Marketing method and device for target customers, electronic equipment and storage medium

Abstract

The application discloses a marketing method and device for a target customer, an electronic device, and a storage medium. The method comprises the following steps: establishing a prediction model based on a customer data set and a marketing data set, the prediction model being used for predicting the sensitivity of the target customer to a marketing action; inputting actual customer data to be predicted and marketing data into the trained prediction model, outputting a prediction result of the sensitivity of the target customer to the marketing action according to the prediction model, and generating a matched marketing scheme according to the prediction result. The application provides a model for matching a customer and a marketing scheme based on multi-modal data, and further provides a method for formulating a personalized marketing scheme for a customer.

Description

Technical Field

[0001] This application relates to the field of internet product marketing technology, and in particular to a marketing method, device, electronic device, and storage medium for target customers. Background Technology

[0002] With the rapid development of wealth management business, high-net-worth clients have become a core customer group for commercial banks. Marketing channels targeting high-net-worth clients are abundant, encompassing online push notifications, telephone outreach, and one-on-one offline visits, among others. However, the costs associated with marketing to this type of client—including manpower, materials, and channel deployment—are high, resulting in an overall expensive marketing investment.

[0003] Currently, traditional bank marketing models generally have significant shortcomings: marketing plans are mostly based on product promotion needs, failing to fully explore customers' real needs, behavioral preferences, and acceptance levels, which can easily lead to pushing mismatched wealth management products and value-added services to customers; at the same time, there is a lack of unified control over the frequency and reach of multi-channel marketing, and frequent and homogeneous marketing can easily cause excessive disturbance to customers, resulting not only in poor marketing conversion but also in a high risk of provoking resentment from high-net-worth customers, reducing customer service experience and customer loyalty.

[0004] Existing customer-focused precision marketing solutions often rely solely on simple tiered matching based on basic structured customer data. This results in limitations such as a single data dimension and the inability to integrate multimodal information like online customer behavior, in-person call transcripts, and historical marketing feedback. Consequently, it becomes difficult to accurately predict customer sensitivity to different marketing products and channels. Furthermore, existing solutions do not perform global optimization for independent marketing budgets across multiple channels, making it impossible to automatically assemble suitable multi-channel sequential marketing plans for customers within cost constraints. Consequently, it is difficult to simultaneously reduce marketing costs and improve customer satisfaction.

[0005] In summary, existing marketing methods for high-net-worth clients in banks suffer from technical pain points such as low matching accuracy, waste of marketing resources, poor customer experience, and inability to coordinate and optimize multi-channel budgets. There is an urgent need for a customer marketing method that can accurately predict customer marketing sensitivity and automatically generate suitable marketing mixes under the constraints of channel costs. Summary of the Invention

[0006] This application provides a marketing method, device, electronic device, and storage medium for target customers, adapted to long-term continuous tracking marketing scenarios for high-net-worth customers.

[0007] The embodiments of this application adopt the following technical solutions:

[0008] In a first aspect, embodiments of this application provide a marketing method for target customers, wherein the method includes:

[0009] Based on customer and marketing datasets, a predictive model is built to predict the sensitivity of target customers to marketing actions. The predictive model includes a customer encoder, a marketing encoder, and a predictive interaction layer. The customer encoder extracts and fuses features from customer data, outputting a customer representation vector. The marketing encoder extracts and fuses features from marketing data, outputting a marketing representation vector. The predictive interaction layer fuses the customer representation vector and the marketing representation vector to calculate and output a sensitivity score for the corresponding marketing action.

[0010] The actual customer data and marketing data to be predicted are input into the trained prediction model. Based on the prediction model, the model outputs the prediction results of the target customer's sensitivity to marketing actions, and a matching marketing plan is generated based on the prediction results.

[0011] In some embodiments, the customer encoder is configured with a fusion layer, which is used to unify the dimensions of multimodal customer data and fuse multimodal information. The fusion layer sequentially performs three processing steps: feature alignment, cross-modal feature fusion, and fully connected mapping output.

[0012] The feature alignment step includes: normalizing and linear activation transformations for various modal features of different dimensions using an adapter; setting adaptive weight coefficients to control the contribution intensity of each modal feature; and converting all modal features into a unified dimension.

[0013] The cross-modal feature fusion step includes: employing a cross-modal attention mechanism, using one modal feature as the query vector and the remaining modal features as key vectors to calculate attention; concatenating all attention information corresponding to the modality and fusing it with the original features, and performing normalization processing to obtain the updated modal features;

[0014] The fully connected mapping output step includes: stacking all updated modal features, mining global modal correlations through a multi-head self-attention mechanism; flattening the processed features, and transforming them through activation, random deactivation, and full connection to output a customer representation vector with a fixed dimension.

[0015] In some embodiments, the marketing encoder's internal fusion layer employs the same feature alignment, cross-modal fusion, and fully connected mapping processing logic as the customer encoder, outputting a marketing representation vector that matches the dimension of the customer representation vector.

[0016] In some embodiments, before feature fusion, the marketing encoder first uses the BERT model to vectorize the marketing text materials, and then sends the text vectors and other marketing features into the fusion layer.

[0017] In some embodiments, the predictive interaction layer uses dot product or cosine similarity to calculate a sensitivity score from the customer representation vector and the marketing representation vector.

[0018] In some embodiments, a multi-channel cross-marketing strategy is configured for a single customer, the marketing strategies are sorted according to the degree of marketing fit, and a 0-1 programming model with channel budget constraints is used to solve for the optimal sequential marketing mix to maximize the overall average marketing fit.

[0019] The 0-1 programming model expression is as follows:

[0020] Objective function:

[0021]

[0022] Constraints:

[0023]

[0024] x_i is not all zero, y_j is not all zero, z_k is not all zero.

[0025] in, The weighted matching score for the corresponding marketing strategy; These represent 0-1 decision variables for online channels, face-to-face channels, and telephone marketing channels, respectively; cost represents the implementation cost of a single marketing strategy. The budgets are for marketing costs for online channels, face-to-face channels, and telephone marketing channels, respectively.

[0026] In some embodiments, the target customers are high-net-worth customers, and the customer dataset additionally includes characteristics related to customer asset size, asset planning goals, channel preferences, and risk preferences.

[0027] Secondly, embodiments of this application also provide a marketing device for target customers, wherein the device includes:

[0028] A prediction model building module is used to build a prediction model based on customer and marketing datasets. This prediction model is used to predict the sensitivity of target customers to marketing actions. The prediction model includes a customer encoder, a marketing encoder, and a prediction interaction layer. The customer encoder extracts and fuses features from customer data, outputting a customer representation vector. The marketing encoder extracts and fuses features from marketing data, outputting a marketing representation vector. The prediction interaction layer fuses the customer representation vector and the marketing representation vector, outputting a sensitivity score for the corresponding marketing action.

[0029] The prediction module is used to input the actual customer data and marketing data to be predicted into the trained prediction model, output the prediction result of the target customer's sensitivity to marketing actions based on the prediction model, and generate a matching marketing plan based on the prediction result.

[0030] Thirdly, embodiments of this application also provide an electronic device, including: a processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the above-described method.

[0031] Fourthly, embodiments of this application also provide a computer-readable storage medium that stores one or more programs, which, when executed by an electronic device including multiple applications, cause the electronic device to perform the above-described method.

[0032] The above-mentioned at least one technical solution adopted in the embodiments of this application can achieve the following beneficial effects: a prediction model is established based on a customer dataset and a marketing dataset; the actual customer data and marketing data to be predicted are input into the trained prediction model, the prediction model outputs the prediction result of the target customer's sensitivity to marketing actions, and a matching marketing plan is generated based on the prediction result.

[0033] This application addresses the problem that existing bank marketing solutions are product-promotion-oriented and fail to meet the real needs of high-net-worth clients. Based on multimodal customer data, it accurately predicts the sensitivity of customers to various product and channel marketing activities, matches suitable products, services and marketing methods from the customer's perspective, reduces ineffective push notifications, avoids excessive customer disturbance, and improves the service experience and satisfaction of high-net-worth clients.

[0034] This application overcomes the shortcomings of existing prior art, which only applies to online general customers, can only filter marketing time points, rely on single structured data, and lack multimodal feature fusion modeling. It constructs a multimodal dual-tower prediction model that includes a customer encoder and a marketing encoder, and integrates multi-dimensional information such as customer profiles, behavioral sequences, call interview texts, and historical marketing feedback to accurately identify the real needs of high-net-worth customers in the long-term tracking and maintenance process.

[0035] This application sets independent cost budget constraints for multiple channels such as online, telephone, and offline, and uses a multi-channel joint 0-1 planning model to optimize and generate a sequential marketing mix. Under the budget limit of each channel, it maximizes the overall marketing matching degree, realizes the rational allocation of marketing resources, and reduces the overall marketing cost.

[0036] This application is adapted to long-term, continuous marketing scenarios for high-net-worth clients. Unlike existing technologies that only select clients from predetermined marketing plans, this application can complete the design of customized marketing plan combinations in advance, providing complete technical support for long-term wealth management marketing for high-net-worth clients throughout their entire life cycle, and simultaneously improving marketing conversion rates and customer loyalty. Attached Figure Description

[0037] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0038] Figure 1 This is a flowchart illustrating the marketing method for target customers in an embodiment of this application;

[0039] Figure 2 This is a schematic diagram of the predictive model in the marketing method for target customers in the embodiments of this application;

[0040] Figure 3 This is a schematic diagram of the customer encoder structure in the marketing method for target customers in the embodiments of this application;

[0041] Figure 4 This is a schematic diagram of the marketing device for the target customer in the embodiments of this application;

[0042] Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0044] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0045] It's understandable that in the banking wealth management field, high-net-worth clients refer to individual clients whose financial assets reach a specified threshold. The common classification standard among mainstream commercial banks in China is: Private banking-level high-net-worth clients: personal investable financial assets ≥ RMB 6 million; High-end wealth clients: personal investable financial assets ≥ RMB 1 million. Investable financial assets include: current / time deposits, wealth management products, funds, stocks, bonds, insurance, precious metals, trusts, and other liquid financial assets, excluding tangible fixed assets such as owner-occupied housing and vehicles. However, most related marketing technologies are only applicable to online marketing scenarios, focusing primarily on selecting the optimal marketing reach time for a massive customer base. The marginal return on investment for mass marketing to a large customer base is low, and these existing technologies are all based on reverse-engineering existing marketing plans to select suitable clients, unable to pre-define a complete marketing package for clients. Therefore, their reference value for designing marketing plans for the long-term, full-cycle tracking and maintenance of high-net-worth clients is limited.

[0046] This application provides a marketing method for target customers, such as... Figure 1 The diagram illustrates a marketing method for target customers in an embodiment of this application. The method includes at least the following steps S110 to S120:

[0047] Step S110: Based on the customer dataset and marketing dataset, a prediction model is established. This prediction model is used to predict the sensitivity of target customers to marketing actions. The prediction model includes a customer encoder, a marketing encoder, and a prediction interaction layer. The customer encoder extracts and fuses features from customer data, outputting a customer representation vector. The marketing encoder extracts and fuses features from marketing data, outputting a marketing representation vector. The prediction interaction layer fuses the customer representation vector and the marketing representation vector to calculate and output a sensitivity score for the corresponding marketing action.

[0048] Two types of basic datasets are pre-built: customer dataset and marketing dataset.

[0049] The target customers are the bank's high-net-worth clients. Therefore, in addition to the regular basic data, the customer dataset also includes exclusive features such as customer asset size, asset planning goals, channel preferences, and risk preferences. The customer dataset contains multimodal customer data, including numerical, labeled, sequential, historical marketing responses, and call / face-to-face interview texts. The marketing dataset stores marketing-related data such as various wealth management products, value-added services, online and offline marketing activities, marketing scripts, and promotional copy.

[0050] Cleaning, standardization, and labeling preprocessing were performed on the two types of datasets respectively: For text materials such as marketing copy and service scripts in the marketing data, the BERT pre-trained model was used to vectorize the text and generate fixed-dimensional text feature vectors, which were then merged with structured features such as product duration, minimum purchase amount, and risk level to form complete marketing data input.

[0051] Based on the processed customer and marketing datasets, a predictive model is built. This model is specifically designed to quantify the sensitivity of high-net-worth customers to different marketing actions. The model has a dual-tower structure, consisting of three main parts: a customer encoder, a marketing encoder, and a predictive interaction layer.

[0052] Step S120: Input the actual customer data and marketing data to be predicted into the trained prediction model, output the prediction result of the target customer's sensitivity to marketing actions according to the prediction model, and generate a matching marketing plan based on the prediction result.

[0053] In business scenarios, real-time multimodal data of high-net-worth customers to be marketed, as well as marketing data of products and activities to be launched, are collected as inference inputs. The two sets of data are fed into the trained prediction model, which outputs the sensitivity prediction results corresponding to each marketing action. Then, the optimal sequential marketing mix is ​​solved through the aforementioned multi-channel 0-1 planning model to generate a customized marketing plan that meets customer needs, balances channels, and complies with budget constraints. Marketing outreach is carried out in the planned order.

[0054] like Figure 2 As shown, the overall processing flowchart of the dual-tower model for predicting the marketing sensitivity of high-net-worth clients in this application is presented, fully demonstrating the entire process from raw data input, independent encoding of the two branches, vector interaction calculation, to outputting marketing sensitivity prediction results. The model adopts a symmetrical dual-tower architecture, with the left branch specifically processing multimodal data from the client side and the right branch specifically processing marketing data. After both branches are encoded, they are uniformly entered into the interactive prediction module to calculate the matching sensitivity, and finally output quantitative prediction results, providing numerical basis for subsequent optimization of multi-channel marketing solutions.

[0055] exist Figure 2 The top-level module is the data input module, which simultaneously receives two independent data sources: customer data to be modeled / inferred, and marketing data. Customer data includes... Figure 2 The data presented includes five multimodal categories: numerical, tagged, sequential, marketing response, and text. For high-net-worth clients, additional features are provided based on their asset size, asset planning goals, channel preferences, and risk appetite. Marketing data includes heterogeneous marketing data such as financial product parameters, value-added service information, marketing materials from various channels (online / in-person / telephone), marketing costs, activity tags, and marketing copy. Both types of data are fed in parallel into the corresponding encoders in the dual-tower structure below.

[0056] exist Figure 2 The core dual-tower module in the model consists of three sub-modules: customer encoder, marketing encoder, and interaction prediction, forming the core dual-tower structure of the model.

[0057] (1) Customer encoder

[0058] Receive all customer multimodal data and internally execute additional... Figure 2 The documented categorized independent encoding process includes: numerical data standardization + MLP, label data embedding, sequence data Transformer Encoder, marketing response data LSTM, and text data BERT encoding. After the output of five modal features, they are fed into the built-in fusion layer. Through three steps of feature alignment, cross-modal attention fusion, and fully connected mapping, a fixed-dimensional customer representation vector is output, which fully carries all information about high-net-worth customers' assets, behaviors, needs, and marketing acceptance.

[0059] (2) Marketing encoder

[0060] It receives all marketing-side data and reuses the multimodal processing logic that is completely consistent with the customer encoder: structured marketing numerical standardization, marketing classification tag embedding, marketing time sequence encoding, and marketing text vectorization using BERT. All marketing modal features are sent to the same structure fusion layer for processing, and the output is a marketing representation vector that is completely matched with the customer representation vector dimension, uniformly representing the marketing attributes of a single product, service, and channel combination.

[0061] (3) Interactive prediction module

[0062] It simultaneously receives customer representation vectors output by the customer encoder and marketing representation vectors output by the marketing encoder, and provides two vector matching calculation methods: vector dot product and cosine similarity; it quantifies the customer's sensitivity to the current marketing mix through vector similarity calculation, calculates the sensitivity score, and uses it as the model's original output.

[0063] exist Figure 2The underlying module provides the prediction results, while the sensitivity score output by the interactive prediction module is the final prediction result. This value directly reflects the target high-net-worth client's willingness to accept the corresponding product, service, and marketing channel combination. Subsequent business-side components, based on the sensitivity prediction results of all marketing combinations, use a 0-1 programming model with channel-specific budget constraints to solve for the optimal sequential multi-channel marketing plan, completing the construction of customized marketing plans for long-term tracking services of high-net-worth clients. Dual-tower decoupled design: Customer and marketing data are independently encoded. Customer / marketing representations can be updated separately during training and inference phases. During inference, only the marketing data to be matched needs to be updated for rapid batch scoring, adapting to the bank's massive product offerings and multi-channel marketing scenarios. Multimodal unified representation: Both encoders integrate heterogeneous financial data, fully mining the implicit needs of high-net-worth clients, significantly improving prediction accuracy compared to single structured data modeling. Universal vector interactive scoring: Using mature vector similarity calculation methods, a quantifiable sensitivity score is output, providing standardized input for the subsequent global optimization model of the marketing plan, achieving a complete technical closed loop from customer needs identification to automatic marketing plan generation.

[0064] like Figure 3 As shown, the original input is a unified customer dataset. The customer data is divided into five heterogeneous modal data according to data attributes. Each modal data is equipped with a dedicated preprocessing and encoding module to complete feature extraction. After the five modal data are independently encoded, modal features of equal length that can participate in interactive operations are output and uniformly fed into the back-end fusion layer to perform multimodal deep fusion, and finally generate a customer representation vector of unified dimension.

[0065] exist Figure 3 The top-level input includes customer data, which serves as the raw input data source for the entire customer encoder. It corresponds to the high-net-worth customer dataset in this application and contains five types of heterogeneous multimodal information: numerical data, labeled data, sequence data, marketing response data, and text data. It also includes built-in high-net-worth customer-specific asset size, asset planning goals, channel preferences, and risk preference characteristics. All data is compliantly collected from banking business systems, customer questionnaires, offline visit records, and historical marketing ledgers.

[0066] exist Figure 3 The first branch includes numerical data → standardization + MLP. The data content includes continuous numerical indicators such as customer age, total assets, monthly transaction volume, asset size, risk assessment score, single-channel marketing cost, and historical wealth management holding amount, which contain the exclusive asset quantitative characteristics of high-net-worth customers. The processing logic is as follows: first, standardization and scaling are performed to eliminate the difference in dimensions, and then non-linear feature mapping is completed through multilayer perceptron (MLP) to convert the original values ​​into a fixed-length one-dimensional numerical feature vector, which is then output to the fusion layer.

[0067] exist Figure 3The second branch includes labeled data → Embedding layer. Data content: discrete category labels such as customer gender, occupation, risk level, channel preference, asset planning type, product type, etc. Processing logic: the Embedding layer is used to express the discrete labels in low-dimensional dense vectorization, and the category labels are mapped to feature vectors of a unified dimension to realize the digital encoding of classification information and output to the fusion layer.

[0068] exist Figure 3 The third branch includes sequence data → TransformerEncoder. The data content includes long-term behavioral sequences such as customer APP browsing time series, historical product purchase time series, fund inflow and outflow time series, and long-term asset allocation change time series. The processing logic is to use the Transformer encoder to capture long-distance dependencies within the sequence, mine the long-term behavioral change patterns of customers, output the sequence feature vector carrying time series correlation information, and send it into the fusion layer.

[0069] exist Figure 3 The fourth branch includes marketing response data → time series model LSTM. The data content includes historical marketing reach records, clicks, inquiries, contracts, rejections, and other marketing feedback time series data from various channels. The processing logic is to use the LSTM long short-term memory time series model to model the dynamic response patterns of customers to different marketing actions, extract the characteristics of customers' historical marketing sensitive behaviors, and output the time series response feature vector.

[0070] exist Figure 3 The fifth branch includes text data → BERT pre-trained model. Data content includes: financial advisor interview records, customer communication texts, customer needs questionnaire texts, complaint feedback, personalized wealth planning requests, and other natural language texts. Processing logic: call the pre-trained BERT language model to complete text semantic encoding, mine the real financial needs, channel aversion tendencies, and risk demands hidden in the customer's textual expressions, and output text semantic feature vectors.

[0071] exist Figure 3 The underlying output includes a fusion layer. After the five modalities are processed by their respective dedicated encoding modules, they generate independent modal feature vectors with regular dimensions, which are all input into the fusion layer in parallel. The fusion layer processes all multimodal features in a unified manner according to the three-stage process described above: feature alignment → cross-modal attention fusion → fully connected mapping. It eliminates dimensional differences, adaptively allocates the information weights of each modality, and mines cross-modal correlations. Finally, it outputs a fixed-dimensional customer representation vector, which is provided to the subsequent prediction and interaction layer to complete the marketing sensitivity scoring.

[0072] 1) Numerical data: such as age, income, average monthly revenue, transaction frequency, RFM score, etc. 2) Tag data: gender, occupation, education level, risk level, life cycle stage, etc. 3) Sequence data: App browsing sequence, click sequence, etc., over the past 7 / 14 / 30 / 90 days. 4) Marketing response data: historical marketing response records (clicks, conversions, etc.). 5) Text data: face-to-face visits, call records, etc.

[0073] In one embodiment of this application, the customer encoder is configured with a fusion layer. The fusion layer is used to unify the dimensions of multimodal customer data and fuse multimodal information. The fusion layer sequentially performs three processing steps: feature alignment, cross-modal feature fusion, and fully connected mapping output. The feature alignment step includes: for various modal features of different dimensions, normalization and linear activation transformation are performed through an adapter; adaptive weight coefficients are set to control the contribution intensity of each modal feature, and all modal features are converted into a unified dimension. The cross-modal feature fusion step includes: using a cross-modal attention mechanism, using one modal feature as a query vector and the other modal features as key-value vectors to calculate attention; concatenating all attention information corresponding to the modality and fusing it with the original feature, and performing normalization processing to obtain the updated modal feature. The fully connected mapping output step includes: stacking all updated modal features, mining global modal correlations through a multi-head self-attention mechanism; flattening the processed features, and performing activation, random deactivation, and fully connected transformation to output a customer representation vector with a fixed dimension.

[0074] In this embodiment, the customer encoder has a built-in fusion layer to eliminate the problems of inconsistent dimensions of multimodal customer data and large differences in the contribution of information from each modality. The fusion layer strictly follows the three steps of feature alignment, cross-modal feature fusion, and fully connected mapping output.

[0075] The first step is feature alignment: For modal features of different dimensions, such as numerical, text, sequence, and label, normalization and linear activation transformation are performed through an adapter; adaptive weight coefficients are introduced to dynamically adjust the weight ratio of each type of modal feature, so as to unify all modal features to the same dimension.

[0076] The second step is cross-modal feature fusion: a cross-modal attention mechanism is adopted, using the current modal feature as the query vector and the other modal features as key and value vectors to calculate the attention weights; all cross-modal attention results corresponding to the current modality are concatenated, fused with the residuals of the original modal features, and normalized to obtain the updated single-modal features.

[0077] The third step, fully connected mapping output, involves stacking all updated modal features and mining global correlations between different modalities through multi-head self-attention. The feature matrix is ​​then flattened and sequentially processed through activation functions, random deactivation, and fully connected transformations to output a fixed-dimensional customer representation vector.

[0078] The fusion layer processes the multimodal data after processing. Because multimodal data may have different dimensions or different contributions, fusion processing is necessary. The fusion layer mainly includes:

[0079] 1) Alignment: unify the dimensions of data from different modalities.

[0080] For each modality's feature x_i (dimension d_i), after the adapter:

[0081] x'_i=LayerNorm(x_i)

[0082] x'_i = ReLU(W_i·x'_i+b_i), and adaptively control the activation intensity of this mode.

[0083] x'_i=g_i·x'_i

[0084] g_i=σ(w_g·x_i+b_g), and then multimodal data of the same dimension can be output.

[0085] 2) Multimodal data fusion output

[0086] Introducing a cross-modal attention mechanism, we use one modality to query the state of another modality and compute the attention response.

[0087] For each feature vector q=x'_i in modality i (and as Query)

[0088] And the eigenvector K=X'_j in mode j (as key and value):

[0089] Attention_ij=softmax((q·W_Q)·(K·W_K)^T / √d)·(K·W_V)

[0090] Then these attention response pairs are spliced ​​and fused together.

[0091] For modality i, all its cross-modal attention outputs are: atn_i = [Attention_i1, Attention_i2, ..., Attention_iN]

[0092] Updated modal feature: x''_i=LayerNorm(x'_i+Linear(attn_i))

[0093] 3) Multimodal data fully connected mapping output:

[0094] After processing in step 2), each feature has information on multiple modalities. These features are then stacked and combined before entering a fully connected layer, outputting a fixed-dimensional customer feature vector.

[0095] Feature stacking: X''=[x''_1,x''_2,...,x''_N]

[0096] Self-attention connection: X'''=MultiHeadSelfAttention(X'')+X''

[0097] Fully connected mapping C = Dropout(ReLU(W_final·Flatten(X''')+b_final))

[0098] Final output: C∈ℝ^{256}# Fixed-dimensional customer representation vector

[0099] In one embodiment of this application, the marketing encoder's internal fusion layer employs the same feature alignment, cross-modal fusion, and fully connected mapping processing logic as the customer encoder, outputting a marketing representation vector that matches the dimension of the customer representation vector.

[0100] In this embodiment, the internal fusion layer of the marketing encoder reuses the same processing flow as the customer encoder, sequentially performing feature alignment, cross-modal feature fusion, and fully connected mapping output.

[0101] The marketing encoder takes into account multimodal marketing data such as product attributes, campaign tags, marketing text vectors, and historical marketing conversion records. After a complete fusion process, it outputs a marketing representation vector that perfectly matches the customer representation vector dimension, ensuring that subsequent interaction layers can directly perform vector matching calculations.

[0102] The marketing encoder operates on the same principle as the customer encoder, but the input feature data differs. Specific data is as follows:

[0103] 1) Numerical data: Product attributes such as annualized rate of return, term, minimum purchase amount, etc.

[0104] 2) Tag-based data: Marketing campaign attributes, such as campaign type, incentive intensity, and distribution channels, as well as tag-based product attributes, such as product type and risk rating.

[0105] 3) Marketing response data: Historical marketing response records of the same type (clicks, conversions, etc.)

[0106] 4) Text data: marketing materials, sales scripts, etc.

[0107] In the predictive interaction layer, the principle of the dual-tower model is applied to fuse marketing features and customer features for calculation, using dot product or cosine similarity. Specifically, the customer encoder outputs a fixed-dimensional customer feature vector; the marketing encoder outputs a fixed-dimensional marketing feature vector (with the same structure as the customer encoder). In the predictive interaction layer, the principle of the dual-tower model is applied to fuse marketing features and customer features for calculation, using dot product or cosine similarity.

[0108] In one embodiment of this application, before feature fusion, the marketing encoder first uses the BERT model to vectorize the marketing text material, and the text vector and other marketing features are sent to the fusion layer.

[0109] In this embodiment, before the marketing encoder performs feature fusion, text-based marketing materials such as marketing copy, financial management scripts, and event introductions are preprocessed: the pre-trained BERT model is called to complete the text encoding and generate a fixed-length text feature vector; the text vector is concatenated with structured marketing features such as product duration, risk level, and event cost, and then sent together to the marketing encoder fusion layer for subsequent processing.

[0110] In one embodiment of this application, the predictive interaction layer uses dot product or cosine similarity to calculate a sensitivity score from the customer representation vector and the marketing representation vector.

[0111] In this embodiment, the predictive interaction layer provides two vector matching calculation modes: vector dot product calculation and cosine similarity calculation. The customer representation vector output by the customer encoder and the marketing representation vector output by the marketing encoder are input into the predictive interaction layer. A sensitivity score is obtained by selecting either calculation method. The higher the score, the stronger the target customer's willingness to accept this product and channel combination.

[0112] In one embodiment of this application, a multi-channel cross-marketing strategy is configured for a single customer. The marketing strategies are ranked according to their marketing fit, and a 0-1 programming model with channel-specific budget constraints is used to solve for the optimal sequential marketing mix, maximizing the overall average marketing fit. The expression of the 0-1 programming model is as follows: Objective function:

[0113] Constraints: x_i is not all zero, y_j is not all zero, z_k is not all zero, where, The weighted matching score for the corresponding marketing strategy; These represent 0-1 decision variables for online channels, face-to-face channels, and telephone marketing channels, respectively; cost represents the implementation cost of a single marketing strategy. The budgets are for marketing costs for online channels, face-to-face channels, and telephone marketing channels, respectively.

[0114] After obtaining the matching degree between the marketing set and the customer set, a sequential marketing mix is ​​created for each customer across multiple marketing channels, based on the matching degree. To maximize the average marketing matching degree within a given marketing cost, a 0-1 programming model is applied, with the following formula:

[0115] Maxsum(score_i*case_i*x_i)+sum(score_j*case_j*y_j)+sum(score_k*case_k*z_k)

[0116] St.sum(cost_i*x_i)<=C1, sum(cost_j*y_j)<=C2, sum(cost_k*z_k)<=C3;

[0117] x_i is not all zero, y_j is not all zero, z_k is not all zero, where score*case is the matching score of the corresponding strategy, x, y, z are the target values ​​of different channels (e.g., online channels, face-to-face channels, telephone marketing channels), and cost is the cost of this strategy.

[0118] This application embodiment targets three cross-channel marketing strategies for a single high-net-worth client: online push notifications, offline private visits, and dedicated telephone follow-ups. First, all marketing strategies are ranked from highest to lowest based on the sensitivity scores output by the model. Then, a 0-1 programming model with independent budget constraints for each channel is used to solve for the optimal sequential marketing mix, aiming to maximize the overall weighted matching degree of the client. After solving, a multi-channel ordered marketing plan is output, balancing multi-channel reach, cost control, and customer preferences.

[0119] In one embodiment of this application, the target customer is a high-net-worth customer, and the customer dataset additionally includes customer asset size, asset planning goals, channel preferences, and risk preference-related characteristics.

[0120] In addition to regular multimodal data, the corresponding customer dataset also collects exclusive features such as the customer's total asset size, medium- and long-term asset allocation planning goals, preferred channels for reaching customers, risk tolerance level, and offline service participation records. All exclusive features and basic data are input into the customer encoder to accurately identify the long-term wealth management needs of high-net-worth customers.

[0121] First, the definition of high-net-worth clients is clearly defined. In this embodiment, high-net-worth clients specifically refer to individual clients who have opened personal accounts at banks and whose investable financial assets reach a preset threshold. Investable financial assets specifically include liquid financial assets such as current deposits, time deposits, various wealth management products, public and private funds, stocks, bonds, insurance products, precious metals, and trust plans held by the client at the bank and other financial institutions. This excludes tangible fixed assets such as the client's own residential property, commercial property, vehicles, and luxury goods. In accordance with the service standards of the bank's private banking business, the threshold for investable financial assets for high-net-worth clients in this embodiment is set at no less than RMB 6 million. These clients are characterized by large asset size, diverse wealth planning needs, high demands for marketing service experience, strong demand for one-on-one offline exclusive services, and significantly higher marketing reach costs than ordinary clients. They also have extremely low tolerance for excessive or irrelevant marketing, which is one of the core reasons for the targeted optimization of the marketing plan in this application.

[0122] Secondly, a dedicated customer dataset tailored to high-net-worth clients is constructed. In this embodiment, the customer dataset includes not only conventional multimodal customer data (such as numerical data like age, average monthly transaction volume, and transaction frequency; tagged data like gender, occupation, and education level; sequence-based data like app browsing sequences and transaction behavior sequences; text-based data like face-to-face interview records and call records; and historical marketing response data), but also additionally collects and supplements high-net-worth client-specific characteristic data. Specifically, this includes the following four core characteristics. All characteristics are collected through compliant channels such as the bank's internal business systems, customer face-to-face interview records, and wealth management questionnaires to ensure the authenticity, completeness, and compliance of the data:

[0123] First, the characteristics of the client's asset size, specifically including the client's total assets, the breakdown of investable financial assets (such as the proportion of deposits, wealth management products, and equity assets), and asset change trends (the increase or decrease of assets in the past 3 months, 6 months, and 12 months), are used to accurately assess the client's wealth strength and current asset allocation status, providing a basis for subsequent product recommendations and service matching.

[0124] Second, the characteristics of asset planning objectives. This characteristic is a combination of textual and tagged data. The tagged data includes the core direction of the client's wealth planning (such as wealth preservation, wealth appreciation, children's education planning, retirement planning, estate planning, etc.), while the textual data is the personalized asset planning needs that the client clearly states in face-to-face interviews and questionnaires (such as "I hope to allocate low-risk and stable products for retirement savings" or "I plan to allocate equity assets in the next 1-2 years to increase returns"). This is used to uncover the client's long-term wealth management needs and avoid pushing products or services that are not in line with the client's plan.

[0125] Third, channel preference characteristics, specifically including customer preferred marketing outreach channels (such as online banking apps, offline private banking exclusive face-to-face visits, dedicated financial advisor telephone follow-ups, email pushes, etc.), usage frequency and feedback evaluation of each channel (such as whether they dislike telephone marketing, whether they prefer offline one-on-one communication), etc., are used to optimize marketing outreach methods, avoid channels that customers dislike, and improve the marketing experience.

[0126] Fourth, risk preference characteristics. These characteristics are collected through bank wealth management risk assessment questionnaires. Specifically, they include the customer's risk tolerance level (such as conservative, stable, balanced, growth-oriented, and aggressive), the degree of risk tolerance for various financial products (such as whether they accept principal loss and the maximum acceptable loss ratio), past investment experience in high-risk products, and feedback. This data is used to match wealth management products with appropriate risk levels, reducing ineffective marketing and customer resentment caused by product risks not matching customer preferences.

[0127] Finally, the aforementioned high-net-worth client-specific feature data, together with conventional multimodal client data, constitute a complete client dataset. After data cleaning, standardization, and other preprocessing, the dataset is uniformly input into the client encoder, where it participates in feature extraction and fusion along with other modal data. Ultimately, a client representation vector that accurately reflects the needs, preferences, and asset status of high-net-worth clients is generated. This provides core data support for subsequent marketing action sensitivity prediction and customized marketing plan generation, effectively solving the technical pain point that existing technologies cannot accurately identify the real needs of high-net-worth clients in long-term follow-up marketing services.

[0128] Similarly, it can be understood that high-net-worth clients are only a preferred embodiment of this application and are not intended to limit the scope of protection of this application.

[0129] This application embodiment also provides a marketing device 400 for target customers, such as... Figure 4 As shown, a schematic diagram of the structure of a marketing device for target customers in an embodiment of this application is provided. The marketing device 400 for target customers includes at least: a prediction model building module 410 and a prediction module 420, wherein:

[0130] In one embodiment of this application, the prediction model building module 410 is specifically used to: build a prediction model based on a customer dataset and a marketing dataset. The prediction model is used to predict the sensitivity of target customers to marketing actions. The prediction model includes a customer encoder, a marketing encoder, and a prediction interaction layer. The customer encoder is used to extract and fuse features from customer data and output a customer representation vector. The marketing encoder is used to extract and fuse features from marketing data and output a marketing representation vector. The prediction interaction layer fuses and calculates the customer representation vector and the marketing representation vector to output a sensitivity score for the corresponding marketing action.

[0131] Two types of basic datasets are pre-built: customer dataset and marketing dataset.

[0132] The target customers are the bank's high-net-worth clients. Therefore, in addition to the regular basic data, the customer dataset also includes exclusive features such as customer asset size, asset planning goals, channel preferences, and risk preferences. The customer dataset contains multimodal customer data, including numerical, labeled, sequential, historical marketing responses, and call / face-to-face interview texts. The marketing dataset stores marketing-related data such as various wealth management products, value-added services, online and offline marketing activities, marketing scripts, and promotional copy.

[0133] Cleaning, standardization, and labeling preprocessing were performed on the two types of datasets respectively: For text materials such as marketing copy and service scripts in the marketing data, the BERT pre-trained model was used to vectorize the text and generate fixed-dimensional text feature vectors, which were then merged with structured features such as product duration, minimum purchase amount, and risk level to form complete marketing data input.

[0134] Based on the processed customer and marketing datasets, a predictive model is built. This model is specifically designed to quantify the sensitivity of high-net-worth customers to different marketing actions. The model has a dual-tower structure, consisting of three main parts: a customer encoder, a marketing encoder, and a predictive interaction layer.

[0135] In one embodiment of this application, the prediction module 420 is specifically used to: input the actual customer data and marketing data to be predicted into the trained prediction model, output the prediction result of the target customer's sensitivity to marketing actions according to the prediction model, and generate a matching marketing plan based on the prediction result.

[0136] In business scenarios, real-time multimodal data of high-net-worth customers to be marketed, as well as marketing data of products and activities to be launched, are collected as inference inputs. The two sets of data are fed into the trained prediction model, which outputs the sensitivity prediction results corresponding to each marketing action. Then, the optimal sequential marketing mix is ​​solved through the aforementioned multi-channel 0-1 planning model to generate a customized marketing plan that meets customer needs, balances channels, and complies with budget constraints. Marketing outreach is carried out in the planned order.

[0137] In one embodiment of this application, the customer encoder in the prediction model building module is configured with a fusion layer. The fusion layer is used to unify the dimensions of multimodal customer data and fuse multimodal information. The fusion layer sequentially performs three processing steps: feature alignment, cross-modal feature fusion, and fully connected mapping output.

[0138] The feature alignment step includes: normalizing and linear activation transformations for various modal features of different dimensions using an adapter; setting adaptive weight coefficients to control the contribution intensity of each modal feature; and converting all modal features into a unified dimension.

[0139] The cross-modal feature fusion step includes: employing a cross-modal attention mechanism, using one modal feature as the query vector and the remaining modal features as key vectors to calculate attention; concatenating all attention information corresponding to the modality and fusing it with the original features, and performing normalization processing to obtain the updated modal features;

[0140] The fully connected mapping output step includes: stacking all updated modal features, mining global modal correlations through a multi-head self-attention mechanism; flattening the processed features, and transforming them through activation, random deactivation, and full connection to output a customer representation vector with a fixed dimension.

[0141] In one embodiment of this application, the marketing encoder's internal fusion layer in the prediction model building module adopts the same feature alignment, cross-modal fusion, and fully connected mapping processing logic as the customer encoder, and outputs a marketing representation vector that matches the dimension of the customer representation vector.

[0142] In one embodiment of this application, before feature fusion, the marketing encoder in the prediction model building module first uses the BERT model to vectorize the marketing text materials, and the text vectors and other marketing features are sent to the fusion layer.

[0143] In one embodiment of this application, the prediction interaction layer in the prediction model building module uses dot product or cosine similarity to calculate the sensitivity score of the customer representation vector and the marketing representation vector.

[0144] In one embodiment of this application, the prediction model building module configures multi-channel cross-marketing strategies for a single customer, sorts the marketing strategies according to their marketing match degree, and uses a 0-1 programming model with channel budget constraints to solve for the optimal sequential marketing mix, thereby maximizing the overall average marketing match degree.

[0145] The 0-1 programming model expression is as follows:

[0146] Objective function:

[0147]

[0148] Constraints:

[0149]

[0150] x_i is not all zero, y_j is not all zero, z_k is not all zero.

[0151] in, The weighted matching score for the corresponding marketing strategy; These represent 0-1 decision variables for online channels, face-to-face channels, and telephone marketing channels, respectively; cost represents the implementation cost of a single marketing strategy. The budgets are for marketing costs for online channels, face-to-face channels, and telephone marketing channels, respectively.

[0152] In one embodiment of this application, the target customer is a high-net-worth customer, and the customer dataset additionally includes customer asset size, asset planning goals, channel preferences, and risk preference-related characteristics.

[0153] It is understood that the marketing device for the target customer described above can realize each step of the marketing method for the target customer provided in the foregoing embodiments. The relevant explanations of the marketing method for the target customer are applicable to the marketing device for the target customer, and will not be repeated here.

[0154] Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Please refer to it. Figure 5 At the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations.

[0155] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0156] Memory is used to store programs. Specifically, programs may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0157] The processor reads the corresponding computer program from non-volatile memory into main memory and then executes it, forming a marketing device for the target customer at the logical level. The processor executes the program stored in memory and specifically performs the following operations:

[0158] Based on customer and marketing datasets, a predictive model is built to predict the sensitivity of target customers to marketing actions. The predictive model includes a customer encoder, a marketing encoder, and a predictive interaction layer. The customer encoder extracts and fuses features from customer data, outputting a customer representation vector. The marketing encoder extracts and fuses features from marketing data, outputting a marketing representation vector. The predictive interaction layer fuses the customer representation vector and the marketing representation vector to calculate and output a sensitivity score for the corresponding marketing action.

[0159] The actual customer data and marketing data to be predicted are input into the trained prediction model. Based on the prediction model, the model outputs the prediction results of the target customer's sensitivity to marketing actions, and a matching marketing plan is generated based on the prediction results.

[0160] The above is as stated in this application. Figure 1The method executed by the marketing device targeting the customer disclosed in the illustrated embodiment can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0161] The electronic device can also perform Figure 1 Methods for implementing marketing devices targeting specific customers, and methods for achieving marketing devices targeting specific customers in [the context of marketing]. Figure 1 The functions of the embodiments shown are not described again in this application.

[0162] This application also proposes a computer-readable storage medium that stores one or more programs, the programs including instructions that, when executed by an electronic device including multiple applications, enable the electronic device to perform... Figure 1 The method executed by the marketing device targeting the customer in the illustrated embodiment is specifically used to perform:

[0163] Based on customer and marketing datasets, a predictive model is built to predict the sensitivity of target customers to marketing actions. The predictive model includes a customer encoder, a marketing encoder, and a predictive interaction layer. The customer encoder extracts and fuses features from customer data, outputting a customer representation vector. The marketing encoder extracts and fuses features from marketing data, outputting a marketing representation vector. The predictive interaction layer fuses the customer representation vector and the marketing representation vector to calculate and output a sensitivity score for the corresponding marketing action.

[0164] The actual customer data and marketing data to be predicted are input into the trained prediction model. Based on the prediction model, the model outputs the prediction results of the target customer's sensitivity to marketing actions, and a matching marketing plan is generated based on the prediction results.

[0165] Those skilled in the art will understand 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.

[0166] 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.

[0167] 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.

[0168] 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.

[0169] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0170] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0171] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0172] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0173] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application 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.

[0174] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A marketing method targeting specific customers, wherein, The method includes: Based on customer and marketing datasets, a predictive model is built to predict the sensitivity of target customers to marketing actions. The predictive model includes a customer encoder, a marketing encoder, and a predictive interaction layer. The customer encoder extracts and fuses features from customer data, outputting a customer representation vector. The marketing encoder extracts and fuses features from marketing data, outputting a marketing representation vector. The predictive interaction layer fuses the customer representation vector and the marketing representation vector to calculate and output a sensitivity score for the corresponding marketing action. The actual customer data and marketing data to be predicted are input into the trained prediction model. Based on the prediction model, the model outputs the prediction results of the target customer's sensitivity to marketing actions, and a matching marketing plan is generated based on the prediction results.

2. The marketing method for target customers according to claim 1, characterized in that, The customer encoder is configured with a fusion layer, which is used to unify the dimensions of multimodal customer data and fuse multimodal information. The fusion layer sequentially performs three processing steps: feature alignment, cross-modal feature fusion, and fully connected mapping output. The feature alignment step includes: normalizing and linear activation transformations for various modal features of different dimensions using an adapter; setting adaptive weight coefficients to control the contribution intensity of each modal feature; and converting all modal features into a unified dimension. The cross-modal feature fusion step includes: employing a cross-modal attention mechanism, using one modal feature as the query vector and the remaining modal features as key vectors to calculate attention; concatenating all attention information corresponding to the modality and fusing it with the original features, and performing normalization processing to obtain the updated modal features; The fully connected mapping output step includes: stacking all updated modal features, mining global modal correlations through a multi-head self-attention mechanism; flattening the processed features, and transforming them through activation, random deactivation, and full connection to output a customer representation vector with a fixed dimension.

3. The marketing method for target customers according to claim 1, characterized in that, The marketing encoder's internal fusion layer uses the same feature alignment, cross-modal fusion, and fully connected mapping processing logic as the customer encoder, outputting a marketing representation vector that matches the dimension of the customer representation vector.

4. The marketing method for target customers according to claim 3, characterized in that, Before feature fusion, the marketing encoder first uses the BERT model to vectorize the marketing text materials, and then sends the text vectors and other marketing features into the fusion layer.

5. The marketing method for target customers according to claim 1, characterized in that, The predictive interaction layer uses dot product or cosine similarity to calculate sensitivity scores from customer representation vectors and marketing representation vectors.

6. The marketing method for target customers according to claim 1, characterized in that, Configure multi-channel cross-marketing strategies for individual customers, rank marketing strategies according to their marketing fit, and use a 0-1 programming model with budget constraints for each channel to solve for the optimal sequential marketing mix, so as to maximize the overall average marketing fit. The 0-1 programming model expression is as follows: Objective function: Constraints: x_i is not all zero, y_j is not all zero, z_k is not all zero. in, The weighted matching score for the corresponding marketing strategy; These represent 0-1 decision variables for online channels, face-to-face channels, and telephone marketing channels, respectively; cost represents the implementation cost of a single marketing strategy. The budgets are for marketing costs for online channels, face-to-face channels, and telephone marketing channels, respectively.

7. The marketing method for target customers according to claim 1, characterized in that, The target customers are high-net-worth individuals, and the customer dataset additionally includes characteristics related to customer asset size, asset planning goals, channel preferences, and risk appetite.

8. A marketing device targeting specific customers, wherein, The device includes: A prediction model building module is used to build a prediction model based on customer and marketing datasets. This prediction model is used to predict the sensitivity of target customers to marketing actions. The prediction model includes a customer encoder, a marketing encoder, and a prediction interaction layer. The customer encoder extracts and fuses features from customer data, outputting a customer representation vector. The marketing encoder extracts and fuses features from marketing data, outputting a marketing representation vector. The prediction interaction layer fuses the customer representation vector and the marketing representation vector, outputting a sensitivity score for the corresponding marketing action. The prediction module is used to input the actual customer data and marketing data to be predicted into the trained prediction model, output the prediction result of the target customer's sensitivity to marketing actions based on the prediction model, and generate a matching marketing plan based on the prediction result.

9. An electronic device, comprising: processor; as well as A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium storing one or more programs, which, when executed by an electronic device including a plurality of applications, cause the electronic device to perform the method of any one of claims 1 to 7.

Images