Customer screening device, customer screening method, and computer program product

Abstract

The invention provides a customer screening apparatus, a customer screening method, and a computer program product. The customer screening device for performing customer screening based on the game adversarial network comprises a recessive code generation module which generates a recessive code vector based on pre-prepared data of a target customer, and the recessive code vector maps features which are meaningful to judgment of the target customer; the virtual customer information generation module is used for generating virtual customer information based on the implicit coding vector and the random noise vector by utilizing a generative network in the game adversarial network; the game confrontation module is used for discriminating the virtual customer data based on the real customer data by using a confrontation network in the game confrontation network; and the auxiliary network module is used for providing training feedback for the generative network by using an auxiliary neural network according to the implicit coding vector and the virtual customer data so as to improve the correspondence between the generated virtual customer information and the implicit coding vector, and alternately training the generative network and the adversarial network based on a loss function so as to improve the correspondence between the generated virtual customer information and the implicit coding vector. The loss function comprises mutual information operation between a nonlinear difference function of the generative network and the adversarial network and an output function of the generative network and an output function of the auxiliary neural network.

Description

Technical Field

[0001] This invention primarily relates to the field of computer applications, but can also be applied to the financial field, particularly to methods and apparatus for screening customer groups. Background Technology

[0002] In various industries, customer screening is commonly conducted to accurately allocate limited resources, avoid indiscriminate investment, match the needs of different customer groups, improve service acceptance and customer loyalty, and ultimately retain high-value customers, identify potential customers, and maximize overall efficiency. In the financial industry, customer screening is particularly important.

[0003] Currently, target customer screening in the financial industry still relies heavily on manual methods. For example, when a client is established, the manager corresponding to the client's sponsoring institution categorizes the client and assigns them to account managers within their jurisdiction—a manual assignment of client marketing. The main method for this customer categorization currently involves filtering and classifying target customer groups based on data reporting platforms, then managers select target customers in batches from lists provided by the platform or categorize customers based on external transaction information, and then assign account managers. This customer screening method is overly reliant on manual assessment; if managers cannot process or screen customer groups in a timely manner, it may lead to failed customer marketing efforts and ultimately customer churn.

[0004] On the other hand, with the rapid development of artificial intelligence technology, the introduction of large-scale model algorithms to accurately screen target customer groups based on customer characteristic information has become a current research hotspot. In this regard, game-theoretic adversarial networks are a key research direction.

[0005] Generative Adversarial Networks (GANs), also known as Game Theory Adversarial Networks, consist of a generator and an adversary (also called a discriminator). The technical concept of GANs originates from zero-sum games in game theory. The generator is responsible for generating data samples, while the adversary is responsible for determining whether the input samples are real data or fake data generated by the generator. During training, the generator and adversary train together through a game-like interaction. The generator strives to generate more realistic samples to deceive the adversary, while the adversary continuously improves its discriminative ability. The final training result is that the data generated by the generator becomes increasingly realistic, and the adversary's discriminative ability correspondingly increases, until a Nash equilibrium is reached.

[0006] However, the inventors discovered obstacles when applying adversarial networks to customer segmentation. Because the distribution of a specific customer group within the overall customer base (typically, the distribution of customers requiring further development within the long-tail customer group) is highly uneven, the generator cannot perceive the underlying characteristics of the training data during model training. This leads to model convergence difficulties, ultimately resulting in the generator failing to accurately generate data for the target category. Furthermore, when the adversarial network performs discriminative classification, the existence of two imbalanced data categories with completely different distributions and characteristics results in low reliability of the classification results. This is because data classification tends to favor the majority of samples, and after model training, it only roughly determines the data category based on the learned features of the majority of samples, ignoring the unique features of the minority samples. These problems lead to poor screening performance when targeting specific customer types, such as the long-tail customer group.

[0007] The inventors proposed this invention in view of the aforementioned problems. However, it should be noted that the above description is intended to provide background or context for embodiments of the present invention and should not be construed as an admission of prior art simply because it is included in the background section. Summary of the Invention

[0008] The inventors of this invention proposed this invention in view of the problems in the aforementioned background art. In this invention, game-theoretic adversarial networks are applied to customer screening, and the structure of game-theoretic adversarial networks is improved in many ways.

[0009] In one aspect of the present invention, a customer screening device is provided for customer screening based on a game-theoretic adversarial network, comprising: a latent coding generation module that generates a latent coding vector based on pre-prepared target customer data, wherein the generated latent coding vector maps features that are meaningful for judging the target customer in customer screening; a virtual customer information generation module that uses a generator network in the game-theoretic adversarial network to generate virtual customer information based on the latent coding vector generated by the latent coding generation module and a randomly generated noise vector; a game-theoretic adversarial module that uses an adversarial network in the game-theoretic adversarial network to discriminate the virtual customer data generated by the virtual customer information generation module based on input real customer data; and an auxiliary network module that uses an auxiliary neural network to provide training feedback to the generator network based on the latent coding vector output by the latent coding generation module and the virtual customer data generated by the virtual customer information generation module, thereby improving the correspondence between the virtual customer information generated by the generator network and the latent coding vector, wherein the generator network and the adversarial network are trained alternately based on a loss function, wherein the loss function includes mutual information operations between the nonlinear difference function of the generator network and the adversarial network and the output function of the generator network and the output function of the auxiliary neural network.

[0010] As an example, the implicit coding generation module uses a lightweight fully connected neural network to generate implicit coding vectors.

[0011] As an example, the virtual customer information generation module concatenates the implicit coding vector and the noise vector and inputs them into the generation network.

[0012] As an example, the virtual customer information generation module generates and outputs feature vectors as virtual customer information.

[0013] As an example, the training of generative networks and adversarial networks is carried out with the goal of maximizing the results of mutual information computation.

[0014] As an example, the loss function is constructed to capture the semantic information contained in the implicit coding vector.

[0015] On the other hand, the present invention provides a customer screening method based on a game-theoretic adversarial network, comprising the following iterative steps: an implicit encoding generation step, generating an implicit encoding vector based on pre-prepared target customer data, wherein the generated implicit encoding vector maps features meaningful for judging the target customer in customer screening; a virtual customer information generation step, using a generator network in the game-theoretic adversarial network, generating virtual customer information based on the implicit encoding vector generated in the implicit encoding generation step and a randomly generated noise vector; a game-theoretic adversarial step, using an adversarial network in the game-theoretic adversarial network, judging the virtual customer data generated in the virtual customer information generation step based on input real customer data; and an auxiliary network training step, which uses an auxiliary neural network to provide training feedback to the generator network based on the implicit encoding vector generated in the implicit encoding generation step and the virtual customer data generated in the virtual customer information generation step, so as to improve the correspondence between the virtual customer information generated by the generator network and the implicit encoding vector, wherein the generator network and the adversarial network are trained alternately based on a loss function, wherein the loss function includes mutual information operations between the nonlinear difference function of the generator network and the adversarial network and the output function of the generator network and the output function of the auxiliary neural network.

[0016] This loss function can be constructed to capture the semantic information contained in the latent encoded vector. After the game adversarial network is trained, it can be used to filter real customer data. As an example, the generator network and the adversarial network are trained with the goal of maximizing the result of mutual information computation.

[0017] In another aspect, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described customer screening method.

[0018] In another aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described customer screening method.

[0019] In another aspect, the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the above-described customer screening method.

[0020] The proportion of high-potential customers among a bank's long-tail customer base is relatively low, making it difficult to achieve satisfactory results when using general neural network models for training and customer screening. This invention significantly alleviates this sample imbalance problem. When applied to screening a bank's long-tail customer base, this invention automatically extracts deep-level value features, and the trained model can accurately screen for long-tail customers. By using this invention, the customer marketing pressure on frontline bank staff can be significantly reduced, and marketing efficiency can be improved. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0022] Figure 1 A schematic structure of a game-theoretic adversarial network is shown.

[0023] Figure 2 This is a schematic diagram illustrating the training process of a game adversarial network.

[0024] Figure 3 This is a schematic diagram of a game-theoretic adversarial network structure in one embodiment of the present invention.

[0025] Figure 4 This is a schematic structural block diagram of a customer screening device according to one embodiment of the present invention.

[0026] Figure 5 This is a schematic flowchart of a customer screening method in one embodiment of the present invention.

[0027] Figure 6 This is a schematic structural block diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, some embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the illustrative embodiments and descriptions described herein are for explaining the present invention, but are not intended to limit the present invention.

[0029] Furthermore, the following description uses the application of this invention to long-tail customer screening in banks as an example, but this invention can certainly be applied to other scenarios as well.

[0030] "Long-tail customers" refer to the large number of customers within a bank's customer base, but whose individual asset size and business contribution are relatively low. The core characteristics of this group can be summarized as "three more and three less": more people, larger total scale, wider coverage, and fewer assets, lower revenue per customer, and lower service costs per customer. To accurately allocate limited resources to high-potential customers, avoid the cost waste caused by indiscriminate service, and simultaneously tap into the incremental value within this customer group, it is necessary to screen long-tail customers, selecting those who require refined services, ultimately improving customer satisfaction and retention rates.

[0031] In various embodiments of the present invention, game-theoretic adversarial networks are applied to screen specific customer groups for banks, such as long-tail customers. To facilitate understanding of the present invention, a brief explanation of game-theoretic adversarial networks will be provided below.

[0032] Figure 1 The diagram shown illustrates a typical game-theoretic adversarial network. Figure 1 As shown, the adversarial network consists of two parts: the left side is the generator network (also called the generator), and the data input is a noise vector, i.e., randomly generated input data. After this data reaches the input layer, it passes through the neural units in the input layer and the fully connected neural units in the hidden layer to learn the information in the data. After passing through multiple hidden layer neurons, the network learns enough features of the original input data and finally outputs the generated data. Figure 1 The right side of the image represents an adversarial network (also known as an adversarial processor). The input data for an adversarial network includes data generated through iterative training of a generative network and the original real dataset. Generally, an adversarial network consists of input layer neurons, multiple hidden layer neurons, fully connected neurons, and finally, the network is trained to produce its output.

[0033] For game-theoretic adversarial networks, the core training logic can be understood as a game process. For example... Figure 2As shown, the original input data is a noisy variable. After being input into the generator network, it passes through the input layer neurons and various hidden layers. The neural network then learns the features of the input data and performs calculations. After passing through the hidden layers, the entire model stores the potential patterns in the data and finally outputs the generated data. Subsequently, the generated data and the real distribution data are used as the input set for the adversarial network. After being processed by the adversarial network, the final data label is output. The entire model is continuously iterated and optimized so that the generated data sufficiently matches the distribution of the original data, and the adversarial network can maximize the distinction between generated and real data. Finally, when the model converges, the generated data is almost identical to the real data, and the classification effect of the adversarial network reaches its optimal level.

[0034] However, the inventors discovered that because the target customers (i.e., customers that need to be developed) in the long-tail customer group are a small proportion and their distribution is very uneven, when applying game adversarial networks to screen long-tail customers, when the input noisy data is difficult to control, it often requires a lot of resources in model training. The data type of the generated data is also uncontrollable. If training data for imbalanced samples is input, and the generator network simultaneously simulates and generates data that is almost identical to the distribution of real data, it will lead to more severe data skew. Obviously, the classification effect of the adversarial network does not meet the requirements.

[0035] This invention addresses the aforementioned problems. In one embodiment, when constructing a game-theoretic adversarial network for long-tail customer screening, the input data structure of the adversarial network is improved. The input data of the generator network includes a random noise vector and a latent encoding vector that maps semantic features to the characteristics of the target customer. An auxiliary neural network is also added to train the similarity between the latent semantic encoding and the virtual customer data generated by the generator network. Based on this, the adversarial network is improved so that the loss function can capture the latent encoding, thereby enabling the model to learn the semantic feature information of the input.

[0036] The final network architecture is as follows: Figure 3 As shown. In brief, the generative network generates virtual customer data based on implicit encoding and noise vectors. An auxiliary neural network (also referred to as the auxiliary network below) verifies the semantic matching degree between the generated customer data and the implicit encoding, constraining the generation direction of the generative network. An adversarial network receives real customer data and the virtual customer data generated by the generative network, judges their authenticity, and forms a game-like adversarial relationship with the generative network. This process iterates repeatedly, ultimately allowing the generative network to generate samples that closely resemble real customer characteristics and conform to the target semantic features under the dual constraints of "deceiving the adversarial network" and "satisfying the matching degree of the auxiliary network." Simultaneously, the adversarial network achieves optimal discriminative ability in the game-like adversarial process.

[0037] The present invention will now be described in more detail with reference to some embodiments. The present invention can be implemented in many different forms, including customer screening devices, customer screening methods, computer equipment, computer storage media, computer program products, etc. It is understood that the description of one embodiment below is also applicable to other embodiments, and the descriptions of various embodiments can also be combined with each other.

[0038] Example 1

[0039] Figure 4 A schematic block diagram of a customer screening device according to an embodiment of the present invention is shown. Figure 4 As shown, the customer screening device in this embodiment includes: an implicit coding generation module 401, a virtual customer information generation module 402, a game-theoretic module 403, and an auxiliary network module 404. Each module can be implemented by running appropriate programs on an information processing device such as a personal computer or server, and will not be described in detail here.

[0040] The implicit coding generation module 401 is configured to generate the implicit coding vector (also sometimes referred to as implicit coding) in this invention. Implicit coding is a low-dimensional dense vector obtained by encoding the semantic features of the target customer (i.e., long-tail customers deemed to have potential value) and is used to transmit the core attribute constraints of the target customer to the generator and adversary. Although implicit coding reflects or relates to the characteristics of the target customer, it does not directly correspond to the original customer data. Instead, it retains key features and masks redundant information through compression mapping.

[0041] Specifically, the implicit coding generation module 401 generates implicit codes based on pre-prepared original customer data. The original customer data will be described below.

[0042] The raw customer data can be extracted from internal legitimate business data and publicly available related data, focusing on target customers deemed to have development value. In other words, it is multi-dimensional business data strongly correlated with the long-tail customer screening target. When collecting customer data, dimensions such as customer characteristics, screening rules, and business scenarios should be covered. For example, data as shown in Table 1 below can be collected:

[0043]

[0044] Table 1

[0045] First, core features strongly correlated with development value are extracted from raw customer data. These features are then quantified and normalized to obtain a value feature vector. For example, as a mere illustration, five core features are extracted and, through quantization and normalization, a five-dimensional value feature vector is obtained:

[0046] Potential Lifetime Value (Potential LTV): Total revenue over the next 3-5 years predicted based on customer historical transaction data and product usage behavior (normalized to the [0, 1] range).

[0047] Service cost-benefit ratio: Customer's projected revenue for the next year ÷ service costs (such as operating costs and risk control costs).

[0048] Cross-purchase probability: The probability of new business conversion predicted based on the customer's historical product usage records (e.g., whether they are likely to purchase foreign exchange wealth management products if they have already used cross-border payments);

[0049] Transaction frequency dispersion: Number of transactions in the past 12 months ÷ Average number of transactions of top clients in the industry during the same period;

[0050] Business scenario fit: The degree to which the customer's business matches the target scenario (e.g., the proportion of cross-border transactions in cross-border business scenarios, and the proportion of online transactions in fintech scenarios).

[0051] Finally, this value feature vector can be in the form of [0.9, 0.85, 0.8, 1.0, 0.82]. This feature vector reflects the attributes of target customers with high development value.

[0052] In this specification, such a value feature vector may also be referred to as customer data. The implicit coding generation module 401 receives such a value feature vector. As an example, the implicit coding generation module 401 may have a lightweight fully connected neural network; for example, corresponding to a 5-dimensional value feature vector, the neural network's input layer may have 5 neurons, the hidden layer may have 16 neurons (e.g., a ReLU activation function), and the output layer may have 16 neurons (e.g., a Sigmoid activation function). Through this fully connected neural network, an implicit coding vector is output based on the value feature vector. This implicit coding vector, for example, is 16-dimensional and carries, or maps, features associated with the target customer.

[0053] The above example illustrates the generation of implicit coding vectors from raw customer data using a lightweight fully connected neural network. However, this invention is not limited to this; other methods can also be used to generate the implicit coding vectors used in the game-adversarial network of this embodiment based on customer data. Furthermore, the extracted customer information items described above are merely illustrative examples; various customer information can be extracted according to specific business scenarios. The key here is mapping the features of target customers (i.e., customers with development value in the long-tail customer group) into this implicit coding vector.

[0054] Alternatively, during the use of the customer screening device in this embodiment, the implicit coding vector can be continuously updated based on the customer data of new target customers according to the processing described above.

[0055] The virtual customer information generation module 402 includes a generator network in a game adversarial network. The virtual customer information generation module 402 receives a latent coding vector from the latent coding generation module 401, generates a random noise vector, and concatenates these two vectors before inputting them into the generator network.

[0056] Specifically, the input to the generator network is a multidimensional vector. In this embodiment, this multidimensional vector consists of two parts: a random noise vector and a latent encoding vector that maps semantic features (business features that are meaningful for determining whether a customer is a target customer).

[0057] As an example, the random noise vector is a 30-dimensional vector, where each dimension represents a variable value distributed in the interval [0,1]. Each dimension of the random noise vector has no explicit semantic meaning and does not correspond to any specific business features (such as deposit amount, customer tags, etc.). Its purpose is simply to inject irregular diversity into the generated samples, preventing the generated results from becoming monotonous due to over-reliance on the semantic control of implicit coding. Specifically, each dimension of the random noise vector is a random value in the interval [0,1]. These values ​​themselves have no practical meaning and exist only as noise signals. For example, for two identical implicit codes (corresponding to features of the same type of target customer), when paired with different random noise vectors, the generator network will output samples with similar features but different details. For instance, both are high-deposit VIP customers, but one may have purchased financial product A, while the other may have purchased financial product B. In addition, the random noise vector can also avoid overfitting to semantic constraints. If the generated samples are entirely controlled by the semantic dimension of implicit encoding, the model may mechanically replicate the feature combinations corresponding to the implicit encoding, resulting in highly similar generated samples. For example, all "high-deposit VIP" customers might have the same age and spending frequency, leading to a loss of the natural diversity of real data. Introducing uncontrollable subtle fluctuations through random noise vectors makes the generated samples closer to the randomness of real data, just as the behavior of real customers naturally exhibits irregular differences.

[0058] In this example, the virtual customer information generation module 402 concatenates the implicit encoded vector and the random noise vector as input to the generator network in the game adversarial network. In this invention, the implicit encoded vector can be considered a controllable skeleton, mapped to semantic features to determine the core type and key features of the sample, while the random noise vector is the random flesh and blood, supplementing subtle differences and ensuring sample diversity. In this invention, by combining the implicit encoded vector and the random noise vector as described above, virtual target customer data with controllable types and closely approximating real-world distributions can be generated.

[0059] As a specific example, after the generator network receives the concatenated vector (latently encoded vector + random noise vector), it transforms the low-dimensional vector into virtual customer data with real business characteristics through the mapping and decoding process of a multi-layer neural network.

[0060] The first step is vector mapping. Fully connected or convolutional layers in the generator network perform non-linear transformations on the concatenated vectors, mapping the original low-dimensional input vectors to a higher-dimensional intermediate feature space. For example, the first fully connected layer transforms a 35-dimensional input vector into a 256-dimensional vector, and then an activation function (such as ReLU) is used to obtain 256-dimensional intermediate features. This amplification of feature dimensions reserves expression space for subsequent generation of multi-dimensional customer features (such as deposit amount, product preferences, and tag attributes).

[0061] Next is feature differentiation, which separates controllable features from random features. Specifically, the intermediate network differentiates features based on the different functions of the implicit encoding and random noise in the input vector. For controllable features, the focus is on the semantic information carried by the implicit encoding, such as "high deposit amount" and "VIP label," and these features are enhanced through specific network layers, such as attention mechanism layers. For example, from the 256-dimensional intermediate features, sub-features related to deposit amount (such as continuous values ​​mapped to [500,000, 2,000,000]) and sub-features related to "VIP label" (such as discrete values ​​mapped to 1 / 0) are extracted. For random features, the perturbation information of random noise is used to add personalized details to the controllable features, avoiding completely identical generated samples. For example, under the same implicit encoding of "high deposit VIP," random noise will cause the generated "number of financial products held" to fluctuate randomly within the range of [3, 8], and the "recent transaction time" to be randomly selected within the last 3 months.

[0062] At this point, the virtual customer information generation module 402 has generated virtual customer data. This virtual customer data is a vector, and the virtual customer information generation module 402 outputs this vector to the game-playing adversarial module 403.

[0063] If needed, the virtual customer data vector can also be decoded into structured customer data. This can be done at the top layer of the network, such as the output layer, to convert the virtual customer data vector into virtual customer data conforming to the business format. For example, such virtual customer data can contain two types of features: continuous features, such as deposit amount (outputting a specific value, e.g., 1.285 million yuan), average monthly transaction frequency (e.g., 15 times), which are mapped to the numerical range of real business transactions through a linear layer; and discrete features, such as customer tags (VIP=1 / 0), main held products ("fixed deposit + fund", "wealth management + insurance", etc.), which are output as classification probabilities through a softmax layer and then converted into specific tags.

[0064] In this embodiment, by introducing semantic features corresponding to the characteristics of the target customer into the input of the generator network, the input vector contains both controllable implicitly encoded information and random noise information. The generator network then performs multi-layer nonlinear transformations to convert the data into customer data with realistic distribution characteristics. The implicitly encoded vector ensures that the generated data conforms to the core characteristics of the target customer group, while the random noise vector ensures sufficient detail and diversity in the data. The combination of these two elements allows the generated virtual customer data to both alleviate sample imbalance and closely resemble real business scenarios.

[0065] The adversarial module 403 receives virtual customer data as vectors from the virtual customer information generation module 402, and also receives real customer data (training data). These two data points are used as input to the adversarial network, which then determines whether the input data is real customer data or generated virtual customer data. Each virtual customer data vector can be assigned a pre-set label of 0.

[0066] An adversarial network can be considered a binary classifier based on customer features. Its working process can be roughly divided into four closed-loop steps: input layer feature reception, hidden layer feature extraction, output layer probability determination, loss calculation and parameter optimization. Throughout the process, it dynamically plays against the generator network, ultimately achieving the dual goals of accurately distinguishing between real customer data and virtual customer data and forcing the generation of high-quality virtual customer data.

[0067] Specifically, the input layer of the adversarial network receives the aforementioned real and virtual customer data, and converts them into tensors that the network can compute. At this point, the input layer is only responsible for feature propagation and does not change the feature dimensions.

[0068] Hidden layers, through multiple convolutional / fully connected layers and activation functions, gradually capture the subtle differences between real and virtual customer data from the original features. For example, shallow hidden layers map the input vector to a higher dimension through fully connected layers and use the ReLU activation function to extract common customer features such as "average order value" and "purchase frequency". Deeper hidden layers further enhance the feature abstraction through fully connected layers (e.g., 128-dimensional → 256-dimensional → 64-dimensional), add batch normalization to alleviate overfitting, and adapt to the data sparsity of long-tail customers. For example, they extract features unique to long-tail customers such as "niche category purchase preferences + purchase interval patterns" and "feature logic loopholes in the generation of virtual customer data". For example, the generative network sometimes mistakenly binds "high average order value" with "high frequency of purchase", while real target customers are mostly "high average order value + low frequency of purchase". Through multiple iterations, the original customer data is transformed into a high-order feature vector that can distinguish between real and fake data. For example, the high-order features of real target customers may be a logical combination of "niche categories → low frequency → high average order value", while the virtual customer data generated in the early stage may lack this logical consistency. The hidden layer will amplify this difference.

[0069] The output layer performs binary classification probability determination. For example, based on the high-order feature vector output by the hidden layer, the output layer maps the feature vector to a single probability value P (P∈[0,1]) through the Sigmoid function.

[0070] If P ≥ 0.5, the adversarial network considers the input data to be "real customer data" and sets the label to 1;

[0071] If P < 0.5, the adversarial network considers the input data to be "virtual customer data" and sets the label to 0.

[0072] Adversarial networks (ANNs) do not output results all at once. Instead, they continuously optimize their discriminative capabilities through a closed loop of "loss calculation → parameter adjustment," while simultaneously driving the generative network to improve the quality of virtual customer data. Specifically, the gradient of the loss function with respect to the parameters (weights, biases) of each layer of the ANN is calculated using the gradient descent algorithm, and then the parameters are updated along the negative direction of the gradient. During parameter updates, an attention mechanism can be used to appropriately weight the weights corresponding to the features of the target customer, ensuring that the ANN prioritizes the differentiated features of the target customer and avoids ignoring the features of the target customer due to the high proportion of mainstream customer data. Moreover, the ANN and the generative network are trained alternately, forming the following closed loop.

[0073] Step 1: Fix the parameters of the generator network and train the adversarial network. That is, input real customer data and real customer data into the adversarial network, calculate the loss, and update the parameters of the adversarial network until the discrimination accuracy of the adversarial network reaches a preset threshold (e.g., 80%).

[0074] Step 2: Fix the parameters of the adversarial network and train the generator network. As mentioned above, the generator network receives the implicit coding vector and the random noise vector, generates virtual customer data, and inputs it into the adversarial network to obtain the prediction probability. The parameters of the generator network are then updated through the loss function.

[0075] Repeat steps 1-2 above until Nash equilibrium is reached, meaning the adversarial network can no longer distinguish between real and virtual customer data (the discrimination accuracy is approximately 50%). At this point, the adversarial network's feature extraction capability is optimal, accurately capturing the core features of the target customer, while the distribution of the virtual customer data generated by the generator network is consistent with that of the real customer data. In the early stages of training, the virtual customer data generated by the generator network may differ significantly from the real customer data, and the P-value of the output layer will be significantly biased towards 0. As the adversarial network iterates, the quality of the virtual customer data gradually improves, and the P-value will gradually approach 0.5, thus reaching Nash equilibrium.

[0076] The customer screening device in this embodiment also includes an auxiliary network module 404, which has an auxiliary neural network. The auxiliary network module 404 receives a latent coding vector from the latent coding generation module 401 and a virtual customer data vector from the virtual customer information generation module 402, and uses these two as inputs to the auxiliary neural network.

[0077] The inventors of this invention recognized that in screening long-tail customers, relying solely on the game between generator networks and adversarial networks can sometimes lead to the problem that "the surface characteristics of virtual customer data resemble real data, but lack the implicit characteristics of real target customers." For example, virtual customer data may mimic the surface characteristic of "low consumption frequency" but fail to capture the implicit characteristic of "niche product preferences," ultimately resulting in insufficient accuracy in the selected customers.

[0078] Therefore, in this embodiment, an auxiliary network module 404 is set in the customer screening device. An auxiliary neural network is added to the game-theoretic network described above to train the similarity between the implicit semantic encoding and the virtual customer data generated by the generator network. This allows the generator network to learn the deep semantic features of real target customers, avoiding the generation of virtual customer data that is superficially valid but ineffective at a deeper level. By strengthening the generator network's capture of the aforementioned implicit semantic features through the auxiliary neural network, the generated virtual customer data is made to better match the deep features of real customers, rather than merely remaining at the level of surface feature similarity, thereby improving the accuracy of long-tail customer screening. In this embodiment, this auxiliary neural network does not participate in the authenticity judgment of customer data, but focuses on similarity learning. The training objective of this auxiliary neural network is to minimize the distance between the input implicit semantic encoding and the virtual customer data. Essentially, it enables the generator network to learn to generate virtual customer data based on implicit semantic encoding, rather than merely imitating the surface distribution of real customer data.

[0079] Specifically, the implicit encoding vector and the virtual customer data vector generated by the generator network are input to the auxiliary network, and the encoding layer of the auxiliary network calculates the semantic encoding of the virtual customer data. Furthermore, by defining a loss function, the difference between the input implicit semantic encoding and the calculated semantic encoding of the virtual customer data is calculated. The parameters of the auxiliary network are updated through backpropagation, and the loss signal is fed back to the generator network.

[0080] Ultimately, during training, the generative network must not only fool the adversarial network (i.e., have realistic surface features) but also meet the similarity requirements (deep semantic fit) of the auxiliary network, so that the generated virtual customer data not only matches the surface features of real customers but also possesses the deep features of target customers.

[0081] Furthermore, in this embodiment, based on the network architecture described above, implemented by the virtual customer information generation module 402, the game-playing adversarial module 403, and the auxiliary network module 404, a special design was made for the loss function. The inventors recognized that when using game-playing adversarial networks for screening long-tail customers, conventionally known loss functions cannot capture latent semantic features and can lead to problems such as gradient explosion or overfitting. Therefore, a loss function specifically suitable for long-tail customer screening in this embodiment and capable of learning input semantic feature information was proposed, namely, a customer feature standardization function based on mutual information. The design concept of the loss function Loss in this embodiment can be represented by Equation 1 below.

[0082] …(Formula 1)

[0083] In Equation 1, G represents the generator network, D represents the adversarial network, Q represents the auxiliary network, c represents the implicit coding vector input to the generator network, n represents the noise vector input to the generator network, and x represents real customer data.

[0084] In Equation 1, minmax represents the minimax game, which is also the core optimization logic of the generative adversarial network model. The nonlinear difference function representing the two parts, generator network G and adversarial network D, represents the game loss of adversarial network D and generator network G under input data x. In this embodiment, The entities that minimize this loss are the generator network G and the auxiliary network Q; that is, the generator network G and the auxiliary network Q work together to adjust their own parameters and jointly minimize the loss. Minimize. Furthermore, this also implicitly means that the entity maximizing is the adversarial network D, meaning that the adversarial network D will adjust its own parameters to minimize the loss. maximize.

[0085] In Equation 1, as described above This represents the game loss between the generator network G and the adversarial network D under the input data x. This represents the output function for generating customer data from the network. This represents the output function of the auxiliary neural network that controls the generation of samples and the loss of the input. I() represents the mutual information function, which utilizes mutual information theory to... and , In this context, the nonlinear difference function of the generative adversarial mechanism is used to calculate mutual information terms between the data generated by the generative network, maximizing their correlation. In this embodiment, the loss function of the entire network structure is designed as described above to achieve accurate screening of the long-tail customer group.

[0086] Specifically, in Equation 1 The calculation is as follows:

[0087]

[0088] That is, It is the joint loss function of the generator network G and the auxiliary network Q. Through the mechanism of "distribution fitting + entropy constraint", it maximizes the correlation between the latent coding vector c and the real customer data x. For auxiliary neural networks The logarithm of the output conditional probability distribution is used in this embodiment to make the probability distribution fitting process smoother and easier to converge, because the logarithmic function can transform the product into a summation, simplifying gradient calculation. This represents the probability that, given customer data x, the implicit encoding is c. Additionally, Entropy, representing the latent code c, is used to measure the uncertainty of the latent code c. In this embodiment, [the following is added] This is to avoid the auxiliary network Q from overfitting a single type of implicit code c, and to ensure that the generated customer data covers more types of long-tail customers.

[0089] In addition, in Equation 1 The calculation is as follows:

[0090]

[0091] This formula means that, taking semantic features n and implicit codes c as input, the generator network outputs virtual customer information, and the loss function of this generation process is... A logarithmic approach is used to ensure that the generated virtual customer information highly matches the target in terms of semantic features and actual attributes. Input variables The distribution output corresponding to (i.e., the hidden coding vector and the random noise vector), where y is the generator network. The output.

[0092] In Equation 1 The expression is as follows.

[0093]

[0094] The data distribution representing the input noise, This indicates the distribution of the input long-tail customer data. This represents the output of data generated by simulating the distribution of long-tail customer data. The input represents the real long-tail data, and E represents the information entropy value.

[0095] In this embodiment, the system loss function is designed as described above. Specifically, in this embodiment, the mutual information term is calculated using the mutual information function on the nonlinear difference function between the generator network G and the adversarial network D, the output function of the generator network G, and the output function of the auxiliary network. The model is trained with the goal of maximizing this mutual information term. The model training reaches... After the function converges, the virtual customer data generated by the generator network is both realistic, satisfying the game requirements of the game adversarial network, and accurate, conforming to the essential characteristics of the actual target customers.

[0096] The operation of the above-described customer screening device based on a game-theoretic adversarial network will now be described in more detail with reference to the customer screening method in Embodiment 2. Furthermore, the description of the customer screening method in Embodiment 2 below also applies to Embodiment 1.

[0097] Example 2

[0098] In Example 2, the present invention is implemented as a method for screening customers based on game-theoretic adversarial networks. The customer screening method in Example 2 generally includes a model training phase and an actual screening phase. The following description focuses on the model training process to explain the customer screening method of this example.

[0099] Before proceeding, the training and test sets used in the training will be explained. These sets can be collected at any time before model training, or they can be datasets prepared or obtained in advance through other means. The acquisition of this data itself does not constitute part of the customer screening method of this embodiment; it is only described below for the purpose of understanding the invention.

[0100] Data can be collected from legitimate internal bank business data or publicly available relevant data. The key to data collection is extracting business data that is highly relevant to the long-tail customer screening target. Details are the same as explained in Table 1 above and will not be repeated here. Additionally, corresponding marketing feedback from account managers should be collected, such as marketing success / failure and customer responsiveness, and converted into labels corresponding to the data entries. This label can be a binary value, for example, 0 or 1, where 0 represents "non-target customer" and 1 represents "target customer". The principle of data collection here is to ensure that the data covers both customer characteristics and marketing results, providing a basis for the subsequent model to learn "which customers are worth marketing to".

[0101] Next, data visualization can be performed to make a preliminary assessment of data quality. For example, histograms, box plots, and heatmaps of missing values ​​can be used to visualize the data. This allows for a direct visual confirmation of the overall data distribution, such as the numerical range of features and the proportion of each category. It also allows for the direct identification of missing data and outliers.

[0102] Next, data cleaning can be performed. For example, missing value handling can be done, manually filling in missing information based on human experience. For instance, based on business experience, missing "account opening duration" for "high-quality new customers" could be filled with 3 months. For continuous features without clear business rules, such as deposit amount, the median value of the data distribution can be used to fill in the missing data. In short, while preserving business logic, statistical values ​​are used to ensure data integrity. Outlier handling can also be performed, removing data with abnormal distributions. For example, a sample with a deposit amount of 1 billion yuan far exceeds the normal customer range and can be identified as an anomaly and deleted. This processing is to avoid outliers interfering with the model's learning of normal customer characteristics, especially since this involves long-tail customers, where outliers may obscure their true distribution.

[0103] Additionally, discontinuous data can be transformed by normalizing various user characteristics into numerical data. For example, discrete categorical features such as "age stratification" and "product type" can be converted into numerical values ​​using one-hot encoding or label encoding. For instance, "35-45 years old" → 1, "45-55 years old" → 2. For continuous features such as deposit amount, normalization to [0,1] or standardization (mean 0, variance 1) is used to avoid features of different magnitudes affecting model weights. The data after this processing corresponds to the customer value vector described in Example 1. The vector form meets the model input requirements because the neural network model only accepts numerical input. Thus, as an example, the final dataset contains multiple 5-dimensional vectors, and each vector has an associated label.

[0104] After collecting the customer dataset as described above, or obtaining such a dataset through other means, the dataset can be divided into a training set and a test set. The training set is used for model training, while the test set can be used to simulate real-world scenarios to evaluate the model's training performance and generalization ability, preventing the model from simply memorizing training data and failing to identify new customers. In one example, a 7:3 ratio is used, with 70% of the data used as the training set and the remaining 30% as the test set. Of course, any other ratio can be used to divide the collected data into training and test sets.

[0105] Figure 5 The outline of the customer screening method in Embodiment 2 is shown below. Figure 5 The steps of this customer screening method are explained in detail.

[0106] In step 510, a latent coding vector is generated based on the training data, for example using a lightweight fully connected neural network. This latent coding vector maps features associated with the attributes of the target customer. The specific generation process can be found in the description of the latent coding generation module 401 in Embodiment 1 above, and will not be repeated here.

[0107] In step 520, the implicit encoding vector generated in step 510 is received, and a random noise vector is generated. These two vectors are concatenated and input into the generator network in the game adversarial network. Specifically, the input to the generator network is a multidimensional vector. In this embodiment, this multidimensional vector consists of two parts: a random noise vector and an implicit encoding vector that maps semantic features (features that are meaningful for determining whether a customer is a target customer).

[0108] After the generated network receives the spliced ​​vector (latently encoded vector + random noise vector), the low-dimensional vector is transformed into virtual customer data with real business characteristics through the mapping and decoding process of a multi-layer neural network.

[0109] The first step is vector mapping. Fully connected or convolutional layers in the generator network perform non-linear transformations on the concatenated vectors, mapping the original low-dimensional input vectors to a higher-dimensional intermediate feature space. For example, the first fully connected layer transforms a 35-dimensional input vector into a 256-dimensional vector, and then an activation function (such as ReLU) is used to obtain 256-dimensional intermediate features. This amplification of feature dimensions reserves expression space for subsequent generation of multi-dimensional customer features (such as deposit amount, product preferences, and tag attributes).

[0110] Next is feature differentiation, which separates controllable features from random features. Specifically, the intermediate network differentiates features based on the different functions of the implicit encoding and random noise in the input vector. For controllable features, the focus is on the semantic information carried by the implicit encoding, such as "high deposit amount" and "VIP label," and these features are enhanced through specific network layers, such as attention mechanism layers. For example, from the 256-dimensional intermediate features, sub-features related to deposit amount (such as continuous values ​​mapped to [500,000, 2,000,000]) and sub-features related to "VIP label" (such as discrete values ​​mapped to 1 / 0) are extracted. For random features, the perturbation information of random noise is used to add personalized details to the controllable features, avoiding completely identical generated samples. For example, under the same implicit encoding of "high deposit VIP," random noise will cause the generated "number of financial products held" to fluctuate randomly within the range of [3, 8], and the "recent transaction time" to be randomly selected within the last 3 months.

[0111] The specific processing in step 520 can be referred to the description of the virtual customer information generation module 402 in Embodiment 1. This generates virtual customer data as a vector. This virtual customer data is then output to the auxiliary network and the adversarial network.

[0112] Step 540 corresponds to the processing performed by the auxiliary network module 404 in Embodiment 1. In step 540, the similarity between the implicit semantic encoding and the virtual customer data generated by the generator network is trained, allowing the generator network to learn the deep semantic features of real target customers and avoid generating virtual customer data that is superficially qualified but ineffective at a deeper level. By strengthening the generator network's capture of the aforementioned implicit semantic features through the auxiliary neural network, the generated virtual customer data is made to better match the deep features of real customers, rather than merely remaining at the level of superficial feature similarity, thereby improving the accuracy of long-tail customer group screening.

[0113] In step 540, the implicit coding vector input to the generator network and the virtual customer data vector generated by the generator network are input to the auxiliary network, and the coding layer of the auxiliary network calculates the semantic code of the virtual customer data. Furthermore, using the loss function described in Example 1, the difference between the input implicit semantic code and the calculated semantic code of the virtual customer data is calculated, and the parameters of the auxiliary network are updated through backpropagation, while the loss signal is fed back to the generator network.

[0114] Ultimately, during training, the generative network must not only fool the adversarial network (i.e., have realistic surface features) but also meet the similarity requirements (deep semantic fit) of the auxiliary network, so that the generated virtual customer data not only matches the surface features of real customers but also possesses the deep features of target customers.

[0115] In step 530, the virtual customer data generated in step 520 and the real customer data in the training dataset are input into the adversarial network in the game adversarial network, and the adversarial network determines whether the input data is real customer data or generated virtual customer data.

[0116] Adversarial networks (ANNs) do not output results all at once. Instead, they continuously optimize their discriminative capabilities through a closed loop of "loss calculation → parameter adjustment," while simultaneously driving the generative network to improve the quality of virtual customer data. Specifically, the gradient of the loss function with respect to the parameters (weights, biases) of each layer of the ANN is calculated using the gradient descent algorithm, and then the parameters are updated along the negative direction of the gradient. During parameter updates, an attention mechanism can be used to appropriately weight the weights corresponding to the features of the target customer, ensuring that the ANN prioritizes the differentiated features of the target customer and avoids ignoring the features of the target customer due to the high proportion of mainstream customer data. Moreover, the ANN and the generative network are trained alternately, forming the following closed loop.

[0117] Step 1: Fix the parameters of the generator network and train the adversarial network. That is, input real customer data and real customer data into the adversarial network, calculate the loss, and update the parameters of the adversarial network until the discrimination accuracy of the adversarial network reaches a preset threshold (e.g., 80%).

[0118] Step 2: Fix the parameters of the adversarial network and train the generator network. As mentioned above, the generator network receives the implicit coding vector and the random noise vector, generates virtual customer data, and inputs it into the adversarial network to obtain the prediction probability. The parameters of the generator network are then updated through the loss function.

[0119] Repeat steps 1-2 above until Nash equilibrium is reached, meaning the adversarial network can no longer distinguish between real and virtual customer data (the discrimination accuracy is approximately 50%). At this point, the adversarial network's feature extraction capability is optimal, accurately capturing the core features of the target customer, while the distribution of the virtual customer data generated by the generator network is consistent with that of the real customer data. In the early stages of training, the virtual customer data generated by the generator network may differ significantly from the real customer data, and the P-value of the output layer will be significantly biased towards 0. As the adversarial network iterates, the quality of the virtual customer data gradually improves, and the P-value will gradually approach 0.5, thus reaching Nash equilibrium.

[0120] It is worth mentioning that, as described in Example 1, the loss function used in the aforementioned game adversarial network is specially designed. It calculates the mutual information term by applying a mutual information function to the nonlinear difference function between the generator network and the adversarial network, the output function of the generator network, and the output function of the auxiliary network. The model is trained with the goal of maximizing this mutual information term.

[0121] In addition, after the model training is completed, the aforementioned test data can be used to test the long-tail customer screening effect of the entire system. When the screening accuracy of the test data reaches a preset threshold, the model can be put into actual use. The customer screening method in this embodiment proceeds to step 550 to perform actual customer screening.

[0122] In step 550, the actual long-tail customer data is processed using the same processing and format as the aforementioned training dataset to obtain a long-tail customer data vector. This long-tail customer data vector is then input into an adversarial network, which outputs the probability that the long-tail customer data vector corresponds to a target customer. For example, long-tail customers whose probability value is greater than a preset threshold can be identified as target customers with marketing value.

[0123] Example 3

[0124] This invention can also be implemented as a computer device. Figure 6 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention, such as... Figure 6 As shown, the computer device 600 includes a memory 610, a processor 620, and a bus 630. The processor 620 and the memory 610 communicate with each other via the bus 630. The memory 610 stores a computer program 611, and the processor 620 calls the computer program 611 stored in the memory 610. The computer program 611 is programmed to implement the various steps of the customer screening method described above when executed by the processor 620.

[0125] The present invention can also be implemented as a computer-readable storage medium storing a computer program that, when executed by a processor, implements the various steps of the customer screening method described above.

[0126] The present invention can also be implemented as a computer program product, which includes a computer program that, when executed by a processor, implements the various steps of the customer screening method described above.

[0127] The foregoing has provided a detailed description of some embodiments of the present invention. Those skilled in the art should understand from the foregoing description that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0128] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0129] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0130] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0131] The information collected in the technical solution of this application is information and data authorized by the user or fully authorized by all parties. The collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with the relevant laws, regulations and standards of the relevant countries and regions, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation portals are provided for users to choose to authorize or refuse.

[0132] The acquisition, transmission, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0133] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0134] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A customer screening device for performing customer screening based on a game-theoretic network, comprising: The implicit coding generation module generates implicit coding vectors based on pre-prepared target customer data. These generated implicit coding vectors map features that are meaningful for judging target customers in customer screening. The virtual customer information generation module utilizes the generator network in the game adversarial network to generate virtual customer information based on the implicit coding vector generated by the implicit coding generation module and the randomly generated noise vector. The game-theoretic adversarial module utilizes the adversarial network within the game-theoretic adversarial network to evaluate the virtual customer data generated by the virtual customer information generation module based on the input real customer data; and The auxiliary network module utilizes an auxiliary neural network to provide training feedback to the generator network based on the implicit coding vector output by the implicit coding generation module and the virtual customer data generated by the virtual customer information generation module. This improves the correspondence between the virtual customer information generated by the generator network and the implicit coding vector. Specifically, the generator network and the adversarial network are trained alternately based on a loss function, wherein the loss function includes mutual information operations between the nonlinear difference function of the generator network and the adversarial network, the output function of the generator network, and the output function of the auxiliary neural network.

2. The customer screening device according to claim 1, wherein, The implicit coding generation module uses a lightweight fully connected neural network to generate the implicit coding vector.

3. The customer screening device according to claim 1, wherein, The virtual customer information generation module concatenates the implicit coding vector with the noise vector and inputs the result into the generation network.

4. The customer screening device according to claim 1, wherein, The virtual customer information generation module generates and outputs feature vectors as the virtual customer information.

5. The customer screening device according to any one of claims 1-4, wherein, The generator network and the adversarial network are trained with the goal of maximizing the result of the mutual information operation.

6. The customer screening device according to claim 5, wherein, The loss function is constructed to capture the semantic information contained in the implicit coding vector.

7. A customer screening method based on a game-theoretic adversarial network, comprising the following iteratively executed steps: The implicit coding generation step generates an implicit coding vector based on pre-prepared target customer data. This generated implicit coding vector maps features that are meaningful for judging target customers in customer screening. The virtual customer information generation step utilizes the generator network in the game adversarial network to generate virtual customer information based on the implicit coding vector generated in the implicit coding generation step and the randomly generated noise vector. The game-playing adversarial step utilizes the adversarial network within the game-playing adversarial network to evaluate the virtual customer data generated in the virtual customer information generation step based on the input real customer data; and The auxiliary network training step utilizes an auxiliary neural network to provide training feedback to the generator network based on the implicit coding vector generated in the implicit coding generation step and the virtual customer data generated in the virtual customer information generation step. This improves the correspondence between the virtual customer information generated by the generator network and the implicit coding vector. Specifically, the generator network and the adversarial network are trained alternately based on a loss function, wherein the loss function includes mutual information operations between the nonlinear difference function of the generator network and the adversarial network, the output function of the generator network, and the output function of the auxiliary neural network.

8. The customer screening method according to claim 7 further includes: After the game adversarial network is trained, it is used to filter actual customer data.

9. The customer screening method according to claim 7, wherein, In the implicit coding generation step, the implicit coding vector is generated using a lightweight fully connected neural network.

10. The customer screening method according to claim 7, wherein, In the virtual customer information generation step, the implicit coding vector and the noise vector are concatenated and then input into the generation network.

11. The customer screening method according to claim 7, wherein, In the virtual customer information generation step, a feature vector is generated as the virtual customer information.

12. The customer screening method according to any one of claims 7-11, wherein, The generator network and the adversarial network are trained with the goal of maximizing the result of the mutual information operation.

13. The customer screening method according to claim 12, wherein, The loss function is constructed to capture the semantic information contained in the implicit coding vector.

14. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the customer screening method according to any one of claims 7-13.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the customer screening method according to any one of claims 7-13.

16. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the customer screening method according to any one of claims 7-13.

Images