TWI759620B - Method and apparatus for determining user's financial default risk and computer-readable storage medium and computing device - Google Patents

Abstract

The embodiments of this specification provide a method and device for determining a user's financial default risk. According to an embodiment of the method, different models based on a recurrent neural network are used to process the user's short-term operation sequence and long-term behavior data set sequence, respectively. The user conducts risk assessment and comprehensively processes at least two different processing results to obtain the user's financial default risk result. In this way, on the one hand, more in-depth time series data is used, and on the other hand, the final result is determined by synthesizing the output results of multiple models. This implementation can improve the effectiveness of determining the user's financial default risk.

Description

Method and apparatus for determining user's financial default risk, computer-readable storage medium and computing device

One or more embodiments of this specification relate to the field of computer technology, and in particular, to a method and apparatus for determining a user's risk of financial default by means of a computer.

With the development of computer and Internet technology, more and more businesses are realized through computing platforms, such as commodity trading, money payment, financial lending, insurance claims and so on. However, in the execution and processing of many businesses, there are some user operations that pose certain financial risks to financial platforms or other users, such as requesting first-pay-after-pay services, using Huabei, Baitiao and other overdraft services, applying for loans Wait. This requires pre-assessment and judgment of the user's financial default risk.

In the conventional technology, in order to prevent and reduce the above risks, the user's credit is often evaluated through information such as the user's identity and historical transaction behavior. However, these information are usually static information, which cannot reflect the relationship between user behavior and other information, and for new users, it may not be possible to obtain the corresponding information, so that the credit risk cannot be determined. Therefore, a more effective method is needed, using more network data and more effective evaluation methods, to conduct a comprehensive analysis of users, and to improve the effectiveness of user financial default risk assessment.

One or more embodiments of this specification describe a method and apparatus for determining a user's financial default risk, which can more effectively analyze and evaluate the user's credit risk.

According to a first aspect, a method for determining a user's financial default risk is provided, comprising: acquiring a short-term operation sequence of a user to be evaluated in a first time period, and a long-term behavior data set sequence in a second time period, the second The time period is greater than the first time period, the short-term operation sequence includes a plurality of pieces of operation information related to the operation behavior of the user to be evaluated, arranged in chronological order, and the long-term behavior data set sequence includes chronological order. a plurality of behavior data sets, each behavior data set corresponds to a preset third time period respectively, and the behavior data set includes behavior information related to the transaction behavior of the user to be evaluated; using the first model based on the recurrent neural network to process the short-term operation sequence to obtain a first output result; use a second model based on a recurrent neural network to process the long-term behavior data set sequence to obtain a second output result; at least for the first output result Predetermined processing is performed with the second output result, and the financial default risk of the user to be evaluated is determined according to the processing result.

In some embodiments, the operation information includes at least one of the following: browsing information, click information, logged-in application, logged-in device, and geographic location information.

In some embodiments, the transaction behavior information includes at least one of the following: transaction time, transaction object, and transaction amount.

In some embodiments, the method further includes: obtaining the to-be-evaluated attribute information of the user; processing the attribute information using a prediction model to obtain a third output result; and performing predetermined processing on at least the first output result and the second output result, and determining according to the processing result The financial default risk of the user to be evaluated includes: performing the predetermined processing on the first output result, the second output result and the third output result, and determining the user to be evaluated according to the processing result. financial default risk.

In some embodiments, the predetermined processing includes at least one of the following: averaging; taking a maximum value; inputting a preset logistic regression model as a feature to obtain a logistic regression result.

In some embodiments, the first model/the second model comprises a multi-layer stacked long short term memory model LSTM.

In some embodiments, the training samples of the first model/the second model include a plurality of labeled users, and the labeled users have at least pre-labeled credit labels.

In some embodiments, the plurality of tagging users include a first tagging user, and the first credit tag corresponding to the first tagging user is determined by obtaining the credit of the first tagging user within a predetermined period of time. record; determining the first credit tag based on the credit record.

In some embodiments, the determining the first credit label based on the credit record comprises: determining from the credit record the number of times of trustworthiness and the number of times of untrustworthiness of the first marked user; In the case that the ratio of the times of keeping promises exceeds a preset ratio threshold, it is determined that the first credit tag is a dishonest user.

In some embodiments, the determining the first credit label based on the credit record includes: detecting whether the number of untrustworthy users of the first marking user is zero; in the case that the number of untrustworthy times is non-zero, determining the The first credit label is described as a dishonest user.

According to a second aspect, an apparatus for determining a user's financial default risk is provided, comprising: an acquisition unit configured to acquire a short-term operation sequence of a user to be evaluated in a first time period, and a long-term behavior data set in a second time period sequence, the second time period is longer than the first time period, the short-term operation sequence includes a plurality of pieces of operation information related to the operation behavior of the user to be evaluated arranged in time sequence, and the long-term behavior data set sequence It includes a plurality of behavior data sets arranged in chronological order, each behavior data set corresponds to a preset third time period, and the behavior data set includes behavior information related to the transaction behavior of the user to be evaluated; the first processing unit , which is configured to process the short-term sequence of operations using a first model based on a recurrent neural network to obtain a first output result; a second processing unit is configured to use a second model based on a recurrent neural network to process the short-term operation sequence; processing the long-term behavior data set sequence to obtain a second output result; a determining unit configured to perform predetermined processing on at least the first output result and the second output result, and determine the The financial default risk of the user to be evaluated.

According to a third aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed in a computer, the computer is caused to execute the method of the first aspect.

According to a fourth aspect, a computing device is provided, including a memory and a processor, wherein the memory stores executable program codes, and when the processor executes the executable program codes, it realizes method of the first aspect. Through the method and device for determining a user's financial default risk provided by the embodiments of this specification, different models based on the recurrent neural network are used to process the user's short-term operation sequence and long-term behavior data set sequence, and perform risk assessment on the user. At least two different processing results are comprehensively processed to obtain the user's financial default risk result. In this way, on the one hand, more in-depth time series data is used, and on the other hand, the final result is determined by synthesizing the output results of multiple models. Thus, the effectiveness of determining the user's financial default risk can be improved.

在表1示出的操作日誌示意表格中，每一行代表一條操作資訊。相應的短期操作序列可以表示為[瀏覽商品頁面；點擊購物車頁面；……]，還可以添加時間資訊，表示為[10點01：瀏覽商品頁面:10點20：點擊購物車頁面；……]。在一些實施例中，操作序列還可以用向量表示，例如[

，

，

……]，其中

可以表示瀏覽商品頁面的操作向量，如[1，0，0，1……]，

可以表示點擊購物車頁面，如[0，1，0，0……]，等等。在一個實施例中，還可以透過詞向量模型(如word2vec)將各條操作資訊藉由詞向量表示，在此不再贅述。如此，可以獲取到用戶在短期內的各種操作相關的具有時序特徵的資訊。 上述的長期行為資料集序列可以包括，按照時間順序排列的多個行為資料集。在第二時間段內，可以將預設的第三時間段(例如1天)作為一個週期。通常，第二時間段可以是第三時間段的整數倍。對於這樣的週期，可以獲取每個週期內與用戶的交易行為有關的各種資訊，將這樣的資訊整理為一個資料集。具體地，在一個實施例中，可以對用戶在上述週期內的交易行為資訊進行統計，形成資料集。 在一個實施例中，與用戶的交易行為相關的行為資訊可以包括，在購物平臺瀏覽商品、下單、所購買商品類型、商品價格、是否支付、付款金額等行為。 在另一個實施例中，與用戶的交易行為相關的行為資訊可以包括，是否透過金融借貸平臺進行借款、借款額度、是否還款等等。 在另一個實施例中，與用戶的交易行為相關的行為資訊可以包括，與其他用戶之間的轉帳行為、轉帳金額等等。 在更多實施例中，與用戶的交易行為相關的行為資訊還可以包括更多資訊，在此不再贅述。 進一步地，可以對每個週期內的用戶交易行為資訊進行統計，形成相應週期內的行為資料集。例如，{瀏覽商品頁面：5次；支付：1次……}。其中，5次、1次等還可以替換成與次數正相關的權重值。瀏覽商品頁面、支付等還可以替換成字元或詞向量。如此，對於多個週期的行為資料集，可以構成具有時序特徵的長期行為資料集序列。藉由長期行為資料集序列，可以體現用戶長期的交易行為習慣等。例如每個月從固定日期(如工資到賬的日期)購買行為較多，交易金額相對較大，逐漸衰減到無購買行為、或較少購買行為、交易金額相對較小等，直到下個月的上述固定日期再重複這樣的過程。 可以理解的是，在一些實現中，上述第一時間段、第二時間段都可以是，從當前時間開始向前回溯的時間段。 藉由以上描述可知，上述短期操作序列和長期行為資料集序列都是具有時序特徵的資料序列。對於這樣的序列，可以藉由循環神經網路模型來處理，進而從時序的角度對用戶的行為進行預測。 可以理解的是，循環神經網路(RNN，Recurrent Neural Networks)是一種時間遞迴神經網路，可用於處理序列資料。在RNN中，一個序列當前的輸出與其前面的輸出相關聯。具體地，RNN會對前面的資訊進行記憶並應用於當前輸出的計算中，即隱藏層之間的節點是有連接的，並且隱藏層的輸入不僅包括輸入層的輸出還包括上一時刻隱藏層的輸出。如圖3示出的循環神經網路時序示意圖中，第t次的隱含層狀態可以表示為：S_t =f(U*X_t +W*S_t-1 )； 其中，X_t 為第t次輸入層的狀態，S_t-1 為第t-1次隱含層狀態，f為計算函數，W、U為權重。如此，RNN將之前的狀態循環回當前輸入，考慮了歷史輸入的影響，因而適合於具有時序的資料序列。 更進一步地，在一個實施例中，在RNN架構下，可以採用長短期記憶模型(LSTM，Long Short Term Memory)進行上述序列資料的處理。 如前所述，在RNN中當前隱含層狀態依賴於之前的狀態輸出，因此在處理時，需要將當前的隱含態的計算與前n次的計算關聯，即S_t = f(U*X_t + W₁ *S_t-1 + W₂ *S_t-2 + … + W_n *S_t-n )。隨著n的增大，計算量呈指數式增長，導致模型訓練的時間大幅增加。為此，提出LSTM模型來解決長期依賴的問題。 The solution provided in this specification will be described below with reference to the accompanying drawings. FIG. 1 is a schematic diagram of an implementation scenario of an embodiment disclosed in this specification. As shown in Figure 1, the user can perform various operations through the terminal, such as browsing web pages, clicking on hyperlinks on the page, etc. It is also possible to conduct various exchanges and interactions with the background server through the network, such as performing various interactions with Lending-related behaviors, such as applying for a loan, repaying, applying for deferral of repayment, etc. Correspondingly, the terminal can record the user's operation information through logs, etc., and the background server can also record the user's behavior related to the service provided by the background server. For example, a user applies for a loan through "Ant Huabei" under Alipay, and then repays the loan every month. At this time, the online platform is the Alipay platform, and the background server can be the Alipay platform server. It can be understood that the background server may be a centralized server, a distributed server, or a plurality of servers that are completely independent of each other, which is not limited here. In order to evaluate the user's credit risk, the computing platform can obtain the corresponding user information from various lending platforms or terminals, and the computing platform can conduct a comprehensive analysis to determine the user's financial default risk. In the embodiment disclosed in the specification, after the computing platform obtains relevant user data, it can use machine learning and recurrent neural network, and adopt a framework combining multiple neural network models to conduct a comprehensive analysis of these data, and then conduct financial risk analysis. Evaluate. The above computing platform can be any device, equipment and system with computing and processing capabilities, such as a server, which can be used as an independent computing platform or integrated into a background server that provides support for certain services (such as lending). in the device. More specifically, on the one hand, the computing platform can obtain the operation sequence formed by various operations of the user in a short period of time (such as 24 hours), and process it through the first model based on the recurrent neural network to mine the short-term operation behavior of the user. to get the first output result. On the other hand, the computing platform can also obtain the user's behavior data sets in several periods (for example, one day), and each behavior data set can include the user's behavior statistical results in the corresponding period. The second model of the road processes the long-term behavior data set sequence formed by these behavior data sets in chronological order, and obtains the second output result. Further, the computing platform can also comprehensively process at least the first output result and the second output result, so as to perform a final financial risk assessment. The specific process by which the computing platform assesses the risk of financial default is described below. Figure 2 illustrates a flow chart of a method of determining a user's risk of financial default, according to one embodiment. The execution body of the method can be any system, device, device, platform or server with computing and processing capabilities, such as the computing platform shown in FIG. 1 . More specifically, for example, it may be a lending server that supports lending services. As shown in FIG. 2 , the method includes the following steps: Step 21: Acquire the short-term operation sequence of the user to be evaluated in the first time period and the long-term behavior data set sequence in the second time period, wherein the second time period is greater than the first time period. For a period of time, the short-term operation sequence includes multiple pieces of operation information related to the operation behavior of the user to be evaluated, arranged in chronological order, and the long-term behavior data set sequence includes multiple behavior data sets arranged in chronological order, and each behavior data set corresponds to In the preset third time period, each behavior data set includes behavior information related to the transaction behavior of the user to be evaluated; step 22, using the first model based on the recurrent neural network to process the short-term operation sequence to obtain the first output Results; Step 23, use the second model based on the recurrent neural network to process the long-term behavior data set sequence to obtain the second output result; Step 24, at least perform predetermined processing on the first output result and the second output result, and according to The results are processed to determine the financial default risk of the user to be evaluated. First, in step 21, the short-term operation sequence of the user to be evaluated in the first time period and the long-term behavior data set sequence in the second time period are acquired. Here, the first period of time may be a relatively short period of time, such as 1 day (24 hours), 12 hours, and the like. The second period of time may be a relatively long period of time, such as 1 month, 3 months, 1 year, and so on. The second period of time may be substantially greater than the first period of time. In this way, for the convenience of description, the first time period may correspond to a short term, and the second time period may correspond to a long term. The short-term operation sequence may include a plurality of pieces of operation information related to the operation behavior of the user to be evaluated, arranged in chronological order. Operational information may include, but is not limited to, one or more of browsing information, click information, logged-in applications, logged-in devices, geographic location information, transaction information, and the like. Browsing information may include, for example, pages browsed, page URLs, website function variable names, and the like. The click information may include, for example, the clicked hyperlink, the page corresponding to the clicked hyperlink, the clicked button for submitting the form, and the like. The logged-in device may be, for example, a desktop computer, a notebook computer, a tablet computer, a smart phone, and the like. Geolocation information can be determined from the location information of the logged in device. For example, the positioning information of the device can be determined according to the SIM card on the smart phone that is currently closest to the communication base station, the IP address of the computer accessing the network, the software/hardware positioning device (such as GPS positioning system) on the device, etc. . In one implementation, the above-mentioned operation information may be obtained from the operation log of the device currently logged in by the user. In another implementation, the above-mentioned operation information can also be obtained through the operation logs of different devices logged in with the same user ID during the login period of the user ID. In some implementations, the above-mentioned operation information can also be obtained through the user IDs of similar users (or the same user) on different platforms determined according to the big data, and the operation logs during login to the corresponding platform. This is not limited. The short-term operation sequence can arrange the pieces of operation information in the first time period according to the time sequence of occurrence. Please refer to Table 1, assuming the following operation information is obtained from the operation log: Table 1 Operation log representation

In the operation log schematic table shown in Table 1, each row represents a piece of operation information. The corresponding short-term operation sequence can be expressed as [Browse the product page; click on the shopping cart page; ...], and time information can also be added, expressed as [10:01: browse the product page: 10:20: click on the shopping cart page; ... ]. In some embodiments, the sequence of operations can also be represented by a vector, such as [

,

,

……],in

It can represent the operation vector of browsing product pages, such as [1, 0, 0, 1...],

It can represent a click on the shopping cart page, such as [0, 1, 0, 0...], and so on. In one embodiment, each piece of operation information can also be represented by a word vector through a word vector model (such as word2vec), which will not be repeated here. In this way, information with time-series characteristics related to various operations of the user in a short period of time can be obtained. The above-mentioned long-term behavior data set sequence may include a plurality of behavior data sets arranged in time sequence. In the second time period, a preset third time period (for example, 1 day) may be used as a period. Typically, the second time period may be an integer multiple of the third time period. For such a cycle, various information related to the user's transaction behavior in each cycle can be obtained, and such information can be organized into a data set. Specifically, in one embodiment, the user's transaction behavior information in the above-mentioned period may be counted to form a data set. In one embodiment, the behavior information related to the user's transaction behavior may include behaviors such as browsing products on the shopping platform, placing an order, the type of the purchased product, the price of the product, whether to pay, and the amount of payment. In another embodiment, the behavior information related to the user's transaction behavior may include whether to borrow money through the financial lending platform, the amount of the loan, whether to repay the loan, and the like. In another embodiment, the behavior information related to the user's transaction behavior may include the transfer behavior with other users, the transfer amount, and the like. In more embodiments, the behavior information related to the user's transaction behavior may further include more information, which will not be described herein again. Further, statistics on the user's transaction behavior information in each cycle can be performed to form a behavior data set in the corresponding cycle. For example, {View product page: 5 times; Pay: 1 time...}. Among them, 5 times, 1 time, etc. can also be replaced with weight values that are positively related to the times. Browsing product pages, paying, etc. can also be replaced with character or word vectors. In this way, for the behavior data sets of multiple periods, a long-term behavior data set sequence with time series characteristics can be formed. Through the long-term behavior data set sequence, the user's long-term transaction behavior habits can be reflected. For example, every month from a fixed date (such as the date of payment of wages), there are more purchases, and the transaction amount is relatively large, and gradually decays to no purchase behavior, or less purchase behavior, and the transaction amount is relatively small, etc., until the next month. This process is repeated for the above fixed dates. It can be understood that, in some implementations, the above-mentioned first time period and second time period may both be time periods that are traced back from the current time. It can be seen from the above description that the above-mentioned short-term operation sequence and long-term behavior data set sequence are data sequences with time series characteristics. For such a sequence, it can be processed by a recurrent neural network model, and then the user's behavior can be predicted from the perspective of time series. It can be understood that Recurrent Neural Networks (RNN) are a time-recurrent neural network that can be used to process sequential data. In an RNN, the current output of a sequence is related to its previous output. Specifically, the RNN will memorize the previous information and apply it to the calculation of the current output, that is, the nodes between the hidden layers are connected, and the input of the hidden layer includes not only the output of the input layer but also the hidden layer at the previous moment. Output. In the schematic diagram of the recurrent neural network shown in Figure 3, the hidden layer state of the t-th time can be expressed as: S _t =f(U*X _t +W*S _t-1 ); where X _t is the th The state of the t-th input layer, S _t-1 is the t-1th hidden layer state, f is the calculation function, and W and U are the weights. In this way, the RNN loops the previous state back to the current input, considering the influence of historical input, and thus is suitable for data sequences with time series. Further, in an embodiment, under the RNN architecture, a long short term memory model (LSTM, Long Short Term Memory) can be used to process the above sequence data. As mentioned above, the current hidden layer state in RNN depends on the previous state output, so when processing, it is necessary to associate the calculation of the current hidden state with the previous n calculations, that is, S _t = f(U* X _t + W ₁ *S _t-1 + W ₂ *S _t-2 + … + W _n *S _tn ). As n increases, the amount of computation grows exponentially, resulting in a substantial increase in model training time. To this end, an LSTM model is proposed to solve the problem of long-term dependencies.

In the LSTM model, some information that is no longer needed is discarded by setting a "forgetting gate" that allows information to selectively pass through, so as to judge and mask the input unnecessary interference information, so as to better understand the data sequence. Perform analytical processing.

Please refer to Figure 3, for a neuron A in LSTM, X _t-1 , X _t , X _t+1 represent the input at time t-1, time t and time t+1 respectively, S _t-1 , S _t and S _t+1 represent the state of the neuron at time t-1, time t and time t+1, respectively, while C _t-1 , C _t and C _t+1 represent time t-1, time t and time t+1 respectively. Output at time t+1, where: S _t =g(U*X _t +W*C _t-1 +b _s );C _t =f(V*S _t-1 +b _c );S _t+1 =g(U*X _t +W*C _t +b _s ); C _t+1 =f(V*S _t +b _c ); where U, W, and V are weights.

It can be seen that in the LSTM model, the current state of each neuron A is jointly determined by the input at the current moment and the output at the previous moment, and the current output of each neuron A is related to the state at the previous moment. By analyzing the data sequence through the LSTM model, information can be selectively memorized and long-distance data dependencies can be mined.

In one embodiment, a multi-layer superimposed LSTM model may also be used to process a sequence composed of multiple pieces of data according to time series.

In the embodiment shown in FIG. 2 , the above-mentioned short-term operation sequence and long-term behavior data set sequence are processed through step 22 and step 23 respectively. Can It should be understood that, step 22 and step 23 may be executed concurrently, or may be executed sequentially, which is not limited in this embodiment of the present specification.

Specifically, in step 22, a first recurrent neural network is used to process the above-mentioned short-term operation sequence to obtain a first output result. For each piece of data in the short-term operation sequence, input the LSTM model in turn. For each piece of data, each element corresponds to each neuron in the input layer, and the current output of each neuron affects the output of the next piece of data after it is logged into the neuron.

Figure 4 shows a schematic diagram of processing a short-term sequence of operations using a recurrent neural network, according to one embodiment. As shown in Figure 4, serial numbers 1, 2, 3, 4, 5, 6, 7... represent each piece of information in the short-term operation sequence. For example, serial number 1 corresponds to the first operation information: browse shopping page a, serial number 2 corresponds to The second operation information: pay 30 yuan, and so on. In the example of Figure 4, the recurrent neural network includes a multi-layer stacked LSTM model. As an illustration, Figure 4 only shows the time series of the multi-layer stacked LSTM model at time t-1, t, and t+1. In Figure 4, at time t-1, the multi-layered LSTM model accepts a piece of operation information with a serial number of 3. The operation information is processed by the LSTM model, and the output of each layer is recorded. At time t, the LSTM model accepts a piece of operation information with sequence number 4. When the LSTM model processes this piece of operation information, it also considers the output records at time t-1 in each layer. And so on until the above short-term sequence of operations is fully processed.

The first output result can be obtained through the processing of the multi-layer recurrent neural network. The first output result is an intermediate result related to each piece of operation information in the short-term operation sequence and its generation order. In one embodiment, the first output may be a user credit/risk score. At this time, the recurrent neural network for processing the above-mentioned short-term operation sequence can be trained with multiple labeled users as samples. These multiple labeled users correspond to historical short-term operation sequences and pre-labeled credit labels, respectively. Wherein, the historical short-term operation sequence of the marked user and the above-mentioned short-term operation sequence may have a consistent time entry point. The time entry point here can be understood as the timing for determining the user's financial default risk in this embodiment. For example, for a lending platform (such as "Huabei"), the above short-term operation sequence is the sequence obtained by the user's operation information within 24 hours before the platform's service is activated. For users of the service, the historical short-term operation sequence of the marked user is the sequence obtained by the corresponding user's operation information within 24 hours before the service is activated on the platform. The time entry point in this example is that the user activates the service on the platform. The credit label corresponding to the marked user may be marked according to the user's credit status on the platform or other platforms. The credit records of the corresponding users within a predetermined period of time may be acquired, and the corresponding credit tags may be determined based on the credit records. The predetermined time period here may be, for example, 3 months, 6 months, 1 year, etc. after the corresponding platform starts the service. It can be understood that the labeling information for users may include labels such as "untrustworthy users" and "trustworthy users", and the labels may also be represented by numerical values (eg, 1, 0). The user's annotation information can be determined manually or through a computer. In one embodiment, the number of times of trustworthiness and the number of times of untrustworthiness of the user may be determined from the user's credit record, and when the ratio of the number of times of untrustworthiness to the number of times of trustworthiness exceeds a preset ratio threshold (eg 1:1), determine the corresponding user 's credit label is "Untrustworthy User". Otherwise, determine the credit label of the corresponding user as "trustworthy user". In another embodiment, it may be detected whether the number of times of untrustworthiness of the user is zero, and if the number of untrustworthy times of the user is non-zero, it is determined that the credit label of the corresponding user is "untrustworthy user". Otherwise, determine the credit label of the corresponding user as "trustworthy user". It is worth noting that, in the embodiment of this specification, although the labeling information of the user as the sample can be determined through a computer, the user's financial default risk cannot be evaluated directly by using this labeling method. This is because the method for determining the user's annotation information is simple and rough, and often corresponds to one of two results to represent the user's financial default risk bias, and the embodiments of this specification aim to provide a solution for determining the user's financial default risk. This scheme is universal and can conduct risk assessment more accurately. For new users without credit records, old users with less credit records, etc., more accurate risk estimates can be given. Information methods cannot be achieved. For distinction, the recurrent neural network trained above is referred to as the first recurrent neural network. Through this step, the impact of the short-term operation information of the user to be evaluated on the user's risk can be mined. In addition to using the first recurrent neural network to process the short-term operation sequence of the user to be evaluated, on the other hand, through step 23, the second recurrent neural network is also used to process the long-term behavior data set sequence to obtain the second output result. For each behavioral dataset in the long-term behavioral dataset sequence, the LSTM model can be input in turn. For a behavior data set, each element corresponds to each neuron in the input layer, and the current output of each neuron affects the output of the next data entry after the neuron is entered. 5 illustrates a schematic diagram of processing a sequence of long-term behavioral datasets using a second recurrent neural network, according to one embodiment. As shown in Figure 5, the serial number is used to represent the behavior data set in each short period according to the date No. 3, No. 4... For example, No. 3 corresponds to the behavior data set of the 3rd day, and No. 4 corresponds to the behavior data of the 4th day. set, etc. As an illustration, Figure 5 only shows the time series of the multi-layer stacked LSTM model at time t-1, t, and t+1. In Figure 5, at time t-1, the multi-layered LSTM model accepts the behavior data set on the 3rd day. The behavior dataset is processed by the LSTM model, and the output of each layer is recorded. At time t, the LSTM model accepts the behavior data set for the 4th day. When the LSTM model processes this behavioral dataset, it also considers the output records at time t-1 in each layer. And so on, until the above long-term behavior dataset sequence is completely processed. After processing by the second recurrent neural network, the second output result can be obtained. The second output is an intermediate result related to the timing of the individual behavioral datasets of the long-term behavioral dataset sequence. In one embodiment, the second output may be another user credit/risk score. It can be understood that the recurrent neural network processing the above-mentioned long-term behavior data set sequence can also be trained with multiple labeled users as samples. These multiple labeled users correspond to a sequence of historical long-term behavior data sets and pre-labeled credit labels, respectively. Among them, the historical long-term behavior data set sequence of the labeled user and the long-term behavior data set sequence of the user to be evaluated also have a consistent time entry point. For example, for a lending platform (such as "Huabei"), the above long-term behavior data set sequence is the user's behavior information within 1 month before the platform's service is opened, and the sequence is obtained by chronological order, then the marked user can be in the For users who have activated the service on the platform, the sequence of the historical long-term behavior data set of the marked user is the sequence obtained by the behavior information of the corresponding user within one month before the service is activated on the platform. The time entry point here is that the user activates the service on the platform. The labeling method for labeling the credit label corresponding to the user is the same as the labeling method in step 22, and details are not repeated here. Through this step, the impact of the long-term behavior information of the user to be evaluated on the user risk can be mined. Further, on the basis of step 22 and step 23, through step 24, at least based on the predetermined processing of the first output result and the second output result, to determine the financial default risk of the user to be evaluated according to the processing result. It is understood that the predetermined process is a pre-stored process method. As shown in Figure 6. Financial default risk can be represented by fractions, decimals, categories, biases, etc., which are not limited here. In one embodiment, the processing method may be a simple rule, such as summing, averaging, maximizing, and so on. Taking the average value as an example, the average value of the first output result and the second output result can be used as the magnitude of the financial default risk of the user to be evaluated. Optionally, it is also possible to first judge whether the processed results (such as the first output result and the second output result) include abnormal values (such as exceeding a predetermined range, or being empty, etc.), and if they include abnormal values, then exclude them. take another value. In this way, in the case that one of the short-term operation sequence or long-term behavior sequence of the user to be evaluated cannot be obtained, etc., the financial default risk of the user to be evaluated can still be obtained. In one embodiment, the processing method may further include inputting the above-mentioned output result as a feature into a logistic regression model (such as an LR model), and the logistic regression model is used to perform linear regression on the first output result and the second output result based on the linear regression model. , perform a logical operation so that the output value of the model falls within a preset interval (eg, between 0-1). The logistic regression model may be a pre-trained model or a pre-set model, which is not limited herein. The preset model may be a model that is considered to determine the calculation method and parameters. It is understandable that both short-term operation information and long-term behavior information are unique vertical characteristics of users. Through this unique vertical characteristics, the evaluation of users' financial default risk is more accurate. At the same time, since multiple time series data are processed separately through multiple models, and multiple evaluation results are synthesized, the impact of one-sided evaluation or invalid data on the evaluation results can be avoided. According to a possible design, the above method may further include: acquiring attribute information of the user to be evaluated, processing the attribute information by using a pre-trained prediction model, and obtaining a third output result. Here, the attribute information may be information representing user attributes, such as gender, age, occupation, income, property status, and the like. The training samples for the predictive model can include multiple labeled users. These labeled users respectively correspond to attribute information data sets and credit labels, which will not be repeated here. Wherein, the method for determining the credit label is the same as before, and will not be repeated here. The prediction model may be a model trained by gradient boosting decision tree (GBDT), fully connected neural network, or the like. Take the prediction model trained by the fully connected neural network as an example, please refer to Figure 7. The process of using the fully connected neural network to process the attribute information is shown in Figure 7. The attribute information of a user to be evaluated is input into the fully connected neural network, and the output of each hidden layer of the fully connected neural network is the same as that of the previous layer. output of each neuron. At this time, in step 24, the above-mentioned predetermined processing may also be performed on the first output result, the second output result and the third output result, so as to determine the financial default risk of the user to be evaluated according to the processing result. Because the user's attribute information has the characteristics of generalization, the user's credibility can be evaluated horizontally, and the evaluation angle is wider. Looking back on the above process, when assessing users' financial default risk, different recurrent neural networks are used to process users' short-term operation sequences and long-term behavior data sets sequences according to the users' longitudinal data, to conduct risk assessments on users, and At least two different processing results are comprehensively processed. In this way, on the one hand, using more in-depth data, on the other hand, even if one result is abnormal, another result can be used to determine the final result. Thus, the effectiveness of determining the user's financial default risk can be improved. In addition, the user's attribute information can be used as horizontal information to evaluate the user's risk, and more data can be used to conduct a comprehensive analysis to improve the accuracy of the evaluation results. According to another embodiment, there is also provided an apparatus for determining a user's risk of financial default. Figure 8 shows a schematic block diagram of an apparatus for determining a user's risk of financial default according to one embodiment. As shown in FIG. 8 , the apparatus 800 for determining a user's financial default risk includes: an acquisition unit 81, which is configured to acquire the short-term operation sequence of the user to be evaluated in the first time period, and the long-term behavior in the second time period Data set sequence, the second time period is greater than the first time period, the short-term operation sequence includes a plurality of pieces of operation information related to the operation behavior of the user to be evaluated, and the long-term behavior data set sequence includes the chronological order. a plurality of behavior data sets, each behavior data set corresponds to a preset third time period respectively, and the behavior data set includes behavior information related to the transaction behavior of the user to be evaluated; the first processing unit 82 is configured to utilize the circulatory neural network The first model of the network is used to process the short-term sequence of operations to obtain the first output result; the second processing unit 83 is configured to use the second model based on the recurrent neural network to process the sequence of long-term behavior data sets to obtain the first output. Two output results; the determination unit 84, which is configured to perform predetermined processing on at least the first output result and the second output result, and determine the financial default risk of the user to be evaluated according to the processing results. In one embodiment, the above-mentioned operation information includes at least one of the following: browsing information, click information, logged-in application program, logged-in device, and geographic location information. In one embodiment, the behavior information includes at least one of the following: transaction time, transaction object, and transaction amount. According to a possible design, the apparatus 800 may further include a third processing unit (not shown) configured to: Obtain the attribute information of the user to be evaluated; The above-mentioned attribute information is processed by the prediction model to obtain a third output result. At this time, the determining unit 84 may be further configured to: The above-mentioned predetermined processing is performed on the first output result, the second output result and the third output result, and the financial default risk of the user to be evaluated is determined according to the processing results. In one embodiment, the above predetermined processing may include at least one of the following: average; take the maximum value; Input a preset logistic regression model as a feature to get logistic regression results. According to one embodiment, the first model/second model comprises a multi-layer stacked long short-term memory model LSTM. In one embodiment, the training samples of the first model/second model include multiple labeled users, and the labeled users have at least pre-labeled credit labels. The credit tag can be determined manually or through the device 800 . When the credit tag is determined through the apparatus 800, the apparatus 800 may further include a marking unit (not shown). For the convenience of description, any one of the multiple labeling users is called the first labeling user, and the labeling unit can determine the first credit label corresponding to the first labeling user in the following manner: Obtain the credit record of the first marked user within a predetermined time period; The first credit tag is determined based on the credit history. In a further embodiment, the marking unit may determine the number of times of trustworthiness and the number of times of untrustworthiness of the first marking user from the above-mentioned credit records; in the case that the ratio of the number of untrustworthy times to the number of times of trustworthiness exceeds a preset ratio threshold, determine the first credit label For untrustworthy users.

In another further embodiment, the labeling unit may detect whether the untrustworthy count of the first labeling user is zero; if the untrustworthy count is non-zero, determine that the first credit label is the untrustworthy user.

It should be noted that the apparatus 800 shown in FIG. 8 is an apparatus embodiment corresponding to the method embodiment shown in FIG. 2 , and the corresponding descriptions in the method embodiment shown in FIG. Repeat.

With the above device, the time series data of the user's short-term and long-term operations are fully utilized, and the final result is determined by synthesizing the output results of multiple models. Thus, the effectiveness of determining the user's financial default risk can be improved.

According to another embodiment, there is also provided a computer-readable storage medium on which a computer program is stored, when the computer program is executed in a computer, the computer is caused to perform the method described in conjunction with FIG. 2 .

According to yet another embodiment, a computing device is also provided, including a memory and a processor, wherein executable program codes are stored in the memory, and when the processor executes the executable program codes, a combination of The method described in Figure 2.

It should be appreciated by those skilled in the art that, in one or more of the above examples, the functions described herein may be implemented in hardware, software, firmware, or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium.

The specific embodiments described above further describe the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the The above are only specific embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solutions of the present invention shall be Included in the protection scope of the present invention.

21: Steps

22: Steps

23: Steps

24: Steps

800: Means for determining a user's risk of financial default

81: Get Unit

82: The first processing unit

83: Second processing unit

84: Determine unit

A: Neurons

C: output

U: weight

W: weight

X: Input layer state

t: time

1-7: Information

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort. FIG. 1 shows a schematic diagram of an implementation scenario of an embodiment disclosed in this specification; Figure 2 illustrates a flow chart of a method of determining a user's risk of financial default according to one embodiment; Fig. 3 shows a schematic diagram of a time sequence state of a neuron of a recurrent neural network; 4 shows a schematic diagram of processing a short-term sequence of operations with a first model based on a recurrent neural network; 5 shows a schematic diagram of processing a sequence of long-term behavioral datasets with a second model based on a recurrent neural network; 6 shows a schematic diagram of performing predetermined processing on the first output result and the second output result; 7 shows a schematic diagram of processing attribute information using a prediction model based on a fully connected neural network; Figure 8 shows a schematic block diagram of an apparatus for determining a user's risk of financial default according to one embodiment.

Claims (20)

A method for determining a user's financial default risk, the method comprising: extracting a short-term operation sequence of the user to be evaluated in a first time period through at least one operation log of the user to be evaluated in a first time period: the user to be evaluated The device currently logged in by the user for evaluation, the respective operation logs corresponding to the different devices logged in by the user to be evaluated during the first time period, and the corresponding operation logs of the different platforms logged in by the user to be evaluated during the first time period An operation log, the short-term operation sequence includes a plurality of pieces of operation information related to the operation behavior of the user to be evaluated, arranged in chronological order, and the operation information includes at least one of the following: browsing information, click information, and logged in applications , logged-in equipment, geographic location information; based on the sorting out of the behavior information of the transaction behavior of the user to be evaluated in the second time period greater than the first time period according to the preset third time period, Extracting the long-term behavior data set sequence of the user to be evaluated in the second time period, wherein each behavior data set in the long-term behavior data set sequence corresponds to each third time period and is arranged in chronological order; using Process the short-term sequence of operations based on a first model of a recurrent neural network to obtain a first output; process the sequence of long-term behavioral datasets with a second model based on a recurrent neural network to obtain a second output ; perform predetermined processing on at least the first output result and the second output result, and determine the financial violation of the user to be evaluated according to the processing result about risk. The method according to claim 1, wherein the behavior information includes at least one of the following: transaction time, transaction object, and transaction amount. The method according to claim 1, wherein the method further comprises: acquiring attribute information of the user to be evaluated; processing the attribute information with a prediction model to obtain a third output result; Performing predetermined processing on the first output result and the second output result, and determining the financial default risk of the user to be evaluated according to the processing result includes: processing the first output result, the second output result and the first output result. 3. The predetermined processing is performed on the output result, and the financial default risk of the user to be evaluated is determined according to the processing result. The method according to claim 1 or 3, wherein the predetermined processing includes at least one of the following: obtaining an average value; taking a maximum value; inputting a preset logistic regression model as a feature to obtain a logistic regression result. The method of claim 1, wherein the first model/the first model The second model includes a multi-layer stacked long short-term memory model LSTM. The method of claim 1, wherein the training samples of the first model/the second model include a plurality of labeling users, and the labeling users at least have pre-labeled credit labels. The method according to claim 6, wherein the plurality of labeling users includes a first labeling user, and the first credit label corresponding to the first labeling user is determined by: a credit record over a time period; the first credit tag is determined based on the credit record. The method according to claim 7, wherein the determining the first credit label based on the credit record comprises: determining the number of times of trustworthiness and the number of times of untrustworthiness of the first label user from the credit record; In the case where the ratio of the number of times of untrustworthiness to the number of times of trustworthiness exceeds a preset ratio threshold, determine the first credit tag as a untrustworthy user. The method of claim 7, wherein the determining the first credit label based on the credit record comprises: detecting whether the number of untrustworthiness of the first marked user is zero; when the number of untrustworthy times is non-zero In this case, it is determined that the first credit tag is an untrustworthy user. An apparatus for determining a user's financial default risk, the apparatus comprising: an acquisition unit configured to extract the short-term short-term value of the user to be evaluated in a first time period through at least one of the following operation logs of the user to be evaluated in a first time period Operation sequence: the device currently logged in by the user to be evaluated, the respective operation logs of different devices logged in by the user to be evaluated within the first time period, the user to be evaluated logging in within the first time period The operation logs corresponding to the different platforms of the APP, and the short-term operation sequence includes a plurality of pieces of operation information related to the operation behavior of the user to be evaluated, arranged in chronological order, and the operation information includes at least one of the following: bangs information , click information, logged-in applications, logged-in devices, and geographic location information; and, based on the user to be evaluated in a second time period greater than the first time period, according to the preset third time period. to sort out the behavior information of the transaction behavior, and extract the long-term behavior data set sequence of the user to be evaluated in the second time period, wherein each behavior data set in the long-term behavior data set sequence corresponds to each third time periods, and are arranged in chronological order; a first processing unit, configured to process the short-term operation sequence by using a first model based on a recurrent neural network, and obtain a first output result; a second processing unit, configured to use a recurrent neural network-based The second model of the network processes the long-term behavior data set sequence to obtain a second output result; the determining unit is configured to perform predetermined processing on at least the first output result and the second output result, and determine according to the processing result The financial default risk of the user to be evaluated. The device according to claim 10, wherein the behavior information includes at least one of the following: transaction time, transaction object, and transaction amount. The device according to claim 10, wherein the device further comprises a third processing unit configured to: acquire attribute information of the user to be evaluated; process the attribute information by using a prediction model to obtain a third output result; The determining unit is further configured to: perform the predetermined processing on the first output result, the second output result and the third output result, and determine the financial default of the user to be evaluated according to the processing results risk. The device according to claim 10 or 12, wherein the predetermined processing includes at least one of the following: obtaining an average value; taking a maximum value; inputting a preset logistic regression model as a feature to obtain a logistic regression result. The apparatus of claim 10, wherein the first model/the second model comprises a multi-layer stacked long short-term memory model LSTM. The apparatus of claim 10, wherein the training samples of the first model/the second model include a plurality of labeled users, and the labeled users have at least pre-labeled credit labels. The device according to claim 15, wherein the plurality of labeling users includes a first labeling user, and the device further comprises a labeling unit configured to determine the first labeling user corresponding to the first labeling user in the following manner Credit label: obtain the credit record of the first marked user within a predetermined period of time; determine the first credit label based on the credit record. The device according to claim 16, wherein the marking unit is further configured to: determine the number of times of trustworthiness and the number of times of untrustworthiness of the first marking user from the credit record; In the case that the ratio of the first credit tag exceeds the preset ratio threshold, it is determined that the first credit tag is a dishonest user. The device according to claim 16, wherein the labeling unit is further configured to: detect whether the number of untrustworthiness of the first labeling user is zero; Credit labels for untrustworthy users. A computer-readable storage medium on which a computer program is stored, and when the computer program is executed in a computer, causes the computer to execute the method of any one of claim 1-9. A computing device, comprising a memory and a processor, characterized in that, executable program codes are stored in the memory, and when the processor executes the executable program codes, any one of claim items 1-9 is implemented. one of the methods described.

Images