，則最終計算的加權分數也較高，並且可以認為該模型可以遷移至新場景。
預設遷移條件也可以為其他形式，並且各遷移條件可以單獨使用，也可以搭配使用，本領域技術人員可以根據實際需求靈活地設定，本說明書對此不做具體限定。
此外，預設遷移規則中也可以包括其他具體規則。新場景中所訓練的預測模型的類型，可以由研發人員預先根據經驗或演算法確定並指定，那麼，為了在比較特徵向量的基礎上，進一步地衡量舊場景中各模型，可否遷移至新場景，在獲得待遷移模型的集合時，還可以進一步地獲得預先為新場景預測模型指定的至少一個類型，並且，預設遷移規則中還可以進一步包括：預先指定的至少一個類型中包括該模型的類型。
從特徵向量及模型類型兩個維度，確定舊場景中的某個模型，是否可以遷移至新場景，從而使遷移至新場景的模型可以透過進一步的訓練，更好地應用於新場景。當然，預設遷移規則中也可以包括其他維度的規則，本說明書實施例對此不做限定。
當然，也可以由研發人員指定舊場景遷移至新場景的待遷移模型，並且研發人員在指定時，也可以根據經驗或演算法，透過從特徵向量、模型類型等維度，衡量各個模型是否可以遷移至新場景、以及遷移後的表現、等等。
S102，從所述待遷移模型的集合中選擇至少一個模型，以用於對新場景中的無標籤樣本進行預測標註；
S103，獲得新場景中的初始訓練樣本集，所述初始訓練樣本集中包括無標籤樣本；
S104，利用所選擇的模型，為初始訓練樣本集中的無標籤樣本添加預測標籤；
為了便於描述，將S102至S104結合進行說明。
基於有監督學習訓練樣本時，需要訓練樣本為有標籤樣本。訓練樣本通常可以透過多種方式標註標籤。例如，可以人工進行標註，且人工標註通常較為準確，但是用於模型訓練的訓練樣本資料量通常很大，人工標註效率較低；又如，部分場景中可以根據實際情況產生標籤，如信用卡風險控制場景中，當銀行核實某信用卡被盜後，該信用卡及對應交易均可以被標註為黑樣本，但這種場景中短期內可能無法獲得黑樣本標籤。
本說明書實施例中，從待遷移模型的集合中選擇至少一個模型，用於對新場景中的無標籤樣本進行預測標註，從而提高標註效率、縮短標註週期。
各個待遷移模型均為可遷移至新場景的模型，但具體地，由於各模型輸入的特徵向量與模型類型等存在區別，其中部分模型可以直接較好地應用於新場景，而部分模型則需要更新後才可以較好地應用於新場景，因此，可以從待遷移模型的集合中，選擇在新場景中表現較好的部分模型，用於預測標註。
從所述待遷移模型的集合中選擇至少一個模型具體可以透過多種方式實現。
在本說明書實施例中，可以首先獲得第三特徵集合，該集合中包括：預先指定的用於預測新場景中樣本標籤的若干特徵向量；接著獲得各待遷移模型對應的各特徵集合，其中任一特徵集合中包括：對應模型所輸入的若干特徵向量；根據預設選擇規則，從所述待遷移模型的集合中，選擇至少一個模型。
與S101中確定待遷移模型類似地，在選擇用於預測標註的模型時，也可以從交集中特徵向量的數量、重要特徵向量的數量、及模型類型是否相同等維度，衡量是否選擇某個模型用於預測標註，本說明書在此不再贅述。
此外，僅透過數量或加權分數是否大於預設臨界值、模型類型是否相同等硬性條件，可能會出現待遷移模型的集合中，不存在符合預設選擇規則的模型的情況，因此，還可以預設各種優先級排序條件，並根據排序結果，選擇1個或多個模型用於預測標註。
如S101所述，待遷移模型可以由研發人員指定，而本步驟中從待遷移模型的集合中，選擇用於對新場景中的無標籤樣本進行預測標註的模型時，也可以由研發人員根據經驗或演算法進行選擇，在此不再贅述。
新場景中的初始訓練樣本集中，可以包括需要被標註預測標籤的無標籤樣本，也可以包括已標註實際標籤的有標籤樣本(可以為白樣本和/或黑樣本)，所選擇的模型用於對無標籤樣本進行預測標註。
具體可以透過多種方式添加預測標籤。
在本說明書實施例中，可以預先設定不同數值與不同預測標籤的對應關係，如數值大於或小於某預設值時，對應黑樣本標籤，反之對應白樣本標籤等。對於所選擇的任一模型：將初始訓練樣本集中的無標籤樣本輸入該模型，得到輸出的預測值；對於所輸入的任一無標籤樣本：確定各模型輸出的預測值的權重；計算各預測值的加權和，並確定該加權和對應的預測標籤；為該無標籤樣本添加該預測標籤。
例如，如果僅從待遷移模型的集合中，選擇了1個模型進行預測標註，則可以直接根據該模型輸出的預測值(即等於加權和)，得到對應的預測標籤。
又如，如果從待遷移模型的集合中，選擇了多個模型進行預測標註，則可以預先設定各模型輸出值對應的權重，如表現更好的模型對應權重更高或更低，當然，也可以預先設定各模型權重相同，即相當於未對各模型設定權重。
此外，利用所選擇的模型添加的預測標籤，還可以透過人工檢查及修正，以提高預測標籤的準確度。
對於上述情況，本領域可以根據實際情況靈活地設定，本說明書不做限定。
S105，利用已添加預測標籤的初始訓練樣本集，基於有監督學習演算法，對待遷移模型進行更新，得到可適用於新場景的模型。
對待遷移模型進行更新時，可以僅將初始訓練樣本集中，已添加預測標籤的訓練樣本輸入待遷移模型。
如果新場景中所積累的訓練樣本數量較少，還可以獲得舊場景中的訓練樣本集，該訓練樣本集中包括已添加實際標籤的有標籤樣本；將新場景中的初始樣本集與舊場景中的訓練樣本集合併，利用合併後的訓練樣本集，基於有監督學習演算法，對待遷移模型進行更新。
舊場景中已經積累了大量的訓練樣本，且訓練樣本為已添加實際標籤的有標籤樣本，因此可以在新場景中所積累的訓練樣本數量較少的情況下，用於輔助新場景中對待遷移模型進行更新。
當然，舊場景中的訓練樣本，並不一定完全適用於新場景的模型更新，其中可能有部分樣本與新場景中的訓練樣本相似度較高，而其他樣本的相似度較低，因此可以為新場景中的初始樣本集與舊場景中的訓練樣本集合併後的訓練樣本集中的不同訓練樣本，預設不同的權重。
例如，初始樣本集中的各訓練樣本權重最高，舊場景中的訓練樣本集中，與初始樣本集中的各訓練樣本相似度較高的各訓練樣本權重次之，而相似度較低的各訓練樣本權重最低。
此外，隨著時間的推移，新場景中也將積累到已添加實際標籤的有標籤樣本，從而構成最佳化訓練樣本集，可以獲得新場景中的最佳化訓練樣本集，所述最佳化訓練樣本集中包括已添加實際標籤的有標籤樣本；將已添加預測標籤的初始訓練樣本集與已添加實際標籤的最佳化訓練樣本集合併，利用合併後的訓練樣本集，基於有監督學習演算法，對待遷移模型進行更新。
可以理解的是，根據新場景對預測模型的需求，各個待遷移模型可以直接應用於新場景中，並在應用的同時按照本方案進行更新，得到更適用於新場景的模型，也可以經過一段時間的更新後，再應用於新場景中，並且還可以在應用之後繼續進行更新，本說明書不做限定。
下面結合一個更為具體的實例，對本說明書提供的針對新場景的預測模型訓練方法進行說明。
金融風險控制領域，可以將積累的大量交易資料作為樣本資料，透過機器學習訓練風險控制模型，從而可以基於所訓練的風險控制模型，及時、準確地對新的交易進行風險決策等。
但是，在新場景中建立風險控制模型時，積累訓練模型所需的大量樣本資料，往往需要較長的時間。例如，樣本資料量通常與新場景的交易量及累計時間有關，此外，樣本訓練集中需要包括一定量的黑樣本資料，而以信用卡盜卡等情況作為黑樣本資料時，由於銀行對盜卡業務處理時間較長等原因，使得短時間內無法積累到訓練所需的黑樣本資料量。
針對上述問題，可以將舊場景中已有的風險控制模型，遷移到新場景中。
新場景與舊場景可以為不同國家、地區的交易市場，舊場景中部署使用的風險控制模型可以包括：盜卡風險控制模型、盜帳戶風險控制模型、隱案識別模型、等等，並且這些風險控制模型可以是基於多個國家、地區的交易資料訓練得到。
如圖2所示，可以預先在雲端，基於從各個舊場景中集中的資料，訓練可在新、舊場景中部署使用的多個模型。
其中，盜卡風險控制模型與盜帳戶風險控制模型，分別針對盜信用卡與盜支付帳戶的情況進行風險控制，並且可以透過有監督學習訓練。
隱案識別模型透過輸入針對性更強的特徵向量，用於識別銀行尚未定為案件(即，非顯案)、但具備案件特徵的交易。
例如，如果同一台設備(如手機)、或同一網路環境等，同時使用多個信用卡或支付帳戶，則該設備或環境中存在批量盜卡、盜帳戶的風險較高；又例如，對於已被列為黑名單的設備、帳戶、信用卡、網路環境等，與之發生關聯的設備、帳戶、信用卡、網路環境等，存在盜卡、盜帳戶的風險也較高；在例如，進行了異常交易(如交易量、交易時間、交易地點等異常)的設備、帳戶、信用卡、網路環境等，存在盜卡、盜帳戶的風險也較高；隱案識別模型可以基於上述特徵，將對應的交易識別為黑樣本。
並且，隱案識別模型可以透過無監督學習訓練，從而可以應用於尚無實際案件(標籤)的場景。
當新場景中需要部署盜卡風險控制模型、盜帳戶風險控制模型及隱案識別模型時，可以以模型檔案的形式，將上述模型下發到新場景本地。並且，本地可以直接使用所部署的模型，對交易事件進行打分、進行風險決策等。
雲端下發部署的模型，是透過多個國家、地區的訓練樣本訓練得到，其優點是訓練樣本全面，通用性較強，而其缺點是與透過新場景本地的資料訓練的全新模型相比，並非完全適用於新場景，因此，還需要在新場景積累一定訓練樣本後，更新這些模型。
各模型部署到新場景本地並使用後，從訓練樣本積累的角度，可以分為多個階段。
在第一階段中，可以認為新場景中積累時間較短，如在部署後的1週內，積累的訓練樣本較少，且各樣本均無標籤，無法進行模型的更新。因此，在第一階段中，使用雲端訓練並且未更新的各模型，對新場景中的交易進行風險控制決策。
在第二階段中，如在部署後的1週至1個月之間，可以認為新場景中積累了一定量的訓練樣本，構成初始訓練樣本集，如果結合雲端下發的舊場景中的大量訓練資料，可以對各模型進行更新。但是，由於金融機構處理盜卡、盜帳戶的週期較長，此時還未積累到帶有實際標籤的有標籤樣本，因此可以透過隱案識別模型，為初始訓練樣本集添加預測樣本。
此外，可以對新、舊場景中的訓練樣本設置不同的權重，例如，新場景為馬來西亞的市場，舊場景包括泰國、美國、日本等市場，其中，泰國與馬來西亞的消費水平、習慣更接近，交易資料相似度更高，而美國、日本與馬來西亞的交易資料相似度更低。因此，可以為馬來西亞本地積累的訓練樣本設置最高的權重，為來自泰國的訓練樣本設置較高的權重，而為來自美國、日本的訓練樣本設置更低的權重。從而，透過動態加權的方式，可以在新場景中資料較少的情況下，使更新訓練後的各模型更適於新場景。
在第二階段中更新後的各模型，仍可以用於新場景的交易決策。
在第三階段中，如在部署1個月之後，可以認為新場景中已經積累了足夠量的訓練樣本，並且積累到了帶有實際標籤的有標籤樣本，則可以進一步地更新各模型。更新所使用的訓練樣本，可以僅包括新場景中、帶有實際標籤的訓練樣本，也可以包括新場景中、帶有隱案識別模型添加的預測標籤的訓練樣本，還可以包括舊場景中的大量訓練樣本，等等。
除了透過雲端預先訓練的模型及積累的資料，在新場景中部署與更新風險控制模型，新場景中所積累的資料，也可以上傳至雲端，以用於更新已有模型、訓練其他新模型，並部署到其他新場景等。
可見，應用上述方案，可以將舊場景中部署使用的模型遷移至新場景中，並且在新場景中樣本積累時間較短，因而樣本沒有或只有少數實際標籤的情況下，透過待遷移模型進行標籤預測，從而進一步最佳化待遷移模型，使這些模型更適於在新場景中使用，為新場景提供一種更高效且更準確的預測模型訓練方案。
相應於上述方法實施例，本說明書實施例還提供一種針對新場景的預測模型訓練裝置，參見圖3所示，該裝置可以包括：
待遷移模型獲取模組110，用於獲得待遷移模型的集合，所述待遷移模型為：在舊場景部署使用、且可遷移至新場景的模型；
標註模型選取模組120，用於從所述待遷移模型的集合中選擇至少一個模型，以用於對新場景中的無標籤樣本進行預測標註；
樣本集獲取模組130，用於獲得新場景中的初始訓練樣本集，所述初始訓練樣本集中包括無標籤樣本；
樣本標註模組140，用於利用所選擇的模型，為初始訓練樣本集中的無標籤樣本添加預測標籤；
模型更新模組150，用於利用已添加預測標籤的初始訓練樣本集，基於有監督學習演算法，對待遷移模型進行更新，得到可適用於新場景的模型。
在本說明書提供的一種具體實施方式中，所述待遷移模型獲取模組110，可以包括：
待遷移特徵獲取單元111，用於獲得第一特徵集合，該集合中包括：預先確定的新場景訓練樣本可提取的若干特徵向量；針對在舊場景部署使用的任一模型：獲得第二特徵集合，該集合中包括：該模型所輸入的若干特徵向量；
待遷移模型選取單元112，用於在該模型符合預設遷移規則的情況下，將該模型確定為待遷移模型；所述預設遷移規則包括：第一特徵集合與第二特徵集合的交集中包括的特徵向量滿足預設遷移條件。
在本說明書提供的一種具體實施方式中，所述預設遷移條件可以包括：
交集中包括的特徵向量的數量不小於預設臨界值；和/或根據交集中包括的各特徵向量的預設權重計算的加權分數不小於預設臨界值。
在本說明書提供的一種具體實施方式中，所述待遷移模型獲取模組110，還可以包括：待遷移類型獲取單元113，用於獲得預先為新場景預測模型指定的至少一個類型；
所述預設遷移規則，還可以包括：預先指定的至少一個類型中包括該模型的類型。
在本說明書提供的一種具體實施方式中，所述標註模型選取模組120，可以包括：
標註特徵獲取單元121，用於獲得第三特徵集合，該集合中包括：預先指定的用於預測新場景中樣本標籤的若干特徵向量；獲得各待遷移模型對應的各特徵集合，其中任一特徵集合中包括：對應模型所輸入的若干特徵向量；
標註模型選取單元122，用於根據預設選擇規則，從所述待遷移模型的集合中，選擇至少一個模型。
在本說明書提供的一種具體實施方式中，所述樣本標註模組140，可以包括：
預測值確定單元141，用於對於所選擇的任一模型：將初始訓練樣本集中的無標籤樣本輸入該模型，得到輸出的預測值；
預測標籤確定單元142，用於對於所輸入的任一無標籤樣本：確定各模型輸出的預測值的權重；計算各預測值的加權和，並確定該加權和對應的預測標籤；為該無標籤樣本添加該預測標籤。
在本說明書提供的一種具體實施方式中，所述樣本集獲取模組130，還可以用於：獲得新場景中的最佳化訓練樣本集，所述最佳化訓練樣本集中包括已添加實際標籤的有標籤樣本；
所述模型更新模組150，具體可以用於：將已添加預測標籤的初始訓練樣本集與已添加實際標籤的最佳化訓練樣本集合併，利用合併後的訓練樣本集，基於有監督學習演算法，對待遷移模型進行更新。
在本說明書提供的一種具體實施方式中，所述樣本集獲取模組130，還可以用於：獲得舊場景中的訓練樣本集，該訓練樣本集中包括已添加實際標籤的有標籤樣本；
所述模型更新模組150，具體可以用於：將新場景中的初始樣本集與舊場景中的訓練樣本集合併，利用合併後的訓練樣本集，基於有監督學習演算法，對待遷移模型進行更新。
上述裝置中各個模組的功能和作用的實現過程具體詳見上述方法中對應步驟的實現過程，在此不再贅述。
本說明書實施例還提供一種電腦設備，其至少包括儲存器、處理器及儲存在儲存器上並可在處理器上運行的電腦程式，其中，處理器執行所述程式時實現前述的針對新場景的預測模型訓練方法。該方法至少包括：
獲得待遷移模型的集合，所述待遷移模型為：在舊場景部署使用、且可遷移至新場景的模型；
從所述待遷移模型的集合中選擇至少一個模型，以用於對新場景中的無標籤樣本進行預測標註；
獲得新場景中的初始訓練樣本集，所述初始訓練樣本集中包括無標籤樣本；
利用所選擇的模型，為初始訓練樣本集中的無標籤樣本添加預測標籤；
利用已添加預測標籤的初始訓練樣本集，基於有監督學習演算法，對待遷移模型進行更新，得到可適用於新場景的模型。
圖4顯示了本說明書實施例所提供的一種更為具體的計算設備硬體結構示意圖，該設備可以包括：處理器1010、儲存器1020、輸入/輸出介面1030、通訊介面1040和匯流排1050。其中處理器1010、儲存器1020、輸入/輸出介面1030和通訊介面1040透過匯流排1050實現彼此之間在設備內部的通訊連接。
處理器1010可以採用通用的CPU(Central Processing Unit，中央處理器)、微處理器、特殊應用積體電路(Application Specific Integrated Circuit，ASIC)、或者一個或多個積體電路等方式實現，用於執行相關程式，以實現本說明書實施例所提供的技術方案。
儲存器1020可以採用ROM(Read Only Memory，唯讀記憶體)、RAM(Random Access Memory，隨機存取記憶體)、靜態儲存設備，動態儲存設備等形式實現。儲存器1020可以儲存作業系統和其他應用程式，在透過軟體或者韌體來實現本說明書實施例所提供的技術方案時，相關的程式代碼保存在儲存器1020中，並由處理器1010來呼叫執行。
輸入/輸出介面1030用於連接輸入/輸出模組，以實現資訊輸入及輸出。輸入輸出/模組可以作為組件配置在設備中(圖中未顯示)，也可以外接於設備以提供相應功能。其中輸入設備可以包括鍵盤、滑鼠、觸控螢幕、麥克風、各類感測器等，輸出設備可以包括顯示器、揚聲器、振動器、指示燈等。
通訊介面1040用於連接通訊模組(圖中未顯示)，以實現本設備與其他設備的通訊互動。其中通訊模組可以透過有線方式(例如USB、網線等)實現通訊，也可以透過無線方式(例如移動網路、WIFI、藍牙等)實現通訊。
匯流排1050包括一通路，在設備的各個組件(例如處理器1010、儲存器1020、輸入/輸出介面1030和通訊介面1040)之間傳輸資訊。
需要說明的是，儘管上述設備僅顯示了處理器1010、儲存器1020、輸入/輸出介面1030、通訊介面1040以及匯流排1050，但是在具體實施過程中，該設備還可以包括實現正常運行所必需的其他組件。此外，本領域的技術人員可以理解的是，上述設備中也可以僅包含實現本說明書實施例方案所必需的組件，而不必包含圖中所示的全部組件。
本說明書實施例還提供一種電腦可讀儲存媒體，其上儲存有電腦程式，該程式被處理器執行時實現前述的針對新場景的預測模型訓練方法。該方法至少包括：
獲得待遷移模型的集合，所述待遷移模型為：在舊場景部署使用、且可遷移至新場景的模型；
從所述待遷移模型的集合中選擇至少一個模型，以用於對新場景中的無標籤樣本進行預測標註；
獲得新場景中的初始訓練樣本集，所述初始訓練樣本集中包括無標籤樣本；
利用所選擇的模型，為初始訓練樣本集中的無標籤樣本添加預測標籤；
利用已添加預測標籤的初始訓練樣本集，基於有監督學習演算法，對待遷移模型進行更新，得到可適用於新場景的模型。
電腦可讀媒體包括永久性和非永久性、可移動和非可移動媒體可以由任何方法或技術來實現資訊儲存。資訊可以是電腦可讀指令、資料結構、程式的模組或其他資料。電腦的儲存媒體的例子包括，但不限於相變記憶體(PRAM)、靜態隨機存取記憶體(SRAM)、動態隨機存取記憶體(DRAM)、其他類型的隨機存取記憶體(RAM)、唯讀記憶體(ROM)、電可抹除可程式化唯讀記憶體(EEPROM)、快閃記憶體或其他記憶體技術、唯讀光碟唯讀記憶體(CD-ROM)、數位多功能光碟(DVD)或其他光學儲存、磁盒式磁帶，磁帶磁磁碟儲存或其他磁性儲存設備或任何其他非傳輸媒體，可用於儲存可以被計算設備存取的資訊。按照本文中的界定，電腦可讀媒體不包括暫存電腦可讀媒體(transitory media)，如調變的資料訊號和載波。
透過以上的實施方式的描述可知，本領域的技術人員可以清楚地瞭解到本說明書實施例可借助軟體加必需的通用硬體平臺的方式來實現。基於這樣的理解，本說明書實施例的技術方案本質上或者說對現有技術做出貢獻的部分可以以軟體產品的形式體現出來，該電腦軟體產品可以儲存在儲存媒體中，如ROM/RAM、磁碟、光碟等，包括若干指令用以使得一台電腦設備(可以是個人電腦，伺服器，或者網路設備等)執行本說明書實施例各個實施例或者實施例的某些部分所述的方法。
上述實施例闡明的系統、裝置、模組或單元，具體可以由電腦晶片或實體實現，或者由具有某種功能的產品來實現。一種典型的實現設備為電腦，電腦的具體形式可以是個人電腦、膝上型電腦、蜂巢式電話、相機電話、智慧型手機、個人數位助理、媒體播放器、導航設備、電子郵件收發設備、遊戲控制台、平板電腦、可穿戴設備或者這些設備中的任意幾種設備的組合。
本說明書中的各個實施例均採用遞進的方式描述，各個實施例之間相同相似的部分互相參見即可，每個實施例重點說明的都是與其他實施例的不同之處。尤其，對於裝置實施例而言，由於其基本相似於方法實施例，所以描述得比較簡單，相關之處參見方法實施例的部分說明即可。以上所描述的裝置實施例僅僅是示意性的，其中所述作為分離部件說明的模組可以是或者也可以不是實體上分開的，在實施本說明書實施例方案時可以把各模組的功能在同一個或多個軟體和/或硬體中實現。也可以根據實際的需要選擇其中的部分或者全部模組來實現本實施例方案的目的。本領域普通技術人員在不付出創造性勞動的情況下，即可以理解並實施。
以上所述僅是本說明書實施例的具體實施方式，應當指出，對於本技術領域的普通技術人員來說，在不脫離本說明書實施例原理的前提下，還可以做出若干改進和潤飾，這些改進和潤飾也應視為本說明書實施例的保護範圍。In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present specification, the technical solutions in the embodiments of the present specification will be described in detail in conjunction with the drawings in the embodiments of the present specification. Obviously, the described implementation Examples are only a part of the embodiments of this specification, but not all the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art should fall within the scope of protection.
The embodiments of the present specification provide a prediction model training method for new scenes. Referring to FIG. 1, the method may include the following steps:
S101. Obtain a set of models to be migrated;
Due to the difference between the new scene and the old scene, each model used in the old scene deployment may not be suitable for the new scene, and some may be suitable for the new scene, and can be migrated to the new scene. The model to be migrated is deployed in the old scene A model that is used and can be migrated to a new scene.
This specification does not limit the specific way of obtaining the collection of models to be migrated.
In the embodiment of the present specification, the feature vectors input by each model in the old scene and the feature vectors extractable from the training samples in the new scene can be compared to determine whether each model in the old scene can be migrated to the new scene. Specifically, the first feature set is first obtained, and the set includes: a plurality of feature vectors that can be extracted from the training samples of the new scene in advance; then for any model deployed in the old scene: a second feature set is obtained, and the set It includes: several feature vectors input by the model; if the model meets the preset migration rule, the model is determined as the model to be migrated; the preset migration rule includes: the first feature set and the second feature set The feature vectors included in the intersection meet the preset migration conditions.
The above-mentioned preset migration conditions can be specifically compared in various ways to compare the first feature set and the second feature set from multiple angles.
For example, the preset migration condition may be: the number of feature vectors included in the intersection is not less than a preset critical value, that is, by comparing the number of feature vectors in the intersection between the first feature set and the second feature set, it is determined whether the model can be migrated To the new scene. If the number of feature vectors in the intersection is small, the probability of the model performing poorly in the new scene is high, so it can be considered that the model cannot be migrated to the new scene; otherwise, it is considered that the model can be migrated to the new scene.
For another example, some feature vectors in the new scene are more important for model training, you can focus on whether to include these feature vectors when measuring whether the model in the old scene is suitable for migration to the new scene, so the preset migration conditions can be: The weighted score calculated according to the preset weight of each feature vector included in the intersection is not less than the preset critical value. For feature vectors that are more important for model training, higher weights can be preset, and the more important the preset weight, the higher the weight. Therefore, if the important feature vectors included in the intersection are high, the final calculated weighted score is also high, and it can be considered that the model can be migrated to a new scene.
The preset migration conditions may also be in other forms, and each migration condition may be used alone or in combination. Those skilled in the art can flexibly set it according to actual needs, which is not specifically limited in this specification.
In addition, the preset migration rules may also include other specific rules. The type of prediction model trained in the new scene can be determined and specified by the R&D personnel based on experience or algorithms in advance. Then, in order to further measure each model in the old scene on the basis of comparing feature vectors, can it be migrated to the new scene When obtaining the set of models to be migrated, at least one type specified in advance for the new scene prediction model may be further obtained, and the preset migration rule may further include: at least one type specified in advance includes the model’s Types of.
From the two dimensions of feature vector and model type, determine whether a model in the old scene can be migrated to the new scene, so that the model migrated to the new scene can be better applied to the new scene through further training. Of course, the preset migration rules may also include rules of other dimensions, which is not limited in the embodiments of this specification.
Of course, the R&D personnel can also specify the model to be migrated from the old scene to the new scene, and when the R&D personnel specify it, they can also measure whether each model can be migrated from the dimensions of feature vectors, model types, etc. based on experience or algorithms. To new scenes, performance after migration, etc.
S102. Select at least one model from the set of models to be migrated for predicting and labeling unlabeled samples in a new scene;
S103: Obtain an initial training sample set in a new scene, where the initial training sample set includes unlabeled samples;
S104. Use the selected model to add prediction labels to the unlabeled samples in the initial training sample set;
For ease of description, S102 to S104 will be described in combination.
When training samples are based on supervised learning, the training samples need to be labeled samples. Training samples can usually be labeled in many ways. For example, manual labeling can be performed manually, and manual labeling is usually more accurate, but the amount of training sample data used for model training is usually large, and the efficiency of manual labeling is low; for another example, in some scenarios, labels can be generated according to the actual situation, such as credit card risk In the control scenario, after the bank verifies that a credit card has been stolen, both the credit card and the corresponding transaction can be marked as black samples, but in this scenario, the black sample label may not be obtained in the short term.
In the embodiment of the present specification, at least one model is selected from the set of models to be migrated, which is used to predict and label the unlabeled samples in the new scene, thereby improving the labeling efficiency and shortening the labeling period.
Each model to be migrated is a model that can be migrated to a new scene, but specifically, due to the difference between the feature vectors input by each model and the model type, some models can be directly applied to the new scene, while some models require It can be applied to new scenes only after updating. Therefore, from the set of models to be migrated, some models that perform well in the new scenes can be selected for predictive labeling.
Selecting at least one model from the set of models to be migrated may be implemented in various ways.
In the embodiment of the present specification, a third feature set may be obtained first, and the set includes: pre-specified several feature vectors for predicting sample labels in a new scene; then, each feature set corresponding to each model to be migrated is obtained, any of which A feature set includes: several feature vectors input by the corresponding model; according to a preset selection rule, at least one model is selected from the set of models to be migrated.
Similar to determining the model to be migrated in S101, when selecting a model for predictive labeling, you can also measure whether to select a model from the dimensions of the number of feature vectors in the intersection, the number of important feature vectors, and whether the model type is the same Used for predictive labeling, and will not be repeated here.
In addition, only through the hard conditions such as whether the number or weighted score is greater than the preset threshold and whether the model type is the same, there may be a case where there is no model that meets the preset selection rules in the set of models to be migrated. Set various priority ranking conditions, and select one or more models for predicting labels according to the ranking results.
As described in S101, the model to be migrated can be designated by the R&D personnel, and in this step, when a model for predicting and labeling unlabeled samples in the new scene is selected from the set of models to be migrated, the R&D personnel can also use Experience or algorithm to choose, not repeat them here.
The initial training sample set in the new scenario may include unlabeled samples that need to be labeled with predicted labels, or labeled samples (which may be white samples and/or black samples) that have been labeled with actual labels. The selected model is used for Predict labeling of unlabeled samples.
Specifically, prediction tags can be added in various ways.
In the embodiment of the present specification, the corresponding relationship between different numerical values and different predicted labels may be preset. For example, when the numerical value is greater than or less than a predetermined value, it corresponds to a black sample label, and vice versa to a white sample label. For any selected model: input the unlabeled samples in the initial training sample set into the model to obtain the predicted value of the output; for any input unlabeled samples: determine the weight of the predicted value output by each model; calculate each prediction Weighted sum of values and determine the predicted label corresponding to the weighted sum; add the predicted label to the unlabeled sample.
For example, if only one model is selected from the set of models to be migrated for predictive labeling, the corresponding predicted label can be obtained directly according to the predicted value (that is, equal to the weighted sum) output by the model.
For another example, if multiple models are selected from the set of models to be migrated for predictive labeling, the weights corresponding to the output values of each model can be set in advance. For example, the better performing models have higher or lower corresponding weights. Of course, The weights of the models can be preset to be the same, which means that no weights are set for the models.
In addition, the prediction labels added by the selected model can also be manually checked and corrected to improve the accuracy of the prediction labels.
For the above situation, the field can be flexibly set according to the actual situation, and this specification does not limit it.
S105. Using the initial training sample set to which the prediction label has been added, based on the supervised learning algorithm, the model to be migrated is updated to obtain a model applicable to the new scene.
When updating the model to be migrated, only the initial training sample set may be input, and the training samples to which the prediction labels have been added are input to the model to be migrated.
If the number of training samples accumulated in the new scene is small, the training sample set in the old scene can also be obtained. The training sample set includes labeled samples with actual labels added; the initial sample set in the new scene is compared with the old scene. The training sample set is merged, and the merged training sample set is used to update the migration model based on the supervised learning algorithm.
A large number of training samples have been accumulated in the old scene, and the training samples are labeled samples with actual labels added, so they can be used to assist the migration of the new scene when the number of training samples accumulated in the new scene is small The model is updated.
Of course, the training samples in the old scene may not be fully applicable to the model update of the new scene. Some of the samples may have a higher similarity with the training samples in the new scene, while other samples have a lower similarity, so it can be Different weights are preset for different training samples in the initial sample set in the new scene and the training sample set in the old scene after the training sample set is merged.
For example, the weight of each training sample in the initial sample set is the highest, the weight of each training sample in the training sample set in the old scene with the highest similarity to the training sample set in the initial sample set is second, and the weight of each training sample in the lower similarity is lowest.
In addition, over time, the new scene will also accumulate labeled samples with actual labels added, thereby forming an optimized training sample set, which can obtain the optimized training sample set in the new scene. The training sample set includes labeled samples with actual labels added; the initial training sample set with predicted labels and the optimized training sample set with actual labels are combined, and the combined training sample set is used based on supervised learning Algorithms to update the migration model.
It can be understood that, according to the demand of the new scene for the prediction model, each model to be migrated can be directly applied to the new scene, and be updated according to this scheme at the same time as the application, to obtain a model more suitable for the new scene, or after a period of time After the time is updated, it can be applied to new scenes, and it can continue to be updated after application, which is not limited in this manual.
In the following, a more specific example will be used to explain the prediction model training method for new scenarios provided in this specification.
In the field of financial risk control, a large amount of accumulated transaction data can be used as sample data to train a risk control model through machine learning, so that based on the trained risk control model, risk decisions can be made on new transactions in a timely and accurate manner.
However, when building a risk control model in a new scenario, it often takes a long time to accumulate a large amount of sample data required for training the model. For example, the amount of sample data is usually related to the transaction volume and cumulative time of the new scenario. In addition, the sample training set needs to include a certain amount of black sample data. When credit card stolen cards are used as black sample data, due to the bank’s card stolen business Long processing time and other reasons make it impossible to accumulate the amount of black sample data required for training in a short time.
In response to the above problems, the existing risk control model in the old scenario can be migrated to the new scenario.
The new scenario and the old scenario can be transaction markets in different countries and regions. The risk control models deployed in the old scenario can include: card piracy risk control model, stolen account risk control model, hidden case identification model, etc., and these risks The control model can be trained based on transaction data from multiple countries and regions.
As shown in Figure 2, multiple models that can be deployed and used in new and old scenarios can be trained on the cloud in advance, based on data gathered from various old scenarios.
Among them, the stolen card risk control model and the stolen account risk control model respectively carry out risk control for the situation of stolen credit card and stolen payment accounts, and can be trained through supervised learning.
The hidden case recognition model is used to identify transactions that the bank has not yet determined as a case (ie, non-obvious case), but has the characteristics of the case by inputting a more targeted feature vector.
For example, if multiple credit cards or payment accounts are used at the same time on the same device (such as a mobile phone) or the same network environment, there is a high risk of bulk card theft or account theft in the device or environment; For devices, accounts, credit cards, and network environments that are blacklisted, and associated devices, accounts, credit cards, and network environments, there is a higher risk of card theft and account theft; in, for example, Abnormal transactions (such as abnormal transaction volume, transaction time, transaction location, etc.), equipment, accounts, credit cards, network environment, etc., are also at higher risk of card theft and account theft; the hidden case identification model can be based on the above characteristics and will correspond Of transactions are identified as black samples.
In addition, the hidden case recognition model can be trained through unsupervised learning, so that it can be applied to scenarios where there is no actual case (tag).
When it is necessary to deploy a card theft risk control model, a stolen account risk control model, and a hidden case identification model in a new scenario, the above model can be delivered to the new scenario in the form of a model file. In addition, the deployed model can be used locally to score transaction events and make risk decisions.
The model deployed in the cloud is obtained through training samples from multiple countries and regions. Its advantages are comprehensive training samples and strong versatility. Its disadvantage is that it is compared with a brand new model trained with local scene data in new scenarios. It is not completely applicable to new scenes, therefore, it is necessary to update these models after accumulating certain training samples in the new scene.
After each model is deployed locally in a new scene and used, from the perspective of training sample accumulation, it can be divided into multiple stages.
In the first stage, it can be considered that the accumulation time in the new scene is relatively short. For example, within 1 week after deployment, the accumulated training samples are few, and each sample has no label, and the model cannot be updated. Therefore, in the first stage, each model trained in the cloud and not updated is used to make risk control decisions on transactions in the new scenario.
In the second phase, such as between 1 week and 1 month after deployment, it can be considered that a certain amount of training samples have been accumulated in the new scene to form the initial training sample set. If combined with a large amount of training in the old scene delivered by the cloud Information can be updated on each model. However, because financial institutions have a relatively long period of time to deal with card theft and account theft, at this time, there are no labeled samples with actual labels, so you can add prediction samples to the initial training sample set through the hidden case recognition model.
In addition, different weights can be set for the training samples in the new and old scenes. For example, the new scene is the Malaysian market, and the old scenes include Thailand, the United States, and Japan. Among them, Thailand and Malaysia have closer consumption levels and habits. The similarity of transaction data is higher, while the similarity of transaction data of the United States, Japan and Malaysia is lower. Therefore, you can set the highest weight for the training samples accumulated locally in Malaysia, set higher weights for the training samples from Thailand, and set lower weights for the training samples from the United States and Japan. Therefore, the dynamic weighting method can make the updated models more suitable for the new scene when there is less data in the new scene.
The updated models in the second stage can still be used for trading decisions in new scenarios.
In the third stage, if after 1 month of deployment, it can be considered that a sufficient number of training samples have been accumulated in the new scene and labeled samples with actual labels have been accumulated, then each model can be further updated. The training samples used for the update can include only the training samples with the actual labels in the new scene, or the training samples with the prediction labels added by the hidden case recognition model in the new scene, or the old scenes. A large number of training samples, etc.
In addition to pre-training models and accumulated data in the cloud, deploy and update risk control models in new scenarios. The accumulated data in new scenarios can also be uploaded to the cloud to update existing models and train other new models. And deploy to other new scenarios.
It can be seen that using the above scheme, the models deployed in the old scene can be migrated to the new scene, and the sample accumulation time in the new scene is short, so the sample has no or only a few actual tags, and the tags are transferred through the model to be migrated. Prediction, so as to further optimize the models to be migrated, make these models more suitable for use in new scenes, and provide a more efficient and accurate prediction model training scheme for new scenes.
Corresponding to the above method embodiments, the embodiments of the present specification also provide a prediction model training device for new scenes. Referring to FIG. 3, the device may include:
The model to be migrated acquisition module 110 is used to obtain a collection of models to be migrated. The model to be migrated is: a model deployed in an old scene and can be migrated to a new scene;
The labeling model selection module 120 is used to select at least one model from the set of models to be migrated for predicting and labeling unlabeled samples in a new scene;
The sample set acquisition module 130 is used to obtain an initial training sample set in a new scene, where the initial training sample set includes unlabeled samples;
The sample labeling module 140 is used to add a prediction label to the unlabeled samples in the initial training sample set using the selected model;
The model updating module 150 is used to update the model to be migrated based on the supervised learning algorithm based on the initial training sample set to which the prediction label has been added, to obtain a model applicable to the new scene.
In a specific embodiment provided in this specification, the model acquisition module 110 to be migrated may include:
The feature acquisition unit 111 to be migrated is used to obtain a first feature set, which includes: a number of feature vectors that can be extracted by a predetermined new scene training sample; for any model used in the old scene deployment: obtain a second feature set , The set includes: several feature vectors input by the model;
The model to be migrated selection unit 112 is used to determine the model as the model to be migrated when the model meets the preset migration rule; the preset migration rule includes: the intersection of the first feature set and the second feature set The included feature vector satisfies the preset migration condition.
In a specific implementation provided by this specification, the preset migration condition may include:
The number of feature vectors included in the intersection is not less than a preset critical value; and/or the weighted score calculated according to the preset weight of each feature vector included in the intersection is not less than the preset critical value.
In a specific embodiment provided in this specification, the model-to-be-migrated acquisition module 110 may further include: a type-to-be-migrated acquisition unit 113 for acquiring at least one type specified in advance for a new scene prediction model;
The preset migration rule may further include: at least one type specified in advance includes the type of the model.
In a specific embodiment provided in this specification, the annotation model selection module 120 may include:
Annotated feature acquisition unit 121 is used to obtain a third feature set, which includes: pre-designated several feature vectors for predicting sample labels in a new scene; and each feature set corresponding to each model to be migrated, any of which is a feature The set includes: several feature vectors input by the corresponding model;
The labeling model selection unit 122 is configured to select at least one model from the set of models to be migrated according to a preset selection rule.
In a specific embodiment provided in this specification, the sample labeling module 140 may include:
The predictive value determining unit 141 is used for any selected model: input the unlabeled samples in the initial training sample set to the model to obtain the output predictive value;
The predicted label determination unit 142 is used to determine the weight of the predicted value output by each model for any input label-free samples; calculate the weighted sum of each predicted value, and determine the predicted label corresponding to the weighted sum; for the unlabeled The sample adds the prediction label.
In a specific embodiment provided in this specification, the sample set acquisition module 130 may also be used to: obtain an optimized training sample set in a new scene, where the optimized training sample set includes the added actual label Of labeled samples;
The model updating module 150 may be specifically used to: combine the initial training sample set with added prediction labels and the optimized training sample set with actual labels, and use the combined training sample set based on supervised learning calculus Method to update the migration model.
In a specific embodiment provided in this specification, the sample set acquisition module 130 may also be used to: obtain a training sample set in an old scene, where the training sample set includes labeled samples to which actual labels have been added;
The model updating module 150 may be specifically used to: combine the initial sample set in the new scene with the training sample set in the old scene, and use the combined training sample set based on the supervised learning algorithm to treat the migration model. Update.
For the implementation process of the functions and functions of each module in the above device, please refer to the implementation process of the corresponding steps in the above method for details, which will not be repeated here.
Embodiments of the present specification also provide a computer device, which includes at least a storage, a processor, and a computer program stored on the storage and executable on the processor, wherein the processor implements the program to implement the aforementioned new scenario Predictive model training method. The method includes at least:
Obtain a set of models to be migrated. The models to be migrated are: models deployed in the old scene and can be migrated to the new scene;
Selecting at least one model from the set of models to be migrated for predicting and labeling unlabeled samples in a new scene;
Obtaining an initial training sample set in a new scene, where the initial training sample set includes unlabeled samples;
Use the selected model to add prediction labels to the unlabeled samples in the initial training sample set;
Using the initial training sample set with added prediction labels, based on the supervised learning algorithm, the model to be migrated is updated to obtain a model that is suitable for new scenarios.
4 shows a schematic diagram of a more specific hardware structure of a computing device provided by an embodiment of the present specification. The device may include: a processor 1010, a storage 1020, an input/output interface 1030, a communication interface 1040, and a bus 1050. Among them, the processor 1010, the storage 1020, the input/output interface 1030 and the communication interface 1040 realize the communication connection among the devices through the bus 1050.
The processor 1010 may be implemented by a general-purpose CPU (Central Processing Unit, central processing unit), a microprocessor, an application specific integrated circuit (ASIC), or one or more integrated circuits, etc., for Execute relevant programs to realize the technical solutions provided by the embodiments of this specification.
The storage 1020 may be implemented in the form of ROM (Read Only Memory, read only memory), RAM (Random Access Memory), static storage device, dynamic storage device, or the like. The storage 1020 can store the operating system and other application programs. When the technical solutions provided by the embodiments of the present specification are implemented through software or firmware, related program codes are stored in the storage 1020 and are called and executed by the processor 1010. .
The input/output interface 1030 is used to connect input/output modules to realize information input and output. The input/output/module can be configured as a component in the device (not shown in the figure), or can be externally connected to the device to provide corresponding functions. The input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, and the like.
The communication interface 1040 is used to connect a communication module (not shown in the figure) to realize communication interaction between the device and other devices. Among them, the communication module can realize communication through wired methods (such as USB, network cable, etc.), and can also realize communication through wireless methods (such as mobile network, WIFI, Bluetooth, etc.).
The bus 1050 includes a path for transmitting information between various components of the device (such as the processor 1010, the storage 1020, the input/output interface 1030, and the communication interface 1040).
It should be noted that although the above device only shows the processor 1010, the storage 1020, the input/output interface 1030, the communication interface 1040, and the bus bar 1050, in the specific implementation process, the device may also include necessary for normal operation Other components. In addition, those skilled in the art may understand that the above-mentioned device may also include only the components necessary to implement the solutions of the embodiments of the present specification, rather than including all the components shown in the figures.
The embodiment of the present specification also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the aforementioned prediction model training method for a new scene is realized. The method includes at least:
Obtain a set of models to be migrated, the models to be migrated are: models deployed in the old scene and can be migrated to the new scene;
Selecting at least one model from the set of models to be migrated for predicting and labeling unlabeled samples in a new scene;
Obtaining an initial training sample set in a new scene, where the initial training sample set includes unlabeled samples;
Use the selected model to add prediction labels to the unlabeled samples in the initial training sample set;
Using the initial training sample set with added prediction labels, based on the supervised learning algorithm, the model to be migrated is updated to obtain a model that is suitable for new scenarios.
Computer-readable media, including permanent and non-permanent, removable and non-removable media, can be stored by any method or technology. The information can be computer readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM) , Read-only memory (ROM), electrically erasable and programmable read-only memory (EEPROM), flash memory or other memory technologies, read-only disc read-only memory (CD-ROM), digital multifunction Optical discs (DVD) or other optical storage, magnetic cassette tapes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media can be used to store information that can be accessed by computing devices. According to the definition in this article, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.
It can be known from the description of the above implementation manners that those skilled in the art can clearly understand that the embodiments of this specification can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solutions of the embodiments of this specification can be embodied in the form of software products in essence or part that contributes to the existing technology, and the computer software products can be stored in storage media, such as ROM/RAM, magnetic Discs, optical discs, etc., include several instructions to enable a computer device (which may be a personal computer, server, or network device, etc.) to perform the methods described in the various embodiments of the embodiments of this specification or some parts of the embodiments.
The system, device, module or unit explained in the above embodiments may be implemented by a computer chip or entity, or by a product with a certain function. A typical implementation device is a computer, and the specific form of the computer may be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email sending and receiving device, a game Consoles, tablets, wearable devices, or any combination of these devices.
The embodiments in this specification are described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant part can be referred to the description of the method embodiment. The device embodiments described above are only schematics, wherein the modules described as separate components may or may not be physically separated, and the functions of each module can be Implemented in one or more software and/or hardware. Part or all of the modules may also be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement without paying creative labor.
The above is only a specific implementation manner of the embodiments of this specification. It should be pointed out that for those of ordinary skill in the art, without departing from the principles of the embodiments of this specification, several improvements and retouches can be made. These Improvements and retouching should also be regarded as the scope of protection of the embodiments of this specification.