下面結合附圖,對本發明的實施例進行描述。 本發明實施例提供的神經網路模型訓練方法適用於對深度神經網路(Deep Neural Network, DNN)或者人工神經網路(Artificial Neural Network,ANN)等神經網路模型進行訓練的情況。訓練好的神經網路模型可以用來進行模式識別以及分類的情況,如,可以用來對交易行為進行風險識別。 圖1為本發明一種實施例提供的神經網路模型訓練方法流程圖。所述方法的執行主體可以為具有處理能力的設備:伺服器或者系統或者裝置,如圖1所示,所述方法具體包括: 步驟110,將預先收集的多個樣本資料登錄到梯度提升決策樹(Gradient Boosting Decision Tree,GBDT)中,以確定每個樣本資料在GBDT中對應的路徑資訊。 在執行步驟110之前,可以先訓練好GBDT模型。具體的訓練過程後續進行說明。 步驟110中,以訓練的神經網路模型用於交易行為風險識別的情況為例來說,上述樣本資料可以是指用戶的交易行為資料。具體地,可以是從支付寶系統的後臺資料庫中搜集樣本資料。此處,樣本資料可以歸屬於如下五個類別的用戶資料:1)用戶的歷史行為資訊。如,a,若干天(如,180天)內用戶來電次數;b,最後一次登錄城市;c,最後一次登錄距今時間;d,若干天(如,90天)內登錄次數等。2)用戶的交易資訊。如,a,若干天(如,90天)平均支付金額;b,若干天(如,180天)內支付天數;c,若干天(如,180天)內支付金額;d,最後一次支付距今時間等。3)用戶的基本資訊。如,a,用戶是否單身;b,用戶是否裝修;c,用戶是否已婚;d,用戶年齡;e,用戶註冊時長;f,用戶教育水準等。4)用戶的遠端程序呼叫(Remote Procedure Call,RPC)行為資訊。此處的RPC行為資訊是指用戶在使用用戶端的時候,用戶端與伺服器之間的RPC調用。在一種實現方式中,可以搜集每個用戶在最近一個給定時間視窗的這些操作。如,可以搜集用戶近2天訪問的RPC介面的次數變數。5)用戶的統一資源定位器(Uniform Resourc e Locator,URL)位址資訊。 對上述收集的多個樣本資料,如果某樣本資料與目前用戶不相關或者該樣本資料能給用戶帶來負面影響的,則將該樣本資料分類為正樣本資料。如,某一交易行為由非用戶本人操作的或者對用戶的帳戶帶來一定的損失且報案的,則將該交易行為資料標記為正樣本資料。否則,如果某樣本資料為用戶本人正常的交易行為資料,則將該樣本資料標記為負樣本資料。 需要說明的是,通常負樣本資料比較容易搜集。如,可以很容易從支付寶系統的後臺資料庫中搜集到正常支付行為的資料。所以,樣本資料集合中負樣本資料會占絕大多數的比重,如,大於99.999%。然而,當負樣本資料的比重比較高時,訓練的神經網路模型往往會有偏差,如,只能識別安全的交易行為,而不能識別有風險的交易行為,這影響了交易行為風險識別的準確性。 為了能提升交易行為風險識別的準確性,可以對樣本資料進行預處理。在一種實現方式中,可以對正樣本資料進行升取樣處理;和/或,對負樣本資料進行降取樣處理。其中,對正樣本資料進行升取樣處理可以包括:透過複製等方式增加正樣本資料的數量。對負樣本資料進行降取樣處理可以包括:透過刪除等方式減小負樣本資料的數量。在一個例子中,可以將正樣本資料與負樣本資料的比例調整為1:300。 還需要說明的是,對上述預處理後的樣本資料,還可以為正、負樣本資料添加對應的樣本標籤。具體地,為正樣本資料添加正樣本標籤,為負樣本資料添加負樣本標籤。 步驟110中,將預先收集的多個樣本資料登錄到GBDT中具體可以包括:針對每個樣本資料,可以先根據該樣本資料,確定多個特徵對應的特徵值。之後,將特徵的特徵值輸入到GBDT的決策樹中。 此處的特徵可以歸屬於多個類別。在一種實現方式中,上述特徵中的部分特徵可以採用現有交易行為風險識別模型線上沉澱的模型變數,該模型變數歸屬於如下三個類別:1)用戶的歷史行為資訊。2)用戶的交易資訊。3)用戶的基本資訊。 然而,上述模型變數需要根據業務資料來確定,而業務資料通常來自不同業務部門,其採集和整理需要一定的時間,所以僅透過上述模型變數不能得到用戶最新的狀態,從而也不能對用戶最新的交易行為進行風險識別。為解決該問題,本發明中增加了歸屬於用戶的RPC行為資訊的特徵和歸屬於用戶的URL位址資訊的特徵。 綜上,本發明的特徵可以為歸屬於如下五個類別的特徵:1)用戶的歷史行為資訊。2)用戶的交易資訊。3)用戶的基本資訊。4)用戶的RPC行為資訊。5)用戶的URL位址資訊。其中,每個類別如上所述,在此不復贅述。 對上述設定的特徵,在根據具體的樣本資料,確定其對應的特徵值之後,就可以將特徵值輸入到GBDT中。此處的GBDT可以由多棵決策樹組成,每棵決策樹包括多個節點,每個節點與一個特徵相對應。以一棵決策樹為例來說,該決策樹可以如圖2所示,圖2中,節點1、節點2和節點3分別與特徵:“用戶性別是否是男”、“用戶年齡大於20歲”以及“交易金額是否超過1000元”相對應。在將特徵的特徵值輸入決策樹之後,就可以在決策樹中確定出多條路徑資訊。如,假設樣本資料包含用戶性別是男,用戶年齡大於20歲,交易金額超過1000元時,確定出的路徑資訊可以如圖2中的粗線所示。 作為示例性說明,圖2中只是展示了一條路徑資訊,實際上樣本資料登錄GBDT時,可以確定出多條路徑資訊,本發明在此不復贅述。 需要說明的是,本發明中,在將該特徵值輸入到GBDT之前,還可以將該特徵值表示為one-hot形式的特徵向量。在還確定特徵值對應的特徵向量的情況下,上述將特徵值輸入到GBDT中可以替換為:將特徵值對應的特徵向量輸入到決策樹中,以確定相應的路徑資訊。其中,確定特徵值的特徵向量的過程可以舉例如下: 以特徵為“用戶性別”為例來說,如果用戶性別為男,亦即,特徵的特徵值為“男”,則該特徵值對應的特徵向量可以為:[0 1]。如果用戶性別為女,亦即,特徵的特徵值為“女”,則該特徵值對應的特徵向量可以為:[1 0]。 再以特徵為用戶的RPC行為資訊為例來說,其特徵值對應的特徵向量的確定可以透過如下兩種方式來實現:第一種實現方式中,首先設定規則:出現過則標識為1,否則為0。具體地,假設預設的RPC行為資訊為:a,b和c。而某個樣本資料包含用戶兩天內的RPC行為資訊為:a,a和b,亦即,特徵值為:a,a和b。則對應的特徵向量可以為:[1 1 0]。在另一種實現方式中,可以設定規則:統計預設的RPC行為資訊的頻次,然後歸一化。具體地,假設預設的RPC行為資訊為:a,b和c。而某個樣本資料包含用戶兩天內的RPC行為資訊為:a,a,b,b和c,亦即,特徵值為:a,a,b,b和c。則對應的特徵向量可以為:2,2和1。因為需要歸一化,所以最終的特徵向量為:[0.4 0.4 0.2]。 需要說明的是,上述將特徵值表示為特徵向量屬於傳統的習知技術,在此不復贅述。 需要說明的是,為了提升神經網路模型的準確性,本發明中設定了比較多的特徵,從而會確定多個特徵值。對於越來越多的特徵值,其處理往往需要花費很多的時間,受限於同時觀察的特徵值的個數,人很難對多個特徵值之間的關係進行深入的分析,並手工產生新的特徵值。而本發明透過將樣本資料登錄GBDT來得到路徑資訊,該路徑資訊由於包含了多個特徵值。從而可以大大地減小特徵值的數量,由此可以顯著地減少人工的操作。 步驟120,根據每個樣本資料對應的路徑資訊以及樣本標籤,對神經網路模型進行訓練。 此處的神經網路模型可以包括DNN或者ANN等。其中,DNN最近幾年發展迅速,相比傳統使用的淺層模型(如,邏輯迴歸(Logistic Regression,LR), 隨機森林(Random forest,RF)),DNN有著其特有的先進性:模型表達能力強大,適合大資料和分散式訓練。因此,本說明書中,以訓練DNN為例來進行說明。 在本發明中,DNN的訓練過程可以如圖3所示,圖3中,DNN的輸入層用來輸入GBDT中的各條路徑資訊,而輸出層即可輸出第一預測結果。可以理解的是,針對每個樣本資料,亦即,在將該樣本資料對應的路徑資訊輸入到DNN之後,DNN都會輸出相應的第一預測結果。對樣本集合中的多個樣本資料,若第一預測結果與樣本資料的樣本標籤相符合的概率達到預設閾值,此處的預設閾值可以根據經驗值來予以設定,則可以認為已經得到了最佳化的DNN。 可以理解的是,隨著路徑資訊的個數的不同,圖3中DNN的層數是可以改變的。 透過實驗發明,本發明訓練得到的神經網路模型會比其他模型(LR或者RF)的效果都好。同時特徵處理的時間大大地減少了,整體建模流程變快了很多。 以下對如何訓練GBDT模型進行說明: 在根據每個樣本資料,確定多個特徵對應的特徵值之後,可以將多個特徵對應的特徵值輸入GBDT的各個決策樹中。之後將各個決策樹的結果累加起來以確定第二預測結果。可以理解的是,針對每個樣本資料,GBDT模型都會輸出相應的第二預測結果。對樣本集合中的多個樣本資料,若第二預測結果與樣本資料的樣本標籤相符合的概率達到預設閾值,此處的預設閾值可以根據經驗值來予以設定,則可以認為已經得到了最佳化的GBDT模型。而若第二預測結果與樣本資料的樣本標籤相符合的概率未達到預設閾值,則可以透過調整決策樹的數目、決策樹的深度以及正則化項(用來表示特徵)來繼續執行上述輸入和輸出的操作,直至達到預設閾值為止。 綜上,本發明具有如下幾方面的優點: 1)由於本發明的特徵包括了類別為用戶RPC行為資訊的特徵,因此本發明訓練的神經網路模型能夠滿足時效性要求,亦即,能夠識別用戶最新的交易行為。 2)本發明訓練的神經網路模型的準確性比傳統的淺層模型高。 3)透過將樣本資料登錄GBDT,獲得了路徑資訊。而一條路徑資訊由多個特徵值組合而成,亦即,一條路徑資訊包含了樣本資料的多個維度的資訊,由此,可以極大地減小DNN輸入層輸入的資料量,從而可以提高神經網路模型訓練的效率。 需要說明的是,在透過圖1所示的各步驟訓練得到神經網路模型之後,就可以將該神經網路模型部署到線上,並對用戶的交易行為進行風險識別了。 圖4為本發明提供的交易行為風險識別方法的過程示意圖。如圖4所示,該方法可以包括: 步驟410,獲取用戶的交易行為資料。 此處的交易行為資料與上述樣本資料的定義相同,在此不復贅述。 步驟420,將交易行為資料登錄到梯度提升決策樹GBDT中,以確定交易行為資料在GBDT中對應的路徑資訊。 上述GBDT由多棵決策樹組成,每棵決策樹包括多個節點,每個節點與一個特徵相對應。步驟420中將交易行為資料登錄到梯度提升決策樹GBDT中,以確定交易行為資料在GBDT中對應的路徑資訊的步驟具體上可以包括:根據交易行為資料,確定多個特徵對應的特徵值;根據特徵值,在決策樹中確定路徑資訊。其中,確定路徑資訊的過程可以參照圖2,在此不復贅述。 步驟430,將路徑資訊輸入到神經網路模型中。 亦即,將步驟420中確定的路徑資訊輸入DNN的輸入層中。 步驟440,輸出交易行為風險識別結果。 具體地,由DNN的輸出層輸出交易行為風險識別結果。此處,如果識別結果為風險的交易行為,則可以發起報警。在支付情況下,若識別結果為風險的支付行為,則可以凍結該用戶帳戶以防止財產流失。與上述神經網路模型訓練方法對應地,本發明實施例還提供的一種神經網路模型訓練裝置,如圖5所示,該裝置包括: 確定單元501,用以將預先收集的多個樣本資料登錄到梯度提升決策樹GBDT中,以確定每個樣本資料在GBDT中對應的路徑資訊。 此處,每個樣本資料具有對應的樣本標籤。 訓練單元502,用以根據確定單元501確定的每個樣本資料在GBDT中對應的路徑資訊以及樣本標籤,對神經網路模型進行訓練。 可選地,GBDT由多棵決策樹組成,每棵決策樹包括多個節點,每個節點與一個特徵相對應。 確定單元501具體上用以: 對多個樣本資料中的每個樣本資料,根據樣本資料,確定多個特徵對應的特徵值。 此處,特徵可以包括:用戶的遠端程序呼叫RPC行為資訊和/或用戶的統一資源定位器URL位址資訊。 根據特徵值,在決策樹中確定路徑資訊。 可選地,樣本標籤可以包括:正樣本標籤和負樣本標籤。上述裝置還可以包括: 處理單元503,用以對樣本標籤為正樣本標籤的樣本資料進行升取樣處理;和/或, 對樣本標籤為負樣本標籤的樣本資料進行降取樣處理。 本發明實施例裝置的各功能模組的功能,可以透過上述方法實施例的各步驟來實現,因此,本發明提供的裝置的具體操作過程,在此不復贅述。 本發明提供的神經網路模型訓練裝置,確定單元501將預先收集的多個樣本資料登錄到梯度提升決策樹GBDT中,以確定每個樣本資料在GBDT中對應的路徑資訊。訓練單元502根據每個樣本資料在GBDT中對應的路徑資訊以及樣本標籤,對神經網路模型進行訓練。由此,可以提高神經網路模型訓練的效率。 與上述交易行為風險識別方法對應地,本發明實施例還提供的一種交易行為風險識別裝置,如圖6所示,該裝置包括: 獲取單元601,用以獲取用戶的交易行為資料。 確定單元602,用以將獲取單元601獲取的交易行為資料登錄到梯度提升決策樹GBDT中,以確定交易行為資料在GBDT中對應的路徑資訊。 輸入單元603,用以將確定單元602確定的路徑資訊輸入到神經網路模型中。 輸出單元604,用以輸出交易行為風險識別結果。 可選地,GBDT由多棵決策樹組成,每棵決策樹包括多個節點,每個節點與一個特徵相對應; 確定單元602具體上用以: 根據交易行為資料,確定多個特徵對應的特徵值。 根據特徵值,在決策樹中確定路徑資訊。 其中,特徵可以包括:用戶的遠端程序呼叫RPC行為資訊和/或用戶的統一資源定位器URL位址資訊。 本發明實施例裝置的各功能模組的功能,可以透過上述方法實施例的各步驟來實現,因此,本發明提供的裝置的具體操作過程,在此不復贅述。 本發明提供的交易行為風險識別裝置,可以提高交易行為風險識別的效率和準確性。 本領域技術人員應該可以意識到,在上述一個或多個示例中,本發明所描述的功能可以用硬體、軟體、韌體或它們的任意組合來實現。當使用軟體來實現時,可以將這些功能儲存在電腦可讀媒體中或者作為電腦可讀媒體上的一個或多個指令或代碼來進行傳輸。 以上所述的具體實施方式,對本發明的目的、技術方案和有益效果進行了進一步詳細說明,所應理解的是,以上所述僅為本發明的具體實施方式而已,並不用來限定本發明的保護範圍,凡在本發明的技術方案的基礎之上,所做的任何修改、等同替換、改進等,均應包括在本發明的保護範圍之內。The following describes the embodiments of the present invention with reference to the drawings. The neural network model training method provided by the embodiments of the present invention is suitable for training neural network models such as deep neural networks (Deep Neural Network, DNN) or artificial neural networks (Artificial Neural Network, ANN). The trained neural network model can be used for pattern recognition and classification, for example, it can be used for risk recognition of trading behavior. FIG. 1 is a flowchart of a neural network model training method provided by an embodiment of the present invention. The execution subject of the method may be a device with processing capability: a server or a system or an apparatus. As shown in FIG. 1, the method specifically includes: Step 110, registering a plurality of pre-collected sample data into the gradient lifting decision tree (Gradient Boosting Decision Tree, GBDT), to determine the corresponding path information of each sample data in GBDT. Before performing step 110, the GBDT model can be trained. The specific training process will be explained later. In step 110, taking the case where the trained neural network model is used for transaction behavior risk identification as an example, the above sample data may refer to the user's transaction behavior data. Specifically, sample data can be collected from the background database of the Alipay system. Here, the sample data can be classified into the following five categories of user data: 1) The user's historical behavior information. For example, a, the number of user calls within a few days (eg, 180 days); b, the last login to the city; c, the time since the last login; d, the number of logins within a few days (eg, 90 days), etc. 2) User's transaction information. For example, a, the average payment amount within several days (eg, 90 days); b, the number of days paid within several days (eg, 180 days); c, the amount paid within several days (eg, 180 days); d, the last payment distance Wait this time. 3) Basic user information. For example, a, whether the user is single; b, whether the user is decorated; c, whether the user is married; d, the user's age; e, the length of user registration; f, user education level, etc. 4) User's remote procedure call (Remote Procedure Call, RPC) behavior information. The RPC behavior information here refers to the RPC call between the client and the server when the user uses the client. In one implementation, these operations can be collected for each user at the most recent given time window. For example, it is possible to collect the variable of the number of times the user has visited the RPC interface in the past 2 days. 5) User's Uniform Resourc e Locator (URL) address information. For the multiple sample data collected above, if a sample data is not related to the current user or the sample data can bring negative impact to the user, the sample data is classified as positive sample data. For example, if a certain transaction is operated by a non-user or brings a certain loss to the user's account and is reported, the transaction behavior data is marked as positive sample data. Otherwise, if a sample data is the user's normal trading behavior data, the sample data is marked as negative sample data. It should be noted that usually negative sample data is easier to collect. For example, information on normal payment behavior can be easily collected from the background database of the Alipay system. Therefore, the negative sample data in the sample data set will account for the vast majority, for example, greater than 99.999%. However, when the proportion of negative sample data is relatively high, the trained neural network model often has deviations. For example, it can only identify safe trading behaviors, but not risky trading behaviors. This affects the risk recognition of trading behaviors. accuracy. In order to improve the accuracy of transaction behavior risk identification, sample data can be preprocessed. In one implementation, positive sample data may be up-sampled; and/or negative sample data may be down-sampled. Among them, the up-sampling processing of the positive sample data may include: increasing the number of positive sample data by means of copying. Downsampling of negative sample data may include: reducing the amount of negative sample data by deleting or other methods. In one example, the ratio of positive sample data to negative sample data can be adjusted to 1:300. It should also be noted that the sample data after the above preprocessing can also be added with corresponding sample labels for positive and negative sample data. Specifically, positive sample labels are added to positive sample data, and negative sample labels are added to negative sample data. In step 110, registering a plurality of sample data collected in advance into the GBDT may specifically include: for each sample data, the feature values corresponding to the multiple features may be determined according to the sample data first. After that, the feature value of the feature is input into the decision tree of GBDT. The features here can be classified into multiple categories. In one implementation, some of the above features can use model variables that are deposited on the existing transaction behavior risk identification model line, and the model variables belong to the following three categories: 1) historical behavior information of users. 2) User's transaction information. 3) Basic user information. However, the above model variables need to be determined based on business data, and business data usually come from different business departments, and their collection and sorting takes a certain amount of time, so the user’s latest status cannot be obtained only through the above model variables, and thus the user’s latest Risk identification of trading behavior. To solve this problem, the present invention adds features of RPC behavior information attributable to users and features of URL address information attributable to users. In summary, the features of the present invention can be classified into the following five categories: 1) historical behavior information of users. 2) User's transaction information. 3) Basic user information. 4) User's RPC behavior information. 5) User's URL address information. Each category is as described above and will not be repeated here. After determining the corresponding feature value according to the specific sample data for the above-mentioned set feature, you can input the feature value into GBDT. The GBDT here may be composed of multiple decision trees, and each decision tree includes multiple nodes, and each node corresponds to a feature. Take a decision tree as an example, the decision tree can be shown in Figure 2, in Figure 2, node 1, node 2 and node 3 are respectively associated with the characteristics: "whether the user's gender is male", "user age is greater than 20 years old "And "whether the transaction amount exceeds 1,000 yuan" corresponds. After the feature values of the features are input into the decision tree, multiple pieces of path information can be determined in the decision tree. For example, assuming that the sample data includes that the user's gender is male, the user's age is greater than 20 years old, and the transaction amount exceeds 1,000 yuan, the determined path information can be shown as the thick line in Figure 2. As an exemplary illustration, FIG. 2 only shows one piece of path information. In fact, when the sample data is registered in GBDT, multiple pieces of path information can be determined, and the present invention will not repeat them here. It should be noted that, in the present invention, before inputting the feature value to the GBDT, the feature value may also be expressed as a one-hot feature vector. In the case where the feature vector corresponding to the feature value is also determined, the above input of the feature value into the GBDT may be replaced by: inputting the feature vector corresponding to the feature value into the decision tree to determine the corresponding path information. Among them, the process of determining the feature vector of the feature value can be exemplified as follows: Take the feature as "user gender" for example, if the user's gender is male, that is, the feature value of the feature is "male", then the feature value corresponds to The feature vector can be: [0 1]. If the user's gender is female, that is, the feature value of the feature is "female", the feature vector corresponding to the feature value may be: [1 0]. Taking the RPC behavior information whose characteristics are users as an example, the determination of the characteristic vector corresponding to the characteristic value can be achieved through the following two methods: In the first implementation method, the rule is first set: when it occurs, the identifier is 1, Otherwise it is 0. Specifically, assume that the preset RPC behavior information is: a, b, and c. A certain sample data contains the user's RPC behavior information within two days: a, a and b, that is, the characteristic values are: a, a and b. Then the corresponding feature vector can be: [1 1 0]. In another implementation, a rule can be set: count the frequency of preset RPC behavior information, and then normalize. Specifically, assume that the preset RPC behavior information is: a, b, and c. A certain sample data contains the user's RPC behavior information within two days: a, a, b, b and c, that is, the characteristic value is: a, a, b, b and c. Then the corresponding feature vectors can be: 2, 2 and 1. Because it needs to be normalized, the final feature vector is: [0.4 0.4 0.2]. It should be noted that the above-mentioned representation of feature values as feature vectors belongs to the conventional conventional technology, and will not be repeated here. It should be noted that, in order to improve the accuracy of the neural network model, more features are set in the present invention, so that multiple feature values are determined. For more and more feature values, the processing often takes a lot of time, limited by the number of feature values observed at the same time, it is difficult for people to conduct in-depth analysis of the relationship between multiple feature values and manually generate New feature value. The present invention obtains the path information by registering the sample data in GBDT. The path information includes multiple feature values. Thus, the number of feature values can be greatly reduced, thereby significantly reducing manual operations. Step 120: Train the neural network model according to the path information and sample labels corresponding to each sample data. The neural network model here may include DNN or ANN. Among them, DNN has developed rapidly in recent years. Compared with the traditional shallow models (such as Logistic Regression (LR) and Random forest (RF)), DNN has its unique advanced nature: model expression ability Powerful, suitable for large data and decentralized training. Therefore, in this specification, the training DNN is taken as an example for description. In the present invention, the training process of the DNN can be as shown in FIG. 3, in FIG. 3, the input layer of the DNN is used to input various path information in the GBDT, and the output layer can output the first prediction result. It can be understood that for each sample data, that is, after the path information corresponding to the sample data is input to the DNN, the DNN will output the corresponding first prediction result. For multiple sample data in the sample set, if the probability that the first prediction result matches the sample label of the sample data reaches a preset threshold, the preset threshold here can be set according to the empirical value, it can be considered that it has been obtained Optimized DNN. Understandably, the number of DNN layers in Figure 3 can be changed with the number of path information. Through the experimental invention, the neural network model trained by the present invention will be better than other models (LR or RF). At the same time, the feature processing time is greatly reduced, and the overall modeling process becomes much faster. The following explains how to train the GBDT model: After determining the feature values corresponding to multiple features based on each sample data, you can input the feature values corresponding to multiple features into each decision tree of GBDT. After that, the results of each decision tree are accumulated to determine the second prediction result. It is understandable that for each sample data, the GBDT model will output the corresponding second prediction result. For multiple sample data in the sample set, if the probability that the second prediction result matches the sample label of the sample data reaches a preset threshold, the preset threshold here can be set according to the empirical value, it can be considered that it has been obtained Optimized GBDT model. If the probability that the second prediction result matches the sample label of the sample data does not reach the preset threshold, you can continue to execute the above input by adjusting the number of decision trees, the depth of the decision tree, and regularization terms (used to represent features) And output operations until the preset threshold is reached. In summary, the present invention has the following advantages: 1) Since the features of the present invention include the feature of the user's RPC behavior information, the neural network model trained by the present invention can meet the timeliness requirements, that is, it can recognize User's latest trading behavior. 2) The accuracy of the neural network model trained by the present invention is higher than that of the traditional shallow model. 3) By registering the sample data in GBDT, the route information was obtained. The path information is composed of multiple feature values, that is, the path information contains the information of multiple dimensions of the sample data, which can greatly reduce the amount of data input by the DNN input layer, which can improve nerve The efficiency of network model training. It should be noted that, after training the neural network model through the steps shown in Figure 1, the neural network model can be deployed online and the risk of the user's trading behavior can be identified. FIG. 4 is a schematic diagram of a process of a method for identifying a transaction behavior risk provided by the present invention. As shown in FIG. 4, the method may include: Step 410, obtaining user's transaction behavior data. The transaction behavior data here has the same definition as the above sample data, and will not be repeated here. Step 420: Register the transaction behavior data in the gradient lifting decision tree GBDT to determine the corresponding path information of the transaction behavior data in the GBDT. The above-mentioned GBDT is composed of multiple decision trees, and each decision tree includes multiple nodes, and each node corresponds to a feature. The step of registering the transaction behavior data in the gradient lifting decision tree GBDT in step 420 to determine the corresponding path information of the transaction behavior data in the GBDT may specifically include: according to the transaction behavior data, determining the characteristic values corresponding to multiple characteristics; Eigenvalues to determine path information in the decision tree. Among them, the process of determining the path information can refer to FIG. 2 and will not be repeated here. Step 430, input the path information into the neural network model. That is, the path information determined in step 420 is input into the input layer of the DNN. Step 440: Output the result of transaction behavior risk identification. Specifically, the output layer of the DNN outputs the transaction behavior risk recognition result. Here, if the recognition result is risky trading behavior, an alarm can be initiated. In the case of payment, if the recognition result is a risky payment behavior, the user account may be frozen to prevent the loss of property. Corresponding to the above neural network model training method, a neural network model training device provided in an embodiment of the present invention, as shown in FIG. 5, the device includes: a determination unit 501, which is used to collect a plurality of sample data collected in advance Log into the gradient lifting decision tree GBDT to determine the corresponding path information of each sample data in GBDT. Here, each sample data has a corresponding sample label. The training unit 502 is used to train the neural network model according to the corresponding path information and sample labels in GBDT of each sample data determined by the determining unit 501. Optionally, GBDT consists of multiple decision trees, each of which includes multiple nodes, and each node corresponds to a feature. The determination unit 501 is specifically used to: determine the characteristic values corresponding to the multiple characteristics according to the sample data for each of the multiple sample data. Here, the features may include: user's remote procedure call RPC behavior information and/or user's uniform resource locator URL address information. According to the characteristic value, determine the path information in the decision tree. Alternatively, the sample label may include: a positive sample label and a negative sample label. The above device may further include: a processing unit 503 for upsampling processing of the sample data whose sample label is a positive sample label; and/or downsampling processing for the sample data whose sample label is a negative sample label. The functions of the functional modules of the device according to the embodiments of the present invention can be implemented through the steps of the above method embodiments. Therefore, the specific operation process of the device provided by the present invention will not be repeated here. In the neural network model training device provided by the present invention, the determination unit 501 registers a plurality of sample data collected in advance in the gradient lifting decision tree GBDT to determine the corresponding path information of each sample data in the GBDT. The training unit 502 trains the neural network model according to the corresponding path information and sample labels in GBDT of each sample data. Thus, the efficiency of neural network model training can be improved. Corresponding to the above transaction behavior risk identification method, an embodiment of the present invention also provides a transaction behavior risk identification device. As shown in FIG. 6, the device includes: acquisition unit 601, which is used to acquire user's transaction behavior data. The determining unit 602 is used to register the transaction behavior data obtained by the obtaining unit 601 in the gradient lifting decision tree GBDT to determine the path information corresponding to the transaction behavior data in the GBDT. The input unit 603 is used to input the path information determined by the determining unit 602 into the neural network model. output unit 604 is used to output the risk identification result of transaction behavior. Optionally, the GBDT is composed of multiple decision trees, and each decision tree includes multiple nodes, and each node corresponds to a feature; determination unit 602 is specifically used to: determine characteristics corresponding to multiple features based on transaction behavior data value. According to the characteristic value, determine the path information in the decision tree. Among them, the characteristics may include: the user's remote program call RPC behavior information and/or the user's uniform resource locator URL address information. The functions of the functional modules of the device according to the embodiments of the present invention can be implemented through the steps of the above method embodiments. Therefore, the specific operation process of the device provided by the present invention will not be repeated here. The transaction behavior risk identification device provided by the present invention can improve the efficiency and accuracy of transaction behavior risk identification. Those skilled in the art should realize that in one or more of the above examples, the functions described in the present invention may be implemented by hardware, software, firmware, or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on the computer-readable medium. The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the scope of the present invention. The scope of protection, any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solution of the present invention, shall be included in the scope of protection of the present invention.