A method and device for screening marketing customer groups

Abstract

The purpose of this application is to provide a method and device for screening marketing customer groups. Compared with the prior art, this application collects customer group-related data; obtains a first target value representing the possibility of customer response, a second target value representing the customer's consumption ability, and a third target value representing the customer's fraud risk through a preset model based on the customer group-related data; generates a customer score based on the first target value, the second target value, and the third target value; and screens out marketing customer groups based on the customer score. In this way, in the process of screening marketing customer groups, the uncertain factors from responding to activities to generating effective consumption are fully considered, which increases the possibility that the screened marketing customer groups respond to marketing activities, participate in activities to generate consumption, and consume effectively without fraud risks.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a technology for screening marketing customer groups. Background Art

[0002] Existing marketing customer screening methods usually use response data from similar marketing activities in the past to develop a binary classification model for customer response results through a scoring model or a tree model, and then select the top customers in the currently reachable customer group based on the model scores, and use them as the main marketing customer group to send them event invitations.

[0003] Under the framework of this type of method, it is assumed that the value of each responding customer is equal, but there is still a long distance between responding to marketing activities and participating in activities to generate consumption, and the value that each person can bring to the customer group participating in activities to consume is also different. Therefore, the marketing customer group screened by this method may contain some customers who cannot generate real consumption. Summary of the invention

[0004] The purpose of this application is to provide a method and device for screening marketing customer groups.

[0005] According to one aspect of the present application, a method for screening marketing customer groups is provided, wherein the method comprises:

[0006] Collect customer-related data;

[0007] According to the customer group related data, a first target value representing the possibility of customer response, a second target value representing the customer's consumption capacity, and a third target value representing the customer's fraud risk are obtained through a preset model;

[0008] generating a customer score according to the first target value, the second target value, and the third target value;

[0009] The marketing customer groups are screened out according to the customer ratings.

[0010] Furthermore, the preset model includes a Deep module and an FM module, wherein after collecting customer group related data, the following steps are further included:

[0011] Dividing the customer group related data into text features, continuous features and category features according to data types;

[0012] Wherein, the first target value representing the possibility of customer response, the second target value representing the customer's consumption capacity, and the third target value representing the customer's fraud risk are obtained through a preset model according to the customer group related data, including:

[0013] Performing comprehensive deep information mining on the text features, the continuous features and the category features through the Deep module of the preset model;

[0014] Performing shallow information extraction on the category features through the FM module of the preset model;

[0015] The first target value, the second target value and the third target value are obtained by splicing the shallow information extraction and the deep information mining results.

[0016] Furthermore, before the Deep module of the preset model performs comprehensive deep information mining on the text features, the continuous features and the category features, the method further includes:

[0017] One-hot encoding the text features;

[0018] The one-hot encoded text features are dynamically embedded according to a dynamic embedding method, wherein the parameters of the dynamic embedding method are fixed.

[0019] Furthermore, the preset model further includes a first DNN layer, wherein after the dynamic embedding, it further includes:

[0020] Convert the dynamically embedded text features into sentence vectors;

[0021] Performing a dimensionality reduction operation on the sentence vector according to the first DNN layer;

[0022] The comprehensive deep information mining of the text features, the continuous features and the category features by the Deep module of the preset model includes:

[0023] The Deep module of the preset model performs comprehensive deep information mining on the text features, the continuous features and the category features after the dimensionality reduction operation.

[0024] Furthermore, the preset model further includes an embedding layer, wherein after the customer group-related data is divided into text features, continuous features and category features according to data types, it further includes:

[0025] One-hot encoding the category features;

[0026] The one-hot encoded category features are embedded according to the embedding layer.

[0027] Furthermore, the Deep module includes a sequentially linked Concat layer, a second DNN layer and an expert module.

[0028] The Concat splicing layer is used to splice the text features after the dimensionality reduction operation, the continuous features, and the category features after the embedding operation;

[0029] Wherein, the second DNN layer is used to interact with the splicing result of the Concat splicing layer;

[0030] The expert modules include a first expert module corresponding to a first target value, a second expert module corresponding to a second target value, and a third expert module corresponding to a third target value.

[0031] Further, the FM module includes a first FM module corresponding to the first target value, a second FM module corresponding to the second target value, and a third FM module corresponding to the third target value.

[0032] The first FM module, the second FM module and the third FM module all include sequentially linked linear layers and second-order cross layers.

[0033] Furthermore, the first target value, the second target value and the third target value obtained by splicing the shallow information extraction and the deep information mining results include:

[0034] splicing the output results of the first expert module and the first FM module into the first target value;

[0035] splicing the output results of the second expert module and the second FM module into the second target value;

[0036] The output results of the third expert module and the third FM module are concatenated into the third target value.

[0037] Furthermore, after dividing the customer group related data into the text features, the continuous features and the category features according to the data types, the step further includes:

[0038] Performing data cleaning and data processing on the text features, the continuous features and the category features,

[0039] Wherein, the data cleaning includes one or more of deleting redundant fields, data format processing and extreme value processing;

[0040] The data processing includes one or more of data aggregation, data scaling, data truncation and data type conversion.

[0041] Furthermore, the data type conversion includes:

[0042] The continuous features that need to be binned are converted into the category features.

[0043] Furthermore, the loss function corresponding to the first target value is a binary cross entropy loss:

[0044]

[0045] The loss function corresponding to the second target value is the mean square error loss:

[0046]

[0047] The loss function corresponding to the third target value is the binary cross entropy loss:

[0048]

[0049] The loss function of the preset model is:

[0050] L all =k 1 L 1 +k 2 L 2 +k 3 L 3 ,

[0051] Among them, k 1 , k 2 , k 3 is a hyperparameter.

[0052] Further, the first target value S 1 The second target value S 2 and the third target value S 3 Generating a customer rating S includes:

[0053] S=p 1 S 1 ×p 2 S 2 ×p 3 S 3 ,

[0054] Among them, p 1 、p 2 、p 3 is the scaling parameter.

[0055] Furthermore, the customer group related data includes overdue risk data, wherein the screening of marketing customer groups according to the customer scores includes:

[0056] The marketing customer group is determined comprehensively based on the overdue risk data and the customer score.

[0057] According to another aspect of the present application, a computer-readable medium is provided, on which computer-readable instructions are stored. The computer-readable instructions can be executed by a processor to implement the operations of the aforementioned method.

[0058] According to another aspect of the present application, a device for screening marketing customer groups is also provided, wherein the device comprises:

[0059] one or more processors; and

[0060] A memory storing computer readable instructions which, when executed, cause the processor to perform the operations of the above-described method.

[0061] Compared with the prior art, the present application collects customer group-related data; obtains a first target value representing the possibility of customer response, a second target value representing the customer's consumption capacity, and a third target value representing the customer's fraud risk through a preset model based on the customer group-related data; generates a customer score based on the first target value, the second target value, and the third target value; and selects a marketing customer group based on the customer score. In this way, in the process of selecting a marketing customer group, the uncertain factors between responding to an activity and generating effective consumption are fully considered, which increases the possibility that the selected marketing customer group responds to marketing activities, participates in activities to generate consumption, and consumes effectively without fraud risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0063] Figure 1 A flow chart of a method for screening marketing customer groups according to one aspect of the present application is shown;

[0064] Figure 2 A flow chart of a method for screening marketing customer groups according to a preferred embodiment of the present application is shown;

[0065] Figure 3 A model structure diagram for marketing customer group screening according to the present application is shown.

[0066] The same or similar reference numerals in the drawings represent the same or similar components. DETAILED DESCRIPTION

[0067] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0068] In a typical configuration of the present application, the terminal, the device of the service network and the trusted party all include one or more processors (CPU), input / output interfaces, network interfaces and memories.

[0069] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0070] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include non-transitory media such as modulated data signals and carrier waves.

[0071] In order to further explain the technical means adopted by the present application and the effects achieved, the technical solution of the present application is clearly and completely described below in combination with the accompanying drawings and preferred embodiments.

[0072] Figure 1 A method for screening marketing customer groups provided in one aspect of the present application is shown, wherein the method comprises:

[0073] S11 collects customer-related data;

[0074] S12 obtains a first target value representing the likelihood of customer response, a second target value representing the customer's spending power, and a third target value representing the customer's fraud risk through a preset model according to the customer group-related data;

[0075] S13 generates a customer score according to the first target value, the second target value and the third target value;

[0076] S14 selects marketing customer groups according to the customer scores.

[0077] In this embodiment, in step S11, customer group related data is collected.

[0078] Here, customer group related data includes marketing activity data and pushed customer data. Among them, marketing activity data includes but is not limited to the specific description of marketing activities, activity types, marketing channels and reward mechanisms, etc. Such data comes from the preparation data of previous marketing activities and related activities; pushed customer data corresponds to the marketing activities that provide marketing activity data, and is all customer-related data that can be obtained, such as the portrait data, consumption preference data, consumption behavior data and device data of customers who are pushed and / or respond to such marketing activities. It is understandable that since marketing activity data and pushed customer data are used to train preset models, the higher the similarity between the marketing activities that provide data and the marketing activities of this screening customer group, the more suitable the trained preset model is for the application scenario of this screening customer group.

[0079] Continuing in this step, this application does not provide any restrictive description of specific marketing activities, and the marketing activities may be online and / or offline marketing activities launched by various businesses or organizations.

[0080] In this embodiment, in step S12, a first target value representing the likelihood of customer response, a second target value representing the customer's spending power, and a third target value representing the customer's fraud risk are obtained through a preset model based on the customer group-related data.

[0081] Here, three target variables are defined, and then the possibility of customers responding to marketing activities and making effective consumption is considered from three dimensions. The first target value represents the customer's customer response possibility, that is, whether the customer can participate in the marketing activity after being pushed. This target variable is a binary variable; the second target value represents the customer's consumption ability, that is, the consumption (or consumption amount) that the customer may bring after participating in the activity. This target variable is a continuous variable; the third target value represents the customer's fraud risk, that is, whether the consumption brought by the customer after participating in the activity can be converted into effective income for the initiator of the marketing activity. This target variable is a binary variable. If the customer uses the marketing activities to make improper gains, for example, organized theft, wool pulling, loan fraud, cashing out, brushing orders, brushing good reviews, etc., it is considered fraudulent consumption. Even if such customers participate in marketing activities and make large-scale consumption, such consumption cannot be converted into effective income for the initiator of the marketing activity, but may cause losses to the initiator of the marketing activity.

[0082] In this embodiment, in step S13, a customer score is generated according to the first target value, the second target value, and the third target value.

[0083] In this step, the present application considers customers from the above three dimensions and takes the three target values ​​as the decisive factors for customer scoring, so that in the process of screening marketing customer groups, the uncertain factors from responding to activities to generating effective consumption are fully considered, ensuring that the screened marketing customer groups can respond to marketing activities, generate consumption, and the consumption is effective without fraud risk.

[0084] In this embodiment, in step S14, marketing customer groups are screened out based on the customer scores.

[0085] Here, since the customer rating is a comprehensive rating of the three aspects of response possibility, fraud possibility and consumption ability, the comprehensive rating is used as the basis for screening marketing customer groups, which fully considers the importance differences between individual customers. For example, customer A has a higher response possibility, lower consumption ability and lower fraud possibility; customer B has a higher response possibility, higher consumption ability and lower fraud possibility. According to the actual marketing scenario, although both customer A and customer B can bring effective consumption, customer B is more important and is an important marketing target in the marketing customer group. For another example, customer C is more likely to respond, has lower spending power, and is more likely to commit fraud. When selecting customer C and customer A as marketing targets, a screening strategy can be set according to the needs of the initiator of the marketing activity. For example, if the initiator of the marketing activity believes that for customers with lower spending power, even if the possibility of consumption fraud is higher, they can still be used as marketing targets because they may cause lower losses. In this way, the customer screening strategy can be set so that when the spending power is lower than a certain threshold, the weight of the third target value in the customer score can be adjusted to a preset value to reduce the impact of the possibility of fraud on the customer score. In this way, while ensuring that fraudulent consumption will not cause greater losses, the scope of the marketing customer base can be expanded as much as possible. Similarly, if the initiator of the marketing activity has a low risk resistance ability, the weight of the third target value in the customer score can be adjusted to another preset value to increase the impact of the possibility of fraud on the customer score. In this way, the possibility of fraud is used as an important factor in screening customers, thereby minimizing the adverse effects of fraudulent consumption on the initiator of the marketing activity. According to the above examples, since this application measures the differences in importance between individual customers from three dimensions, the screening process of marketing customer groups fully considers the differences in response possibility, fraud possibility and consumption capacity, and then different customer group screening strategies can be set according to different marketing scenarios to screen out high-quality marketing customer groups within the risk-bearing capacity range.

[0086] In a preferred embodiment, see Figure 2 ,in, Figure 2 Steps S21, S24 and S25 in Figure 1Steps S11, S13 and S14 in the embodiment are the same or substantially the same, so they are not described here in detail and are only included here by reference. The preset model includes a Deep module and an FM module, wherein the step S22 includes: dividing the customer group-related data into text features, continuous features and category features according to data types; the step S23 includes: performing comprehensive deep information mining on the text features, the continuous features and the category features through the Deep module of the preset model; performing shallow information extraction on the category features through the FM module of the preset model; and obtaining the first target value, the second target value and the third target value by splicing the results of the shallow information extraction and the deep information mining.

[0087] Here, since customer group-related data (including marketing activity data and pushed customer data) covers different data types, the acquired training data (including marketing activity data and pushed customer data) is classified according to the data type, and is divided into text features, continuous features and category features. For example, data such as activity content, activity description and customer address are text features, data such as activity reward amount and customer consumption amount are continuous features, and data such as marketing channels and customer gender are category features. Different model processing methods are then used for different types of data.

[0088] Furthermore, after dividing the customer-related data into the text features, the continuous features and the category features according to data types, it also includes: performing data cleaning and data processing on the text features, the continuous features and the category features, wherein the data cleaning includes deleting redundant fields, data format processing and extreme value processing; wherein the data processing includes one or more of data aggregation, data scaling, data truncation and data type conversion.

[0089] Here, data cleaning is performed, including extreme value processing of continuous features (for example, if the extreme value of customer income has a low credibility problem, the extreme value can be deleted or merged into a certain income range, etc.), special symbols of text features (for example, punctuation marks, voice intonation marks and other redundant fields in text features are deleted) and format processing (for example, punctuation marks in text features are unified into Chinese punctuation / English punctuation), etc. After that, feature derivation work is performed, including but not limited to aggregation on time segments (for example, integrating the consumption amount data of users in a certain time period), transforming continuous features that need to be binned into categorical features, performing variable conversion in cross-sectional state (for example, scaling customer income data, such as logarithmic conversion, etc.), truncating text features, etc.

[0090] In an application scenario, for example, if the correlation between customer income (a continuous feature) and customer gender (a categorical feature) needs to be considered, and customer groups are screened based on this correlation, the customer income is binned and converted into a categorical feature, so that the categorical customer income and the categorical customer gender are simultaneously subjected to shallow information extraction in the FM module, thereby obtaining the correlation information between customer income and gender. At the same time, it should be made clear that in order to ensure sufficient information extraction of the input features, for binned features such as customer income, it is also necessary to treat them as continuous features and perform deep information mining through the Deep module.

[0091] Continuing in this step, the preset model includes a Deep module and an FM module, and these two modules are used to mine deep information and extract shallow information respectively. For example, due to the relatively high complexity of text features and continuous features, if the FM module is used to extract shallow information, on the one hand, the data processing is more difficult, and on the other hand, the value of the extracted information is low. Therefore, only the Deep module is used to mine deep information on text features and continuous features; while the category features are relatively less complex, and shallow information can be extracted from them through the FM module, and deep information can be mined through the Deep module. Therefore, in this solution, both aspects of information are extracted from the category features at the same time to obtain as much valuable information as possible.

[0092] Continuing in this step, after extracting deep and shallow information from the training data, the deep information is concatenated with the shallow information to obtain three types of target values.

[0093] Furthermore, before the Deep module of the preset model performs comprehensive deep information mining on the text features, the continuous features and the category features, it also includes: one-hot encoding the text features; dynamically embedding the one-hot encoded text features according to a dynamic embedding method, wherein the parameters of the dynamic embedding method are fixed.

[0094] Here, due to the large amount of information and high complexity of text features, the one-hot encoding amount is huge. Therefore, in the process of embedding text features, if the neural network Embedding layer is used for embedding, the required data processing amount is extremely large, the corresponding time and resource costs are extremely high, and the embedding efficiency is low. Therefore, a dynamic embedding method is used to embed text features, and the embedding efficiency is improved by referencing an external word network. At the same time, due to the data volume of text features in the customer screening scenario, it is difficult to support the continued training of the dynamic embedding module. In order to speed up the model training speed, the parameters of the dynamic embedding are fixed and do not participate in subsequent training and tuning. Here, this application does not provide a restrictive description of the specific dynamic embedding method. The dynamic embedding method that can be used includes but is not limited to BERT (Bidirectional Encoder Representation from Transformers, deep bidirectional pre-training based on semantic understanding).

[0095] Furthermore, the preset model also includes a first DNN layer, wherein, after the dynamic embedding, it also includes: converting the text features after the dynamic embedding into sentence vectors; performing a dimensionality reduction operation on the sentence vectors according to the first DNN layer; wherein, performing comprehensive deep information mining on the text features, the continuous features and the category features through the Deep module of the preset model includes: performing comprehensive deep information mining on the text features, the continuous features and the category features after the dimensionality reduction operation through the Deep module of the preset model.

[0096] Here, after the one-hot encoded text features are embedded, they are converted into sentence vectors. Since the length of the embedded sentence vectors is relatively long (when BERT is used for embedding, the sentence vectors are 768 dimensions), and the data dimensions of continuous features and category features are relatively short, in order to make the importance of each target value obtained by splicing the subsequent deep information with the shallow information comparable, the embedded sentence vectors are passed through the DNN network layer (i.e., the first DNN layer) for data dimensionality reduction. At the same time, the hyperparameters of the first DNN layer can be adjusted during the training process. Here, no restrictive description is given to the specific method of generating sentence vectors and the target dimension range after data dimensionality reduction. In an application scenario, the one-hot encoded text features can be averaged and converted into sentence vectors, and then the sentence vectors are reduced to 5-100 dimensions through the DNN network layer. It should be clear that the specific method of generating sentence vectors and the target dimension range after data dimensionality reduction should be determined according to the actual application scenario.

[0097] Furthermore, the preset model also includes an embedding layer, wherein after dividing the customer-related data into text features, continuous features and category features according to data types, it also includes: one-hot encoding the category features; and embedding the one-hot encoded category features according to the embedding layer.

[0098] Here, due to the low complexity of the category features, the one-hot encoded category features can be directly embedded using the embedding layer, thereby converting the one-hot encoded category features into a dense vector and controlling the vector length to 2-5 dimensions.

[0099] Correspondingly, for continuous features, since each continuous feature is only a scalar, all continuous features are concatenated into a one-dimensional vector, and the length of the vector is the number of continuous features.

[0100] Furthermore, the Deep module includes a sequentially linked Concat splicing layer, a second DNN layer and an expert module, wherein the Concat splicing layer is used to splice the text features after the dimensionality reduction operation, the continuous features and the category features after the embedding operation; wherein the second DNN layer is used to interact with the splicing results of the Concat splicing layer; wherein the expert module includes a first expert module corresponding to the first target value, a second expert module corresponding to the second target value, and a third expert module corresponding to the third target value in parallel.

[0101] Here, the Deep module is responsible for mining deep information. First, the vectors of text features, continuous features and category features processed as above are spliced ​​through the Concat splicing layer; then all the spliced ​​vectors are interacted through the second DNN layer to conduct a comprehensive investigation of the correlation between all features; finally, in order to connect the three target values, three expert modules are set in the Deep module, and each expert module can customize its own processing logic, for example, a multi-layer DNN structure can be used to train and learn all the vectors after interaction.

[0102] Furthermore, the FM module includes a first FM module corresponding to a first target value, a second FM module corresponding to a second target value, and a third FM module corresponding to a third target value, wherein the first FM module, the second FM module and the third FM module all include sequentially linked linear layers and second-order cross layers.

[0103] Here, the FM module is responsible for extracting the interaction of category features, performing first-order linear aggregation of the linear layer and pairwise feature vector crossover between the second order. The FM module composed of the linear layer and the second-order cross layer can well extract the information in the category features. At the same time, in order to match the final three target values, the FM module consists of three FM sub-modules.

[0104] Furthermore, the first target value, the second target value and the third target value obtained by splicing the shallow information extraction and the deep information mining results include: splicing the output results of the first expert module and the first FM module into the first target value; splicing the output results of the second expert module and the second FM module into the second target value; splicing the output results of the third expert module and the third FM module into the third target value.

[0105] Here, the output results of the expert modules and FM submodules corresponding to the three target values ​​are concatenated, and the output vectors are mapped to the dimensions corresponding to each target value through a DNN layer. At the same time, corresponding to the first target value and the third target value, since they are two-component types, the mapped results need to be processed into two categories through the sigmoid function.

[0106] In this solution, since corresponding Deep modules and FM modules are set for each of the three target values, it is equivalent to setting up a response possibility model, a consumption capacity model and a fraud possibility model from three dimensions. The output results of each model correspond to the response possibility, fraud possibility and consumption capacity. The output results of each model are integrated according to the integration strategy to obtain the final customer score.

[0107] Furthermore, the loss function corresponding to the first target value is a binary cross entropy loss:

[0108]

[0109] The loss function corresponding to the second target value is the mean square error loss:

[0110]

[0111] The loss function corresponding to the third target value is the binary cross entropy loss:

[0112]

[0113] The loss function of the preset model is:

[0114] L all =k 1 L 1 +k 2 L2 +k 3 L 3 ,

[0115] Among them, k 1 , k 2 , k 3 is a hyperparameter.

[0116] Here, since the first target value representing the possibility of response and the third target value representing the possibility of fraud are binary variables, the loss functions corresponding to the response possibility model and the fraud possibility model are both binary cross entropy losses. Since the second target value representing the consumption capacity is a continuous variable, the loss function corresponding to the consumption capacity model can be set to mean square error loss, and the integrated result of the three loss functions is used as the loss function of the preset model for the entire customer group screening. At the same time, since there is a difference in the order of magnitude of the loss function between continuous variables and binary variables, in order to balance the difference in order of magnitude, a hyperparameter is set, and the hyperparameter k is adjusted during the model training process according to the parameter adjustment strategy. 1 , k 2 , k 3 Adjustments are made so that the three loss functions are at the same or similar order of magnitude. Here, the parameter adjustment strategy includes but is not limited to the grid search method.

[0117] Further, according to the first target value S 1 The second target value S 2 and the third target value S 3 Generating a customer rating S includes: S = p 1 S 1 ×p 2 S 2 ×p 3 S 3 , where p 1 、p 2 、p 3 is the scaling parameter.

[0118] Here, similarly, since there is a difference in magnitude between the continuous variable and the two-component variable, in order to balance the difference in magnitude, the scaling parameter p is set 1 、p 2 、p 3 , so that the three target values ​​are at the same or similar order of magnitude. It should be clear that the calculation method of the customer score S is related to the representation meaning of each target value. For example, if the first target value S 1 The larger the value, the higher the probability of response. 3 The larger the value, the higher the possibility of fraud, then S = p 1 S 1 ×p 2 S2 ×p 3 (1-S 3 ). After calculating the customer score, you can set a customer score threshold based on the actual application scenario and add customers with a score higher than the customer score threshold to the marketing customer group.

[0119] Furthermore, the customer group related data also includes overdue risk data, wherein screening out the marketing customer group according to the customer score includes: comprehensively determining the marketing customer group according to the overdue risk data and the customer score.

[0120] Here, the overdue risk data of customers who use installment payment or other post-payment methods can be collected, and the marketing customer groups can be cross-delineated in combination with customer scores to reduce the risk of overdue non-repayment after consumption. For example, three overdue risk groups can be defined based on the overdue risk value, and three score groups can be defined based on the customer score value. The marketing customer groups can be determined by cross-delineation. The cross-delineation method in this scenario is as follows:

[0121]

[0122] Compared with the prior art, the present application collects customer group-related data; obtains a first target value representing the possibility of customer response, a second target value representing the customer's consumption capacity, and a third target value representing the customer's fraud risk through a preset model based on the customer group-related data; generates a customer score based on the first target value, the second target value, and the third target value; and selects a marketing customer group based on the customer score. In this way, in the process of selecting a marketing customer group, the uncertain factors between responding to an activity and generating effective consumption are fully considered, which increases the possibility that the selected marketing customer group responds to marketing activities, participates in activities to generate consumption, and consumes effectively without fraud risks.

[0123] In addition, an embodiment of the present application also provides a computer-readable medium on which computer-readable instructions are stored, and the computer-readable instructions can be executed by a processor to implement the aforementioned method.

[0124] The embodiment of the present application also provides a device for screening marketing customer groups, wherein the device includes:

[0125] one or more processors; and

[0126] A memory storing computer readable instructions which, when executed, cause the processor to perform the operations of the aforementioned method.

[0127] For example, the computer readable instructions, when executed, cause the one or more processors to: collect customer segment related data;

[0128] According to the customer group related data, a first target value representing the possibility of customer response, a second target value representing the customer's consumption capacity, and a third target value representing the customer's fraud risk are obtained through a preset model;

[0129] generating a customer score according to the first target value, the second target value, and the third target value;

[0130] The marketing customer groups are screened out according to the customer ratings.

[0131] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved. In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices stated in the device claim can also be implemented by one unit or device through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any particular order.

Claims

1. A method for screening marketing customer groups. in, The method comprises: Collect customer group-related data, including overdue risk data, marketing activity data and pushed customer data. The marketing activity data includes the specific description of the marketing activity, activity type, marketing channel and reward mechanism. The pushed customer data corresponds to the marketing activity that provides the marketing activity data, including the profile data, consumption preference data, consumption behavior data and device data of the customers who are pushed and / or respond to such marketing activities; A first target value, a second target value, and a third target value are obtained through a preset model according to the customer group related data, wherein the first target value represents the customer response possibility of the customer, indicating whether the customer can participate in the marketing activity after being pushed to the customer, the second target value represents the customer consumption capacity, indicating the consumption amount that can be brought by the customer after participating in the activity, and the third target value represents the fraud risk of the customer, indicating whether the consumption brought by the customer after participating in the activity can be converted into effective income for the initiator of the marketing activity; generating a customer score according to the first target value, the second target value, and the third target value; Comprehensively determine the marketing customer group based on the overdue risk data and the customer score; The preset model includes a Deep module and an FM module, wherein after collecting customer group related data, the following steps are further included: Dividing the customer group related data into text features, continuous features and category features according to data types; Performing comprehensive deep information mining on the text features, the continuous features and the category features through the Deep module of the preset model; Performing shallow information extraction on the category features through the FM module of the preset model; The first target value, the second target value and the third target value are obtained by splicing the results of the shallow information extraction and the deep information mining.

2. The method according to claim 1, in, Before the Deep module of the preset model performs comprehensive deep information mining on the text features, the continuous features and the category features, the method further includes: One-hot encoding the text features; The one-hot encoded text features are dynamically embedded according to a dynamic embedding method, wherein the parameters of the dynamic embedding method are fixed.

3. The method according to claim 2, wherein the preset model further comprises a first DNN layer, in, After the dynamic embedding, the following is also included: Convert the dynamically embedded text features into sentence vectors; Performing a dimensionality reduction operation on the sentence vector according to the first DNN layer; The comprehensive deep information mining of the text features, the continuous features and the category features by the Deep module of the preset model includes: The Deep module of the preset model performs comprehensive deep information mining on the text features, the continuous features and the category features after the dimensionality reduction operation.

4. According to the method according to any one of claims 1 to 3, the preset model further includes an embedding layer, in, After the customer group related data is divided into text features, continuous features and category features according to data types, the method further includes: One-hot encoding the category features; The one-hot encoded category features are embedded according to the embedding layer.

5. According to the method of claim 3, the preset model further comprises an embedding layer, in, After the customer group related data is divided into text features, continuous features and category features according to data types, the method further includes: One-hot encoding the category features; The one-hot encoded category features are embedded according to the embedding layer; The Deep module includes a sequentially linked Concat layer, a second DNN layer and an expert module. The Concat splicing layer is used to splice the text features after the dimensionality reduction operation, the continuous features, and the category features after the embedding operation; Wherein, the second DNN layer is used to interact with the splicing result of the Concat splicing layer; The expert modules include a first expert module corresponding to a first target value, a second expert module corresponding to a second target value, and a third expert module corresponding to a third target value.

6. The method according to claim 5, wherein the FM module comprises a first FM module corresponding to the first target value, a second FM module corresponding to the second target value, and a third FM module corresponding to the third target value. in, The first FM module, the second FM module and the third FM module each include a linear layer and a second-order cross layer connected in sequence.

7. The method according to claim 6, in, The first target value, the second target value and the third target value obtained by splicing the shallow information extraction and deep information mining results include: splicing the output results of the first expert module and the first FM module into the first target value; splicing the output results of the second expert module and the second FM module into the second target value; The output results of the third expert module and the third FM module are spliced ​​into the third target value.

8. The method according to claim 1, in, After dividing the customer group related data into text features, continuous features and category features according to data types, the following further comprises: Performing data cleaning and data processing on the text features, the continuous features and the category features, wherein the data cleaning includes one or more of deleting redundant fields, data format processing and extreme value processing; The data processing includes one or more of data aggregation, data scaling, data truncation and data type conversion.

9. The method according to claim 8, in, The data type conversion includes: The continuous features that need to be binned are converted into the category features.

10. The method according to claim 1, in, The loss function corresponding to the first target value is the binary cross entropy loss: The loss function corresponding to the second target value is the mean square error loss: The loss function corresponding to the third target value is the binary cross entropy loss: The loss function of the preset model is: L all =k 1 L 1 +k 2 L 2 +k 3 L 3 , Among them, k 1 , k 2 , k 3 is a hyperparameter.

11. The method according to claim 1, in, The first target value S 1 The second target value S 2 and the third target value S 3 Generating a customer rating S includes: S=p 1 S 1 ×p 2 S 2 ×p 3 S 3 , Among them, p 1 、p 2 、p 3 is the scaling parameter.

12. A computer-readable medium having computer-readable instructions stored thereon, wherein the computer-readable instructions can be executed by a processor to implement the method according to any one of claims 1 to 11.

13. A device for screening marketing customers. in, The equipment includes: one or more processors; and A memory storing computer readable instructions which, when executed, cause the processor to perform the operations of the method of any one of claims 1 to 11.

Images