A new type of digital marketing collaboration method and system suitable for small and medium-sized banks

Abstract

The application discloses a novel digital marketing collaboration method and system suitable for small and medium-sized banks, relates to the field of marketing collaboration, and comprises the following steps: acquiring user attributes of target users; constructing user portraits of the target users according to the user attributes; analyzing demand information through a preset CNN-BiLSTM model to determine demand intentions of the target users; verifying matching degrees of the demand intentions according to the user portraits; converting the demand intentions into marketing task indexes; acquiring all department resource information; determining evaluation indexes of all departments according to the department resource information and the marketing task indexes in combination with a game theory weighting and a multi-criteria compromise solution sorting; comparing the evaluation indexes with a preset evaluation threshold; and if the evaluation indexes are smaller than the preset evaluation threshold, using a preset screening model to determine a marketing task allocation department. The application can effectively improve the marketing efficiency of small and medium-sized banks.

Description

Technical Field

[0001] This application relates to the field of marketing collaboration, and in particular to a novel digital marketing collaboration method and system suitable for small and medium-sized banks. Background Technology

[0002] With the rapid development of the mobile internet, digitalization has become a core demand for development across various industries. This nationwide trend of digitalization also poses a significant challenge to the traditional marketing models of commercial banks. Traditional marketing primarily relies on offline branch channels, such as bank counters where staff directly interact with customers, providing detailed explanations of financial products based on their specific circumstances, answering questions on the spot, and promoting relevant financial products. However, this model is limited by the operating hours and geographical location of bank branches, resulting in a relatively narrow customer reach. This makes it difficult for banks to fully tap into their potential customer base, thereby impacting customer service experience and satisfaction. Summary of the Invention

[0003] This application provides a novel digital marketing collaboration method suitable for small and medium-sized banks, which can improve the marketing efficiency of small and medium-sized banks.

[0004] To achieve the above objectives, the embodiments of this application adopt the following technical solutions:

[0005] Firstly, a novel digital marketing collaboration method suitable for small and medium-sized banks is provided, which includes:

[0006] In response to receiving the target user's request information, the user attributes of the target user are obtained;

[0007] Construct a user profile of the target user based on the user attributes;

[0008] The target user's intent is determined by analyzing the demand information using a pre-defined CNN-BiLSTM model.

[0009] The matching degree of the demand intent is verified based on the user profile.

[0010] If the stated demand intent passes the matching degree verification, the stated demand intent will be converted into marketing task indicators;

[0011] Obtain resource information for all departments;

[0012] The evaluation index for all departments is determined by combining the departmental resource information and the marketing task indicators with game theory weighting and multi-criteria compromise solution ranking.

[0013] The evaluation index is compared with a preset evaluation threshold. If the evaluation index is greater than the preset evaluation threshold, the marketing task allocation department is determined.

[0014] If the evaluation index is less than the preset evaluation threshold, the marketing task allocation department is determined using a preset screening model.

[0015] In one possible implementation of the first aspect, the user attributes include permanent address, user age, and product consumption behavior, and the step of constructing a user profile of the target user based on the user attributes includes the following steps:

[0016] Use GIS technology to obtain the regional consumption activity value corresponding to the permanent address;

[0017] The target user's age is input into a preset consumption capacity prediction model to determine the target user's consumption capacity assessment value;

[0018] Determine the regional preferred products corresponding to the permanent address from the preset product preference matrix;

[0019] The product demand value of the target user is obtained by combining the regional product preferences and the product consumption behavior using a preset decision tree model;

[0020] Construct a feature vector from the regional consumption activity value, the consumption capacity assessment value, and the product demand value;

[0021] Density clustering algorithm is used to perform cluster analysis on the feature vectors to determine the consumption characteristic categories of the target users;

[0022] A user profile of the target user is constructed based on the aforementioned consumption characteristic categories.

[0023] In one possible implementation of the first aspect, the CNN-BiLSTM model includes a first convolutional layer, a second convolutional layer, a max pooling layer, a first multilayer perceptron, and a second multilayer perceptron. The step of analyzing the demand information using the preset CNN-BiLSTM model to determine the target user's demand intent includes the following steps:

[0024] The requirement information is converted into word embedding vectors;

[0025] The word embedding vectors are input into the first convolutional layer and the second convolutional layer respectively to output a first feature map and a second feature map. The first feature map and the second feature map are then filtered using the max pooling layer to determine the first key feature and the second key feature, wherein the convolution kernel of the first convolutional layer is smaller than the convolution kernel of the second convolutional layer.

[0026] The first embedded feature and the second embedded feature are generated by performing a nonlinear transformation on the first key feature and the second key feature using the first multilayer perceptron.

[0027] The BiLSTM module is used to capture the temporal correlation of sentences in the requirement information;

[0028] The sentence temporal correlation, the first embedding feature, and the second embedding feature are used to construct a multi-layer semantic feature;

[0029] The multi-layer semantic features are input into the second multi-layer perceptron to determine the target user's demand intention;

[0030] The first multilayer perceptron is used to process key features, and the second multilayer perceptron is used to process multilayer semantic features.

[0031] In one possible implementation of the first aspect, the matching degree verification of the demand intent based on the user profile includes the following steps:

[0032] Select the corresponding matching expression based on the user profile;

[0033] The named entities of the user profile are determined based on the matching expression, and the entity category of the named entities is determined.

[0034] Based on the entity category, the matching expression, and the named entity, the target attribute of each user profile is determined by traversing each user profile.

[0035] The target attribute and demand intention features are extracted using a pre-defined language training model.

[0036] Calculate the feature matching degree between the target attribute features and the demand intent features, and verify the matching degree of the demand intent based on the feature matching degree.

[0037] In one possible implementation of the first aspect, selecting the corresponding matching expression based on the user profile includes the following steps:

[0038] Select target information of the target user from the user profile, and select a first matching expression from the candidate matching expressions in the preset database based on the target information;

[0039] The number of matches for each user tag in the user profile is determined by iterating through each user tag in the user profile using the first matching expression;

[0040] Determine the number of rule tags corresponding to the first matching expression;

[0041] The first matching similarity is calculated using the number of matches and the number of rule tags;

[0042] Iterate through each of the candidate matching expressions in the preset database and calculate the matching similarity of each candidate matching expression;

[0043] The candidate matching expression with the highest matching similarity is used as the matching expression.

[0044] In one possible implementation of the first aspect, determining the evaluation index of all departments based on the departmental resource information and the marketing task indicators, combined with game theory weighting and multi-criteria compromise solutions, includes the following steps:

[0045] Principal component analysis was used to select departmental resource indicators based on the departmental resource information.

[0046] A decision matrix is ​​constructed by combining the departmental resource indicators and the marketing task indicators, wherein the rows of the decision matrix are evaluation objects and the columns of the decision matrix are evaluation indicators.

[0047] Standardized indicators are obtained by standardizing each element in the decision matrix, and positive and negative indicators are determined based on preset business rules.

[0048] The variability and conflict of indicators are calculated based on the preset variability formula and the preset conflict formula, respectively.

[0049] The amount of information about an indicator is calculated by combining the variability and conflict of the indicators.

[0050] The first weight vector is calculated using a preset weight formula based on the amount of indicator information.

[0051] Calculate the proportion of each evaluation indicator in the decision matrix in the evaluation object, and calculate the indicator information entropy based on the indicator proportion using a preset information entropy formula;

[0052] The difference coefficient of each evaluation indicator is calculated based on the indicator information entropy of each evaluation indicator, and the second weight vector of each evaluation indicator is calculated based on the difference coefficient.

[0053] The optimal weights are determined using game theory by combining the first weight vector and the second weight vector.

[0054] The evaluation index for all departments is determined by ranking them using a multi-criteria compromise solution based on the optimal weights.

[0055] In one possible implementation of the first aspect, determining the evaluation index of all departments by ranking them using the multi-criteria compromise solution based on the optimal weights includes the following steps:

[0056] By iterating through each of the positive and negative indicators, the maximum positive indicator and the minimum negative indicator are determined.

[0057] Based on the aforementioned calculation of the comprehensive index gap value for each matrix element using the preset index gap formula, combined with the maximum positive index, the minimum negative index, each standardized index, and the optimal weight;

[0058] By iterating through the comprehensive gap values ​​of each of the aforementioned indicators, the maximum and minimum indicator gap values ​​are determined.

[0059] The optimal solution distance and the worst solution distance are calculated based on each of the standardized indices, the maximum positive index value, and the minimum negative index value, respectively.

[0060] The proximity is calculated by the distance between the optimal solution and the distance between the worst solution, and an evaluation index is calculated based on the proximity, the maximum index gap value, the minimum index gap value, and the comprehensive index gap value.

[0061] The departments responsible for marketing task allocation are ranked based on the evaluation index.

[0062] In one possible implementation of the first aspect, determining the marketing task allocation department using a preset screening model includes:

[0063] Using a preset screening model, the departmental resource information and the marketing task indicators are matched to obtain basic matching results, quantity matching results, and energy efficiency matching results;

[0064] The marketing task allocation department is determined based on the basic matching results, the quantity matching results, and the energy efficiency matching results.

[0065] Secondly, this application provides a machine-readable storage medium storing instructions for causing a machine to execute the aforementioned novel digital marketing collaboration method applicable to small and medium-sized banks.

[0066] Thirdly, this application provides a novel digital marketing collaboration system suitable for small and medium-sized banks, comprising:

[0067] The memory is configured to store instructions; and

[0068] The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the novel digital marketing collaboration method applicable to small and medium-sized banks described above.

[0069] By employing the aforementioned technical solutions, user profiles are constructed by acquiring target user attributes. A CNN-BiLSTM model is then used to analyze demand information and determine user intent, enabling a more accurate understanding of users' true needs and more precise service delivery, thus improving user experience. The matching degree of user intent is verified based on the user profile. Verified intents are then transformed into marketing task indicators, ensuring a high degree of alignment between marketing tasks and user needs. This avoids resource waste and user aversion caused by indiscriminate marketing, improving the targeting and effectiveness of marketing. By acquiring resource information from all departments and combining game theory weighting and multi-criteria compromise solutions to determine the evaluation index for each department, and comparing the evaluation index with a preset evaluation threshold, the departments responsible for allocating marketing tasks are determined. This approach fully considers the resource status, capabilities, and strengths of each department, achieving rational resource allocation. When the evaluation index is lower than the preset threshold, a preset screening model is used to determine the departments responsible for allocating marketing tasks. This dynamic adjustment mechanism allows for flexible responses to actual situations, ensuring timely and effective processing of marketing tasks and enhancing the bank's marketing effectiveness and competitiveness.

[0070] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0071] Figure 1 A flowchart illustrating a novel digital marketing collaboration method applicable to small and medium-sized banks, provided as an embodiment of this application;

[0072] Figure 2 A flowchart illustrating the process of determining the intent of a requirement is provided in this application embodiment;

[0073] Figure 3 This is a schematic diagram illustrating a process for constructing a user profile, as provided in an embodiment of this application. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0075] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0076] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0077] Figure 1 The illustration schematically depicts a process diagram of a novel digital marketing collaboration method suitable for small and medium-sized banks according to an embodiment of this application. Figure 1 As shown in the figure, this application provides a novel digital marketing collaboration method suitable for small and medium-sized banks, which may include the following steps.

[0078] S110. In response to receiving the target user's request information, obtain the target user's user attributes;

[0079] S120. Construct a user profile of the target user based on user attributes;

[0080] S130. Analyze the demand information using a pre-set CNN-BiLSTM model to determine the target user's demand intent;

[0081] S140. Verify the matching degree of the user's intent based on the user profile;

[0082] S150. If the demand intent passes the matching degree verification, the demand intent will be converted into a marketing task indicator.

[0083] S160. Obtain resource information for all departments;

[0084] S170. Based on departmental resource information and marketing task indicators, combined with game theory weighting and multi-criteria compromise solutions, determine the evaluation index of all departments.

[0085] S180. Compare the evaluation index with the preset evaluation threshold. If the evaluation index is greater than the preset evaluation threshold, determine the marketing task allocation department.

[0086] S190. If the evaluation index is less than the preset evaluation threshold, the marketing task allocation department is determined using the preset screening model.

[0087] First, in response to receiving the target user's request information, the bank system acquires the target user's user attributes. Upon receiving the request information from the target user, the system immediately initiates relevant processes to collect and organize the target user's related data; this data constitutes the user attributes. Request information can be proactively initiated by the user, such as when a user inquires about a certain type of financial product through the bank's online platform, asking about the returns of wealth management products or the application requirements for loan products. Alternatively, it can be passively detected by the bank through data analysis, such as when a user frequently browses the page of a specific financial product, leading the system to determine that the user may have a potential need for that product. User attributes can include basic information, financial behavior data, credit information, and preference information.

[0088] Secondly, user profiles of target users are constructed based on user attributes. These profiles present key information such as user characteristics, needs, and behavioral patterns in an intuitive, visual, and structured manner, forming a comprehensive and three-dimensional description of the target user. Specifically, this involves summarizing previously acquired basic information, financial behavior data, credit information (credit scores, overdue records, etc.), preference information, and other multi-dimensional attribute information. The integrated attribute information is then analyzed in depth to uncover key user characteristics and behavioral patterns. For example, analyzing a user's transaction records reveals a high proportion of monthly spending on dining and entertainment, indicating high consumer activity and a demand for leisure and entertainment. The extracted information is presented in a structured form, typically including multiple dimensions such as basic user characteristics, consumption habits, financial needs, and risk preferences, forming a clear and complete user description. Specifically, user attributes can be categorized according to different dimensions. Based on the categorized characteristics, data analysis techniques and algorithms are used to construct user profile models. Alternatively, pre-trained neural network models can be used to create a tagging system for user information, assigning corresponding labels to each characteristic, such as "high-spending group" or "conservative investor." By using actual business data and user feedback, the constructed user profiles are verified and optimized to ensure that the profiles accurately reflect the real situation of the target users.

[0089] This method analyzes demand information using a pre-designed CNN-BiLSTM model to determine the target user's intent. This involves using a pre-designed deep learning model that integrates Convolutional Neural Networks (CNNs) and Bidirectional Long Short-Term Memory (BiLSTM) networks to deeply analyze the demand information presented by the target user, thereby accurately identifying the user's true intention. CNNs excel at extracting local features from input data. When processing demand information text, they can identify key words and phrases such as "financial management," "loan," and "credit card," which are associated with specific demand intentions. BiLSTMs can capture contextual information in the text and analyze the preceding and following information of the current word, thus gaining a more comprehensive understanding of the user's intent. For example, in the sentence "I have a history of overdue payments and now I want to apply for a small loan," BiLSTM can combine the contextual information of "having a history of overdue payments" and "now wanting to apply for a small loan" to more accurately determine that the user's intent might be "loan application - a small loan considering creditworthiness." The pre-designed CNN-BiLSTM model indicates that it has undergone extensive training and optimization. During training, a large number of labeled demand information samples were used, allowing the model to learn how to extract key features from this information and accurately determine the user's intent. For example, training data might contain user input such as "I want to learn about stable-yield financial products," along with the corresponding label "Financial product consultation - conservative." When receiving demand information from a target user, the model processes it. The demand information can be in text or other forms, and the model analyzes the input text word by word, extracting semantic, grammatical, and contextual information. Through in-depth analysis of the demand information, the model determines the user's intent based on learned knowledge and patterns. For example, if a user inputs "I want to apply for a credit card with a high credit limit and no annual fee," the model, after analysis, determines that the user's intent is "credit card application - high credit limit and no annual fee."

[0090] The matching degree verification of demand intent is performed based on user profiles. This involves comparing and analyzing the target user's demand intent derived from model analysis with the user profile built based on user attributes to determine the degree of fit. This verifies whether the demand intent truly reflects the user's potential needs and characteristics, avoiding inaccurate marketing due to information bias or misunderstanding. A user profile is a comprehensive description of the target user's multi-dimensional characteristics, covering basic information, financial behavior, credit status, preferences, etc.; demand intent is the goal or expectation the user currently wants to achieve, extracted from the user's demand information. Matching degree verification examines whether the demand intent aligns with the user characteristics and demand patterns presented in the user profile. During verification, the demand intent is compared with various dimensions in the user profile, and similarity is calculated for verification. Key features related to the demand intent are extracted from the user profile, such as the user's investment preferences, risk tolerance, and consumption habits. Simultaneously, the specific content of the demand intent is clarified, such as the type, amount, and term of the loan the user wants to apply for, or the type of financial product the user wants to purchase. The demand intent is compared with the key features of the user profile according to a set matching similarity. Automated matching and comparison can be achieved through algorithm development or the use of professional data analysis tools. Based on the matching comparison results, the degree of matching between the demand intent and the user profile is determined.

[0091] If the demand intent passes the matching verification, it is transformed into marketing task indicators. When it is determined that the user's demand intent highly matches their user profile, the bank will convert the abstract demand intent into specific, quantifiable, and actionable marketing task indicators. This allows each department to clarify its work direction and goals, promoting the effective implementation of marketing activities. The core marketing objectives are determined based on the demand intent. For example, if the demand intent is that a user wants to purchase a stable-yield financial product, then the marketing objective might be to successfully recommend and sell that financial product to the user. Marketing objectives are broken down into specific, quantifiable task indicators. These indicators may include sales volume, sales revenue, customer conversion rate, customer satisfaction, etc. Reasonable time nodes are set for each task indicator to monitor and evaluate marketing efforts.

[0092] The evaluation index for each department is determined by acquiring resource information from all departments and combining this information with marketing task indicators, along with game theory weighting and multi-criteria compromise solution ranking. Acquiring departmental resource information requires a comprehensive collection of resources possessed by each department within the bank, including human resources, financial resources, information resources, and technological resources, which can be queried using the bank's internal information management system. Game theory weighting considers the mutual influence and interests of each department in the marketing task, determining the weight of each department in the evaluation through game analysis. Multi-criteria compromise solution ranking is a multi-criteria decision-making method that comprehensively considers multiple decision criteria, seeking a compromise between maximizing group utility and minimizing individual regret to determine the optimal solution. Combining departmental resource information with marketing task indicators, and using these two methods, the evaluation index for each department can be determined by comprehensively considering factors such as resource contribution and collaboration ability in achieving the marketing task indicators. The evaluation index is a quantitative assessment of each department's performance and contribution in the marketing task, used to measure the department's work effectiveness and value.

[0093] The evaluation index is compared with a preset evaluation threshold. If the evaluation index is greater than the preset threshold, the department responsible for marketing tasks is determined. The evaluation index is a quantitative indicator calculated based on the methods described above, comprehensively considering departmental resource information and marketing task indicators, combined with game theory weighting and multi-criteria compromise solution ranking. It is used to measure the comprehensive capabilities and potential contributions of each department in marketing tasks. The preset evaluation threshold is a standard value set by the bank based on its own strategic goals, past experience, and the specific requirements of marketing tasks. Comparing the evaluation index with the preset evaluation threshold determines whether the department's overall performance meets the bank's established standards. If a department's evaluation index is greater than the preset evaluation threshold, it indicates that the department can better meet the needs of marketing tasks in terms of resources and capabilities, and has the ability and advantages to undertake the corresponding marketing tasks. Therefore, it can be determined as the department responsible for marketing task allocation, and the bank can assign marketing tasks to this department for execution.

[0094] If the evaluation index is less than a preset evaluation threshold, the pre-designed screening model is used to determine the department responsible for marketing tasks. This means that when the department evaluation index calculated using the method described above does not reach the preset evaluation threshold set by the bank, the bank will activate a pre-designed screening model. Based on the rules and standards set by this model, the bank will select the department that is more suitable to undertake the marketing tasks from all departments whose evaluation indices are less than the preset threshold. In this embodiment, the pre-designed screening model can be a neural network model. First, the screening model will define a series of dimensions and specific indicators for evaluating departments, determine the basic matching results, quantity matching results, and efficiency matching results, assign corresponding weights to each matching result, and select the department with the highest comprehensive score as the department responsible for marketing tasks based on the comprehensive ranking results. If multiple departments have similar comprehensive scores, further decisions can be made by considering the specific requirements of the marketing tasks and the bank's strategic objectives.

[0095] By acquiring user attributes to construct user profiles and using a CNN-BiLSTM model to analyze demand information and determine demand intent, we can more accurately grasp users' true needs, provide more precise services, and improve user experience. Based on the user profiles, we verify the matching degree of demand intent, transforming verified demand intents into marketing task indicators. This ensures a high degree of alignment between marketing tasks and user needs, avoiding resource waste and user aversion caused by blind marketing, and improving the targeting and effectiveness of marketing. We acquire resource information from all departments and, using game theory weighting and multi-criteria compromise solutions, determine the evaluation index for each department. The comparison between the evaluation index and a preset evaluation threshold determines the departments responsible for allocating marketing tasks. This fully considers the resource status, capabilities, and strengths of each department, achieving rational resource allocation. When the evaluation index is lower than the preset evaluation threshold, a preset screening model is used to determine the departments responsible for allocating marketing tasks. This dynamic adjustment mechanism allows for flexible responses to actual situations, ensuring timely and effective processing of marketing tasks, improving the bank's marketing effectiveness and competitiveness.

[0096] In one embodiment of this invention, user attributes include permanent address, user age, and product consumption behavior. Constructing a user profile of the target user based on these user attributes includes the following steps:

[0097] S210. Use GIS technology to obtain the regional consumption activity value corresponding to the permanent address;

[0098] S220. Input the target user's age into the preset consumption capacity prediction model to determine the target user's consumption capacity assessment value;

[0099] S230. Determine the regional preferred products corresponding to the permanent address in the preset product preference matrix;

[0100] S240. Combine regional product preferences and product consumption behavior to obtain the product demand value of the target user using a pre-set decision tree model;

[0101] S250: Construct a feature vector from regional consumption activity value, consumption capacity assessment value, and product demand value;

[0102] S260. Use density clustering algorithm to perform cluster analysis on feature vectors to determine the consumption feature categories of target users;

[0103] S270. Construct user profiles for target users based on consumption characteristic categories.

[0104] Figure 3 This is a flowchart illustrating a process for generating a user profile, provided in an embodiment of this application. Specifically, the user profile is determined by the target user's regional consumption activity value, consumption capacity assessment value, and product demand value.

[0105] In this embodiment, the permanent address refers to the user's place of residence, used to analyze regional consumption characteristics; the user's age refers to the user's actual age, used to assess spending power and preferences; and product consumption behavior refers to the user's consumption records for different products, reflecting their consumption habits. First, GIS technology is used to obtain the regional consumption activity value corresponding to the permanent address. In this embodiment, the regional consumption activity value is a quantitative value reflecting the intensity of consumption activity within a specific geographical area. The calculation of the regional consumption activity value can be achieved based on transaction data, pedestrian flow data, and POI data. Specifically, the number of transactions and transaction amount for each area within a preset time period are obtained to calculate transaction activity; pedestrian flow activity for each area within a preset time period is calculated based on pedestrian flow data; the distribution and number of different consumption venues are determined based on POI data to calculate POI activity; and the regional consumption activity value is calculated by combining transaction activity, pedestrian flow activity, and POI activity using a weighted summation method. Subsequently, a map visualization tool is used to display the calculated regional consumption activity value on a map, thereby obtaining the regional consumption activity value corresponding to the permanent address.

[0106] The target user's age is input into a preset spending power prediction model to determine the target user's spending power assessment value. In this embodiment, the preset spending power prediction model can be a machine learning model trained based on historical data, used to predict the user's spending power. The target user's age data is input into the trained model, and the model outputs a predicted value, i.e., the user's spending power assessment value, based on the input age and the learned mapping relationship. The spending power assessment value can be a specific spending amount, spending power level, or other indicators that can reflect the user's spending strength.

[0107] In this embodiment, a two-dimensional table is used to determine the regional product preferences corresponding to a user's permanent address within a pre-defined product preference matrix. Rows represent different regions, and columns represent different product types. Each element in the matrix indicates the degree of preference for a particular product within that region. By matching a user's permanent address with this matrix, the product types typically preferred by users in that region can be identified. The permanent address is extracted from the user's basic information and matched with the regional information in the regional product preference matrix to determine the user's region. Based on the matched regional information, the corresponding row is located in the regional product preference matrix, and the product types with higher preference values ​​in that row are identified. These product types represent the products typically preferred by users in that region.

[0108] Combining regional product preferences and product consumption behavior, a pre-set decision tree model is used to obtain the product demand value of the target user. In this embodiment, the pre-trained decision tree model is a machine learning model to predict the user's interest level or demand value for different products. Product consumption behavior refers to the user's past consumption records and behavioral patterns for various products, including purchase frequency, consumption amount, and purchase time. Product demand value is a quantitative indicator of the user's demand for a specific product, reflecting the intensity of the user's potential demand for a product in the future. For the target user, their corresponding feature data is collected and input into the trained decision tree model. The model outputs the user's potential demand value or demand level for various products. Based on the predicted product demand value, personalized product recommendations are provided to the user. Targeted marketing strategies can be formulated based on the level of demand, such as pushing promotional information for relevant products to users with high demand values.

[0109] A feature vector is constructed by combining regional consumption activity value, consumption capacity assessment value, and product demand value into a single vector to comprehensively characterize user consumption features. First, the data is preprocessed, and Z-Score standardization is used to map each feature to a uniform scale. The three processed features are then combined in a fixed order to form a feature vector, such as X=[x1, x2, x3], where x1, x2, and x3 represent the values ​​of the three indicators mentioned above, respectively, which can comprehensively describe user consumption behavior and needs.

[0110] Density clustering is used to cluster feature vectors to determine the consumption characteristic categories of target users. Density clustering is a clustering method based on data density. In this embodiment, consumption characteristic categories refer to the target users' preferences in consumption behavior, spending power, and product demand. The density clustering algorithm determines the consumption characteristic categories of target users by calculating the distance between each sample point and other sample points, using methods such as Euclidean distance. The distance matrix reflects the similarity between sample points. A suitable density parameter is selected, and core points are identified based on this parameter. Then, starting from the core points, the clustering region is gradually expanded until all connected high-density regions are covered. Based on the clustering results, each cluster is assigned a category label, reflecting the similarity of users' consumption behavior, i.e., the consumption characteristic categories.

[0111] Building user profiles based on consumption characteristic categories refers to creating a comprehensive and representative personalized description for each target user based on common characteristics in their consumption behavior, preferences, and capabilities. In other words, it's about constructing user profiles based on consumption characteristic categories to achieve precise marketing and personalized services. In practical applications, the first step is to analyze the user's feature vectors using density clustering algorithms, dividing users into different categories based on their consumption characteristics, such as "high-activity, high-spending-capacity" or "strong regional preference." Next, combining category characteristics with the user's basic attributes and consumption habits, a profile is formed that includes multi-dimensional information such as age, gender, occupation, consumption preferences, and lifestyle habits. For example, for the "high-activity, high-spending-capacity" category, the user profile might be described as "a young working professional aged 25-35 with a relatively high monthly income, frequently shopping in high-end commercial areas, and preferring branded products and high-quality services."

[0112] Another implementation method in this embodiment also includes:

[0113] When a target user changes their permanent address, GIS technology is used to obtain the migration consumption activity value corresponding to the target user's new address.

[0114] The migration index is calculated using a preset consumption migration index formula by combining regional consumption activity value and migrating consumption activity value;

[0115] Input the target user's current age into the preset spending power cycle curve to determine the target user's current spending power value;

[0116] Determine the region-preferred products corresponding to the migration address from the preset product preference matrix;

[0117] In the preset region-product weight matrix, determine the first weight value of the regional preference product corresponding to the migration address and the second weight value of the regional preference product corresponding to the permanent residence address;

[0118] By combining the regional product preferences and product consumption behavior corresponding to the migration address, the migration product demand value of the target user is obtained using a pre-set decision tree model;

[0119] The product demand transition value is calculated by subtracting the product of the migrated product demand value and the second weight value from the product of the product demand value and the first weight value.

[0120] A migration feature vector is constructed based on the migration index, migration product demand value, and product demand transition value.

[0121] Based on the migration feature vector, determine the migration consumption feature category of the target user and update the user profile of the target user.

[0122] When a target user changes their permanent address, GIS technology is used to obtain the migration consumption activity value corresponding to the new address. In this embodiment, the migration consumption activity value refers to the level of consumption activity at the new address determined by GIS technology when the target user's permanent address changes. Consumption activity level is a quantified value of the intensity of consumption activity within a specific geographical area, which can be quantified through bank card spending. In this embodiment, the migration address is defined as the address where the target user's residential address remains unchanged within a preset time period after changing their permanent address, and this address is used to update the information. Furthermore, short-term travel or short trips can be excluded in this embodiment. The preset time period can be one month, one quarter, or one year. Since the target user's age varies, the preset time period will also differ; the older the user, the longer the corresponding preset time period. The calculation of the migration consumption activity value can be achieved by combining transaction data, pedestrian flow data, and POI data. For example, a weighted sum of transaction amount and frequency can be performed, while also considering the influence of pedestrian flow and POI density.

[0123] The migration index is calculated using a preset formula that combines regional consumption activity value and migration consumption activity value. Regional consumption activity value refers to the consumption vitality index of the area where the user's original residence is located. Migration consumption activity value refers to the consumption vitality index of the area where the user has moved to, reflecting the consumption environment after migration. The preset formula for the consumption migration index is as follows:

[0124]

[0125] Where λ is the time decay coefficient and Δt is the time elapsed since the migration;

[0126] The migration index of target users is calculated using a pre-defined formula to quantify the time-sensitive impact of changes in the consumption environment after migration. A positive migration index indicates that the user's consumption activity increases after migration, with a higher value indicating a more significant increase; a negative migration index indicates that the user's consumption activity decreases after migration, with a lower value indicating a more significant decrease. By comprehensively considering changes in regional consumption activity and the decay effect over time, the degree of change in user consumption behavior after migration is quantified.

[0127] The target user's current age is input into a preset consumption capacity cycle curve to determine their current consumption capacity value. In this embodiment, the preset consumption capacity cycle curve is a pre-defined mathematical model, drawn by collecting, analyzing, and integrating data on consumption behavior, income, and consumption habits of a large number of people at different age stages. The consumption capacity cycle curve exhibits certain regularities. For example, in the early stages of life, young people have relatively low consumption capacity. As they age, gain work experience, and increase their income, their consumption capacity gradually increases, potentially reaching a peak in middle age. Afterward, consumption capacity may change again due to continued age, health conditions, retirement, and other factors. The target user's current age is used as an input variable and substituted into the preset consumption capacity cycle curve to obtain a consumption capacity value corresponding to that age. In this embodiment, the consumption capacity value represents approximately how much disposable income the target user has at their current age; it may also be categorized by consumption capacity level to reflect the relative level of their consumption.

[0128] The preferred products for the migration address are determined from a pre-built product preference matrix. In this embodiment, the pre-built product preference matrix describes the degree of preference of consumers in different regions for various types of products, such as the purchase frequency, rating, and market share of consumers in different regions for a certain type of product. This data quantifies the preferences of consumers in different regions for different products. After determining the migration address of the target user, this address is mapped to the corresponding region, for example, through Geographic Information System (GIS) technology. Finally, the product preference information corresponding to that region is searched in the pre-built product preference matrix to obtain the product types preferred by consumers at that migration address.

[0129] In a pre-defined region-product weight matrix, a first weight value for the regionally preferred product corresponding to the migration address and a second weight value for the regionally preferred product corresponding to the permanent address are determined. A pre-defined two-dimensional data table is used, where rows represent different regions and columns represent different product categories; cell values ​​are weight values, such as values ​​between 0 and 1, with higher values ​​indicating a stronger preference for the corresponding product in that region. First, the permanent address, migration address, and the regionally preferred products corresponding to each address are identified. Next, weight values ​​are extracted from the pre-defined region-product weight matrix. In this embodiment, the first weight value represents the weight value of the regionally preferred product corresponding to the migration address in the matrix, and the second weight value represents the weight value of the regionally preferred product corresponding to the permanent address in the matrix.

[0130] By combining the user's regional product preferences and consumption behavior corresponding to their migration address, a pre-defined decision tree model is used to obtain the target user's migration product demand value. The regional product preferences corresponding to the migration address are products favored within the region determined by the user's migration address. This can be determined through market research, sales data analysis, etc., representing product categories or types generally preferred by consumers in that region. Product consumption behavior includes the user's past consumption records, such as purchase frequency, amount spent, purchase time, and purchase channel, reflecting the user's consumption habits and tendencies. The pre-defined decision tree model is a common machine learning model used for classification and regression tasks. This model is trained on historical data and is used to predict the user's potential demand for certain products after migration. The user's regional product preferences and consumption behavior data corresponding to their migration address are used as input to the pre-defined decision tree model. Based on the input data, the decision tree model derives the potential demand value (such as purchase probability or demand intensity) for certain products after migration. The migration product demand value refers to the predicted degree or quantity of potential demand for certain products after the user migrates to the new region. It is derived from a comprehensive analysis of factors such as the user's migration background, product preferences in the new region, and the user's consumption behavior.

[0131] The product demand transition value is calculated by subtracting the product of the migration product demand value and the second weight value from the product of the product demand value and the first weight value. The product demand value represents the potential demand or quantity of a product from a user in the current region, typically assessed based on factors such as user consumption behavior and preferences. The migration product demand value refers to the potential demand for a product predicted after a user migrates to a new region, based on product preferences in the new region and the user's consumption behavior. The first weight value refers to the preferred product weight in the region corresponding to the migration address, reflecting the overall preference of consumers in the new region for the product. The second weight value refers to the preferred product weight in the region corresponding to the permanent residence address, reflecting the overall preference of consumers in the original residence area for the product. Calculating the product of the product demand value and the first weight value combines the potential demand for a product in the current region with the preferred weight of that product in the new region, quantifying the intensity of the potential demand in the new region. Calculating the product of the migration product demand value and the second weight value combines the potential demand after a user migrates to a new region with the preferred weight of that product in the original residence area, quantifying the intensity of the user's demand in the original region. Finally, subtracting these two values ​​yields the product demand transition value. In this embodiment, the product demand transition value represents the magnitude of change in user demand for the product before and after migration. Positive values ​​indicate an increase in demand, while negative values ​​indicate a decrease in demand.

[0132] A migration feature vector is constructed based on the migration index, the migration product demand value, and the product demand jump value. The migration index, the migration product demand value, and the product demand jump value are combined into a vector. For example, the migration feature vector = [migration index, migration product demand value, product demand jump value]. The migration feature vector integrates multiple aspects of user migration behavior and changes in product demand, providing a comprehensive data foundation for subsequent analysis of user consumption characteristics.

[0133] Determining the migration consumption characteristic category of target users based on migration feature vectors and updating their user profiles can be achieved using a pre-trained decision tree model. The decision tree model classifies users into corresponding migration consumption characteristic categories based on different index values ​​within the feature vectors. These categories represent different consumption behavior patterns or group types after classifying users using migration feature vectors, reflecting users' consumption preferences, changing needs, and purchasing behavior characteristics in migration scenarios. Labels related to these migration consumption characteristic categories are added or updated in the user profile. For example, if a user is classified as "high-demand growth type," a preference label for high-end products or services might need to be added to the user profile. The demand intensity for various products in the user profile is updated based on the migration product demand value and product demand leap value in the migration feature vectors. For example, if a user has a high migration index and a positive product demand leap value, their label can be updated to "high-potential migration consumer."

[0134] By building user profiles of target users, banks can achieve significant and comprehensive positive impacts in many aspects, such as precision marketing, product and service optimization, and risk management. This will enable them to discover potential user groups and achieve precision marketing, thereby enhancing their competitive advantage and operational efficiency in the market.

[0135] In one embodiment of this example, the CNN-BiLSTM model includes a first convolutional layer, a second convolutional layer, a max pooling layer, a first multilayer perceptron, and a second multilayer perceptron. Analyzing demand information using the preset CNN-BiLSTM model to determine the target user's intent includes the following steps:

[0136] S310. Convert the demand information into word embedding vectors;

[0137] S320. Input the word embedding vector into the first convolutional layer and the second convolutional layer respectively to output the first feature map and the second feature map. Use the max pooling layer to filter the first feature map and the second feature map to determine the first key feature and the second key feature. The convolution kernel of the first convolutional layer is smaller than the convolution kernel of the second convolutional layer.

[0138] S330. The first key feature and the second key feature are generated by nonlinear transformation of the first key feature and the second key feature through the first multilayer perceptron;

[0139] S340. Use the BiLSTM module to capture the temporal relationship of sentences in the demand information;

[0140] S350. Construct multi-layer semantic features by combining sentence temporal correlation, first embedding features, and second embedding features;

[0141] S360. Input the multi-layer semantic features into the second multi-layer perceptron to determine the target user's needs and intentions;

[0142] S370, wherein the first multilayer perceptron is used to process key features, and the second multilayer perceptron is used to process multilayer semantic features.

[0143] Figure 2 This application provides a flowchart illustrating the process of determining a demand intent, as shown in the embodiments of this application. Figure 2 As shown, firstly, a first feature map and a second feature map are determined. Next, a first key feature and a second key feature are determined using the first and second feature maps, and a first embedding feature and a second embedding feature are generated using nonlinear transformation. By determining the temporal relationship of the sentence, these three elements constitute a multi-layered semantic feature set, which is then used to determine the intended demand.

[0144] Convert the requirement information into word embedding vectors. Converting the requirement information into word embedding vectors means transforming the requirement information presented in natural language form into a numerical vector form that can be understood and processed by a computer through a specific algorithm model. First, split the requirement information text into individual independent words or sub-word units according to certain rules. For example, for the Chinese requirement information "I want a mobile phone with clear photos", the word segmentation result may be "I", "want", "a", "take photos", "clear", "of", "mobile phone". Next, remove the stop words. Stop words refer to words that appear frequently in the text but contribute little to semantic understanding, such as "of", "is", etc. Restore the words to their basic forms or stems to unify words in different forms. Subsequently, select a word embedding model. The Word2Vec model is a neural network-based word embedding model. Input the preprocessed words into the selected word embedding model, and the model will generate a vector with a fixed dimension for each word according to its internal parameters and algorithms. For example, using the Word2Vec model may generate a 300-dimensional vector, and the numerical value on each dimension represents the intensity of the word on a certain semantic feature.

[0145] Word embedding vectors are input into the first and second convolutional layers, respectively, to output first and second feature maps. A max-pooling layer is then used to filter the first and second feature maps to determine the first and second key features. In this embodiment, the convolutional kernel of the first convolutional layer is smaller than that of the second convolutional layer. The convolutional layer is a core component of a convolutional neural network, and its main function is to extract features from the input data through convolutional kernels. In natural language processing tasks, the input word embedding vector can be viewed as a one-dimensional data sequence. The convolutional kernel slides across the word embedding vector with a certain stride, performing convolution operations on data within local regions to capture local features. The word embedding vector is input into the first convolutional layer, and a smaller convolutional kernel (e.g., size 2 or 3) is used for convolution. A smaller convolutional kernel can capture more local semantic and syntactic relationships in the word embedding vector. For example, when processing the word embedding vector of the sentence "I like natural language processing," using a convolutional kernel of size 2 might capture local features between adjacent words such as "I-like" and "like-natural." After convolution, the first feature map is output, containing local feature information extracted from the word embedding vectors. The word embedding vectors are then fed into the second convolutional layer, but with a larger kernel (e.g., size 4 or 5). A larger kernel captures broader contextual information, such as sentence or paragraph-level features. Using the sentence "I like natural language processing" as an example, a kernel of size 4 might capture features in a broader range, such as "I-like-natural-language". After convolution, the second feature map is output, containing more global feature information extracted from the word embedding vectors. Max pooling is a commonly used pooling operation, primarily aimed at reducing data dimensionality and computational cost while preserving important features. Max pooling is performed on the first feature map output from the first convolutional layer. Within each local region of the feature map (e.g., a 2x2 or 3x3 region), the largest feature value is selected as the output for that region. This method filters out the most important local features in the first feature map, yielding the first key feature, which represents crucial local semantic and syntactic information in the word embedding vectors. Max pooling is performed on the second feature map output by the second convolutional layer. Similarly, the largest feature value is selected as the output in each local region of the feature map to filter out the most important global features in the second feature map, thus obtaining the second key feature. The second key feature represents the more critical global context information in the word embedding vector.

[0146] The first layer of a Multilayer Perceptron (MLP) performs nonlinear transformations on the first and second key features to generate first and second embedded features. The MLP is a type of feedforward artificial neural network. In deep learning, nonlinear transformations enable models to capture complex patterns and relationships in input data. By introducing nonlinear activation functions, the MLP can learn and represent nonlinear relationships in the data, thereby extracting higher-level abstract features. The first and second key features output from the first and second convolutional layers are used as input to the MLP. These key features may contain both local and global information from the text. The MLP then performs nonlinear transformations on the input key features through its hidden layers. Each neuron in the hidden layer performs a weighted sum of its inputs and then passes them through a nonlinear activation function to produce the output. This process can be repeated multiple times, further extracting and transforming features each time. After multiple layers of nonlinear transformations, the output layer of the MLP produces the first and second embedded features.

[0147] Capturing the temporal relationships between sentences in the context of demand information using a BiLSTM module refers to using a Bi-LSTM (Bidirectional Long Short-Term Memory) model to perform bidirectional temporal modeling on the sentences in the demand information, thereby capturing the temporal dependencies and relationships between sentences. BiLSTM is an abbreviation for Bidirectional Long Short-Term Memory Network, combining forward and backward LSTM networks. LSTM is a special type of recurrent neural network. The forward LSTM network processes sentences in the normal order (from front to back) of the demand information, reading sentences one by one and updating the hidden state of the current sentence based on its content and the hidden states of previous sentences. The forward LSTM can capture the temporal dependencies between sentences from front to back, i.e., the influence of earlier sentences on later sentences. The backward LSTM network processes sentences in reverse order (from back to front) of the demand information, also reading sentences one by one, but in the reverse order of the forward LSTM. The backward LSTM can capture the temporal dependencies between sentences from back to front, i.e., the influence of later sentences on earlier sentences. The hidden states of the forward LSTM and backward LSTM at each sentence position are concatenated to obtain a sentence representation containing bidirectional temporal information. This concatenation operation allows the representation of each sentence to incorporate information from the sentences before and after it, thereby capturing the contextual relationships in the requirement information more comprehensively.

[0148] The sentence temporal correlation, first embedding feature, and second embedding feature constitute a multi-layer semantic feature. The first embedding feature is generated by performing a non-linear transformation on the first key feature using a first-layer multilayer perceptron (MLP). The first key feature is a local feature extracted by performing convolution and max pooling operations on the word embedding vectors. The second embedding feature is generated by performing a non-linear transformation on the second key feature using a first-layer multilayer perceptron (MLP). The second key feature is a global feature extracted by performing convolution and max pooling operations on the word embedding vectors. The sentence temporal correlation, first embedding feature, and second embedding feature are integrated by using a weighted summation method, assigning a weight to each feature, and then performing a weighted summation to construct the multi-layer semantic feature.

[0149] Multi-layer semantic features are input into a second multi-layer perceptron to determine the target user's intent. The second multi-layer perceptron, a deep learning model, performs non-linear transformations and abstractions on the input multi-layer semantic features. The second multi-layer perceptron performs a weighted summation of the input multi-layer semantic features through its hidden layers and generates the output using a non-linear activation function (such as ReLU). Through multi-layer non-linear transformations, the second multi-layer perceptron can capture complex patterns and relationships within the multi-layer semantic features, obtaining higher-level features. These higher-level feature representations better reflect the key information of the user's intent. Finally, the output layer of the second multi-layer perceptron determines the target user's intent based on these high-level feature representations, associating the neurons of the output layer with different intent categories, achieved through mechanisms such as the softmax function.

[0150] In this embodiment, the first multilayer perceptron is used to process key features, and the second multilayer perceptron is used to process multilayer semantic features. The first multilayer perceptron processes the key features extracted through convolutional layers and max pooling layers, namely the first key feature and the second key feature. These key features represent local and global features in the demand information, respectively, and are the model's initial understanding and abstraction of the demand information. The second multilayer perceptron is responsible for processing multilayer semantic features that integrate sentence temporal relationships, the first embedding feature, and the second embedding feature. Multilayer semantic features are a comprehensive feature representation containing multi-level and multi-dimensional information, which can more comprehensively reflect the semantic content of the demand information.

[0151] The model can more comprehensively understand the semantic content of demand information, including the temporal relationship between sentences, local and global semantic information, etc., thereby more accurately determining the demand intent of the target user, which helps to improve the performance of the model in natural language processing tasks and provide users with more accurate and personalized services.

[0152] In one embodiment of this example, verifying the matching degree of the user's intent based on the user profile includes the following steps:

[0153] S410. Select the corresponding matching expression based on the user profile;

[0154] S420. Determine the named entities and entity categories of the user profile based on the matching expression;

[0155] S430. Based on the entity category, the matching expression, and the named entity, traverse each user profile to determine the target attribute of the user profile.

[0156] S440. Use a pre-set language training model to extract the target attribute features and demand intention features of the target attributes and demand intentions, respectively.

[0157] S450. Calculate the feature matching degree between the target attribute features and the demand intent features, and verify the matching degree of the demand intent based on the feature matching degree.

[0158] Based on the user profile, a corresponding matching expression is selected. In this embodiment, the matching expression refers to a logical formula composed of strings or special characters, used to determine whether the input meets specific conditions. These expressions can be based on string matching, numerical comparison, logical operations, etc. First, understand the information contained in the user profile, such as the user's age, gender, interests, preferences, and behavioral patterns. Based on business needs, determine the user characteristics or behavioral patterns that need to be matched. For example, in a personalized recommendation system, it may be necessary to match user interests; in precision marketing, it may be necessary to match user location or consumption behavior. If an existing matching expression can meet the requirements, it can be directly selected and used. For example, regular expressions can be used to match user attributes to determine user preferences.

[0159] Named entities and their categories are determined based on matching expressions in user profiles. Named entities refer to proper nouns or phrases in a user profile that have specific meaning or identifiers, such as person names or place names. Entity categories refer to the category or type to which a named entity belongs, such as "person name," and are labeled with entity categories. Matching expressions are logical expressions used to extract specific patterns or rules from text. In named entity recognition, matching expressions can be implemented using technologies such as regular expressions and rule engines to identify named entities that conform to specific patterns from user profiles. The corresponding matching expressions are defined according to the named entity categories to be identified. For example, to identify "place names" in a user profile, the matching expression can be defined as "[Beijing|Shanghai|Guangzhou|Shenzhen] etc." or a more complex regular expression to match various possible place name formats. The defined matching expressions are applied to the text data or user tags of the user profile for pattern matching and entity extraction, using Python's `re` module for regular expression matching. Based on the matching results of the matching expressions, named entities are extracted from the user profile and labeled with their corresponding entity categories. For example, if the matching expression identifies a place name "Beijing", it is tagged as a "place name" category.

[0160] The target attributes of a user profile are determined by iterating through each user profile using entity categories, matching expressions, and named entities. Target attributes refer to key features or attributes within the user profile and are the ultimate goal of user profile analysis. Based on business needs and analysis objectives, the target attributes to be extracted can be clearly defined, such as "user interests" or "purchase intention." A corresponding entity category is defined for each target attribute; for example, "user interests" might correspond to entity categories like "technology," "fashion," or "sports." Matching expressions, such as regular expressions, are used to iterate through user tags in each user profile and identify named entities. For example, in a user's historical query records, social media activity, or product reviews, matching expressions are used to extract named entities related to the "technology" interest, such as "artificial intelligence" or "machine learning." Subsequently, the identified named entities are categorized under their corresponding entity categories and then associated with the target attributes. For example, the extracted named entities "artificial intelligence" and "machine learning" are categorized under the "technology" interest category, thus determining the user's target attribute as "technology enthusiast." Based on the extracted entity categories and associated target attributes, the target attributes of the user profile are comprehensively determined. For example, if multiple named entities related to "technology" interests are extracted from the user profile, and no other category of entities dominates, then the user's target attribute can be determined to be "technology enthusiast".

[0161] The target attribute features and demand intent features are extracted using a pre-set language training model. The pre-set language training model refers to an NLP model pre-trained on a large amount of text data. Target attribute features are semantic representations extracted from user profiles that reflect core user characteristics (such as interests, preferences, geographic location, and consumption habits). Demand intent features are semantic representations extracted from user demand information that reflect the user's current needs or intentions. User profile data is input into the language training model, and the encoder layer converts the text into a fixed-length vector representation. This vector representation contains the deep semantic information from the user profile, i.e., the target attribute features. Similarly, user demand information data is input into the language training model, and the encoder layer also converts the text into a vector representation. This vector representation contains the deep semantic information from the user demand information, i.e., the demand intent features.

[0162] The feature matching degree between target attribute features and demand intent features is calculated, and the demand intent is verified based on the feature matching degree. Feature matching degree is an indicator that measures the similarity or correlation between two feature vectors. The feature matching degree can be calculated using cosine similarity. By calculating the cosine similarity between the target attribute feature vector and the demand intent feature vector, a reasonable matching degree threshold can be set according to business needs and actual application scenarios. This threshold is used to judge the degree of consistency between the demand intent and the user's target attribute. The calculated feature matching degree is compared with the set threshold. If the feature matching degree is higher than or equal to the threshold, the demand intent is considered consistent with the user's target attribute, and the demand intent is accurate and reasonable. If the feature matching degree is lower than the threshold, the demand intent is considered to deviate from the user's target attribute, and the demand intent may need further adjustment or optimization.

[0163] By combining user profiling, expression matching, named entity recognition, language training models, and feature matching degree calculation, the system achieves accurate matching and verification of user needs and intentions, which not only improves the accuracy of needs matching and the intelligence level of the system, but also optimizes the user experience.

[0164] In one embodiment of this example, selecting the corresponding matching expression based on the user profile includes the following steps:

[0165] S510. Select target information of the target user in the user profile, and select the first matching expression from the candidate matching expressions in the preset database based on the target information;

[0166] S520. Use the first matching expression to traverse each user tag in the user profile to determine the number of matching user tags;

[0167] S530. Determine the number of rule tags corresponding to the first matching expression;

[0168] S540. Calculate the ratio between the number of matches and the number of rule tags to determine the first match similarity.

[0169] S550. Traverse each candidate matching expression in the preset database and calculate the matching similarity of each candidate matching expression;

[0170] S560. Select the candidate matching expression with the highest similarity as the matching expression.

[0171] The process involves selecting target information from the user profile and then choosing a first matching expression from a pool of candidate matching expressions in a pre-defined database. Target information refers to key features highly relevant to the current business scenario or task objective, selected from the user profile. Candidate matching expressions are predefined sets of logical rules used to match user features with business rules. These rules can be implemented using regular expressions, conditional statements, and other techniques to filter user tags that meet specific conditions from the user profile. The first matching expression is the first matching rule selected from the set of candidate matching expressions. Selecting target information from the user profile can be determined based on relevant features within the user profile, extracting feature fields relevant to the business scenario requirements. For example, extracting fields such as "average monthly spending," "membership level," and "last purchase time." Next, a predefined list of candidate matching expressions is obtained, and a matching expression is selected from this list.

[0172] The first matching expression is used to iterate through each user tag in the user profile to determine the number of matches for each user tag. This first matching expression is then used as a filtering condition to check the user tags in the user profile. For each user in the user profile, each tag is checked to see if it matches the first matching expression. For example, for each user, check if their "average monthly spending" tag is greater than 2000 yuan, or if their "number of logins in the last 30 days" tag is greater than 5 times. The number of user tags that match the first matching expression is counted, thus determining the number of matches for each user tag.

[0173] Determine the number of rule tags corresponding to the first matching expression. The number of rule tags refers to the number of user tags corresponding to the first matching expression. Here, user tags are the basic elements that constitute the expression conditions, such as "age", "city", "spending amount", etc.

[0174] Subsequently, the ratio between the number of matches and the number of rule tags is calculated to determine the first match similarity. This ratio quantifies the similarity between the first matching expression and the target user or data record. The number of matches refers to the number of features or elements in the object to be analyzed that match the preset rule tags; the number of rule tags refers to the total number of tags used for matching in the preset rule system. The quantitative indicator calculated using this ratio measures the degree of matching between the object to be analyzed and the rule tags.

[0175] The process iterates through each candidate matching expression in the preset database and calculates the matching similarity for each expression. Iterating through the database involves accessing each candidate matching expression sequentially. Then, using each candidate matching expression in the preset database, the process iterates through each user tag in the user profile and determines the number of rule tags and user matches for each expression. Finally, the matching similarity for each candidate matching expression is calculated using the ratio between the number of matches and the number of rule tags.

[0176] The candidate matching expression with the highest similarity is selected as the final matching expression. In other words, the best matching expression is selected from multiple candidate rules.

[0177] By selecting corresponding matching expressions based on user profiles, we can not only improve the accuracy of rule matching and ensure the efficient use of business resources, but also optimize business operation efficiency, respond quickly to market changes, enhance user experience, provide personalized services, and reduce interference.

[0178] In one embodiment of this example, determining the evaluation index of all departments based on departmental resource information and marketing task indicators, combined with game theory weighting and multi-criteria compromise solution ranking, includes the following steps:

[0179] S601. Use principal component analysis to select departmental resource indicators based on departmental resource information;

[0180] S602. Construct a decision matrix by combining departmental resource indicators and marketing task indicators. The decision matrix is ​​the object of behavioral evaluation, and the columns of the decision matrix are the evaluation indicators.

[0181] S603. Standardize each element in the decision matrix to obtain standardized indicators, and determine positive and negative indicators based on preset business rules.

[0182] S604. Calculate the index variability and index conflict based on the preset variability formula and the preset conflict formula, respectively.

[0183] S605. The amount of information about an indicator is calculated by combining the variability and conflict of indicators.

[0184] S606. Calculate the first weight vector based on the amount of indicator information using a preset weight formula.

[0185] S607. Calculate the proportion of each evaluation indicator in the evaluation object in the decision matrix, and calculate the information entropy of the indicator based on the indicator proportion using the preset information entropy formula.

[0186] S608. Calculate the difference coefficient of each evaluation indicator based on the indicator information entropy of each evaluation indicator, and calculate the second weight vector of each evaluation indicator based on the difference coefficient.

[0187] S609. Combine the first and second weight vectors and use game theory to determine the optimal weights;

[0188] S610. Based on the optimal weights, use the multi-criteria compromise solution to rank and determine the evaluation index of all departments.

[0189] Principal component analysis (PCA) is used to screen departmental resource indicators based on departmental resource information. Departmental resource information refers to various data about departmental resources, such as human and material resources in administrative departments. Departmental resource indicators are key dimensions of data extracted from resource information that reflect the characteristics and value of resources. Principal components with a cumulative variance contribution rate reaching a threshold are selected based on their variance contribution rate. These principal components are then used to deduce the original indicators that have the greatest impact on departmental resources. The departmental resource indicators are standardized, and then a correlation coefficient matrix is ​​calculated to analyze the correlation between the resource indicators. Next, the eigenvalues ​​and eigenvectors of the correlation coefficient matrix are calculated. The eigenvalues ​​represent the magnitude of the variance of the principal components, and the eigenvectors define the direction of the principal components. The principal components are sorted from largest to smallest by eigenvalue, and those with a cumulative variance contribution rate reaching the threshold are selected. The contribution of each original indicator to the principal components is analyzed using the principal component loading matrix to identify the original indicators that have the greatest impact on the principal components, thus determining the departmental resource indicators.

[0190] A decision matrix is ​​constructed by combining departmental resource indicators and marketing task indicators. The decision matrix consists of evaluation objects (resources and tasks) and evaluation indicators (columns). In this embodiment, the evaluation objects are resource evaluation objects and task evaluation objects, and the evaluation indicators are resource evaluation indicators and task evaluation indicators. Resource evaluation objects are specific resource carriers within the department, i.e., resource entities that the department can utilize. Marketing task evaluation objects refer to specific marketing tasks, i.e., task subjects that require resource support. Resource evaluation indicators are key dimensions that characterize resource attributes, such as human resource efficiency and equipment value. Marketing task indicators are dimensions that measure task requirements and objectives, such as conversion efficiency. Using evaluation objects (resources / tasks) as rows in the matrix means that each row corresponds to a specific resource or task; using evaluation indicators (resource / task indicators) as columns means that each column corresponds to a measurement dimension (resource characteristics or task requirements).

[0191] Standardized indicators are obtained by standardizing each element in the decision matrix. Positive and negative indicators are then determined based on preset business rules. Z-Score standardization can be used to process each matrix element to obtain standardized indicators. In this embodiment, the preset business rules can be determined according to actual requirements. Positive indicators (benefit-oriented indicators) are those with higher values, such as "resource output efficiency," "task conversion rate," and "audience coverage." Negative indicators (cost-oriented indicators) are those with lower values, such as "resource consumption cost," "task execution deviation rate," and "budget overrun ratio." Positive and negative indicators are determined according to preset business rules. For example, if the business rule requires higher equipment utilization, then equipment utilization (%) is a positive indicator; if the business rule requires lower resource idle rate, then idle rate (%) is a negative indicator.

[0192] The variability and conflict of indicators are calculated based on preset variability and conflict formulas, respectively. Indicator variability reflects the degree of numerical difference of the same indicator across different evaluation objects. The higher the variability, the better the indicator can distinguish the merits of different objects, and the more valuable it is in decision-making. The preset variability formula in this embodiment is as follows:

[0193]

[0194] in, Indicates the variability of the index, x i It is each data point, is the mean, and n is the number of data points.

[0195] The variability of an indicator is determined by calculating the standard deviation of each matrix element, which measures the absolute degree to which the data deviates from the mean, and thus determines the variability of the indicator.

[0196] Subsequently, the conflict factor of the indicators is calculated using a pre-defined conflict factor formula. This formula reflects the consistency of different indicators in evaluating the same object. If two indicators show the same evaluation trend for the same object (e.g., the higher the value of indicator A, the higher the value of indicator B), the conflict factor is low; if the trends are opposite (e.g., the higher the value of indicator A, the lower the value of indicator B), the conflict factor is high, indicating a contradiction between the indicators. First, the Pearson correlation coefficient (r) between the indicators is calculated. ij The correlation coefficient (i) represents the correlation between indicator i and indicator j. The correlation coefficient ranges from -1 to 1; the closer the absolute value is to 1, the stronger the correlation. The preset conflict formula is shown below:

[0197]

[0198] Among them, R j Indicates the conflict of indicators, r jk is the correlation coefficient between indicator j and indicator k, and m is the total number of indicators. Conflict analysis quantifies the independence between indicators by summing the differences between each correlation coefficient and the sum of 1.

[0199] Next, the information content of an indicator is calculated by combining indicator variability and indicator conflict. Indicator information content is a comprehensive indicator that measures the contribution of a single indicator to decision-making. By integrating variability (discriminating ability) and conflict (consistency among indicators), the "effective information value" of an indicator in decision-making is quantified. The information content of an indicator can be calculated using the information content calculation formula, as shown below:

[0200]

[0201] Where C represents the amount of information in the indicator. R represents the variability of the index. j This indicates conflicting indicators.

[0202] By calculating the information content of indicators, it can be determined that indicators with a large amount of information should simultaneously possess high data volatility (variability) and low correlation with other indicators (conflict), thereby providing unique and rich information.

[0203] Subsequently, the first weight vector is calculated using a preset weighting formula based on the information content of the indicators. This quantifies the information value of the indicators and determines their relative importance in decision-making. Indicators with greater information content should be assigned higher weights to reflect their greater contribution to the decision outcome. The preset weighting formula is shown below:

[0204]

[0205] Among them, W i I represents the first weight vector. i Let represent the information content of the i-th indicator, and n represent the total number of indicators.

[0206] The calculation involves determining the proportion of each evaluation indicator in the decision matrix within the evaluated object. Based on this proportion, the information entropy of the indicator is calculated using a pre-defined formula. This process quantifies the uncertainty of the indicator to assess its information value; the lower the information entropy, the lower the uncertainty, the greater the information content, and the more significant the impact on the decision. The indicator proportion represents the relative importance of indicator j within evaluation object i, and its calculation formula is as follows:

[0207]

[0208] in, Indicates the percentage of the indicator, x ij is the original value of indicator j in evaluation object i, and n is the total number of evaluation objects.

[0209] Calculating the indicator proportion involves transforming the original indicator values ​​into relative weights, eliminating dimensional differences, and providing a standardized basis for subsequent calculations. Next, based on the indicator proportions, the information entropy of the indicator is calculated using a pre-defined information entropy formula. Information entropy measures the degree of data dispersion of the indicator. A smaller entropy value indicates more dispersed indicator data (e.g., a department's human resource efficiency is much higher than other departments), indicating that the indicator can effectively distinguish the evaluation objects; the greater the information content, the higher the weight should be. Conversely, a larger entropy value indicates more uniform indicator data (e.g., human resource efficiency is similar across departments), indicating low distinguishability; the smaller the information content, the lower the weight should be. The formula for calculating the indicator information entropy is:

[0210]

[0211] Where k is the normalization constant, This indicates the percentage of the indicator.

[0212] The difference coefficient for each evaluation indicator is calculated based on its information entropy. A second weight vector for each indicator is then calculated based on this difference coefficient. Information entropy measures the dispersion of the indicator data. A larger entropy value indicates a more dispersed distribution of the indicator data, while a smaller entropy value indicates a more concentrated data set. The difference coefficient is calculated using the formula gj = 1 - ej, where ej is the information entropy of indicator j. A larger difference coefficient indicates a more important indicator. For example, if the entropy value of an indicator is ej = 0.7, then the difference coefficient gj = 0.3, indicating that the indicator's weight may be relatively low. The weight vector is obtained after normalizing the difference coefficient, and the weights are positively correlated with the difference coefficient. For example, if the difference coefficient of indicator A is 0.4 and that of indicator B is 0.6, then indicator B has a higher weight (0.6 / 1.0 = 60%). Calculating the second weight vector allows for a more accurate determination of the data's dispersion or information content, and a more precise allocation of weight values.

[0213] By combining the first and second weight vectors and employing game theory to determine the optimal weights, the first and second weight vectors are considered as two "decision-making entities" that reach an equilibrium state through game theory, thus determining the optimal weights. The optimal weights are a "compromise solution" between the first and second weight vectors. Treating the first and second weight vectors as two participants in the game, the optimal combination of them, i.e., the optimal weight vector, is found using game theory methods. Game theory-based weighting methods can comprehensively consider the advantages of different weight allocation methods, avoiding the limitations of a single method, and making weight allocation more scientific and reasonable.

[0214] Based on the optimal weight vector, the Multi-Criterion Compromise Ranking (VIKOR) method is used to rank and evaluate each department, resulting in an evaluation index for each department. The VIKOR method is a multi-criteria decision-making method that considers both maximizing group utility and minimizing individual regret. It can find compromise solutions among multiple evaluation criteria, i.e., the solutions closest to the ideal solution. Through the VIKOR method, a departmental evaluation index that comprehensively considers multiple evaluation criteria can be obtained, which can fully reflect the overall performance of the departments.

[0215] By combining various data analysis methods and decision theories, evaluation indices for all departments are determined. This not only ensures the scientific rigor and objectivity of the evaluation but also enhances its comprehensiveness and integration, providing strong support for departmental performance evaluation and decision-making. Furthermore, this method is flexible and adaptable, capable of being adjusted and optimized according to different business scenarios and needs.

[0216] In one embodiment of this example, determining the evaluation index of all departments by ranking them using a multi-criteria compromise solution based on optimal weights includes the following steps:

[0217] S710. Traverse each positive and negative indicator to determine the maximum positive and minimum negative indicators;

[0218] S720. Based on the preset index gap formula, the comprehensive index gap value of each matrix element is calculated by combining the maximum positive index, the minimum negative index, each standardized index and the optimal weight.

[0219] S730. Iterate through the comprehensive gap values ​​of each indicator to determine the maximum and minimum indicator gap values;

[0220] S740. Calculate the distance to the best solution and the distance to the worst solution based on each standardized index and the maximum positive index value and the minimum negative index respectively;

[0221] S750. Calculate the proximity by using the distance between the best solution and the distance between the worst solution, and calculate the evaluation index based on the proximity, the maximum index gap, the minimum index gap, and the comprehensive index gap.

[0222] S760. Based on the evaluation index, all departments are ranked to determine the departments to be assigned marketing tasks.

[0223] The process iterates through each positive and negative indicator to determine the maximum positive and minimum negative indicators. Positive indicators are those with higher values, while negative indicators are those with lower values. In this embodiment, the determination of positive and negative indicators can be based on preset business rules. Determining these indicators improves resource utilization efficiency, avoids resource redundancy, and enhances the accuracy of resource assessment. Subsequently, all indicators in the decision matrix are examined one by one, distinguishing their positive and negative attributes and extracting key values. The maximum positive indicator is the indicator value with the largest value among all positive indicators, representing the "optimal state" of that indicator. The minimum negative indicator is the indicator value with the largest value among the negative indicators, representing the "worst state" of that indicator.

[0224] This model calculates the comprehensive index gap value for each matrix element using a pre-defined index gap formula, combining the maximum positive index, minimum negative index, each standardized index, and optimal weights. The comprehensive gap value is typically calculated based on weighted distance and is used to quantify the difference between actual and ideal values. It combines the maximum value of the positive index, the minimum value of the negative index, and the maximum value of the negative index. The pre-defined index gap formula is as follows:

[0225]

[0226] in, This represents the overall difference in indicators, where wj represents the optimal weight. It is the ideal value of the j-th indicator (positive indicators take the maximum value, and negative indicators take the minimum value);

[0227] The ideal value of a positive indicator is the maximum value of that indicator (i.e., the "maximum positive indicator"); the ideal value of a negative indicator is the minimum value of that indicator, representing the negative ideal solution (i.e., the worst value).

[0228] The maximum and minimum index gap values ​​are determined by iterating through the comprehensive gap values ​​of each indicator. These are the comprehensive differences between each evaluation object and the ideal state calculated in the previous steps. The comprehensive gap value of each evaluation object is checked one by one. The maximum index gap value is the maximum value among all the comprehensive gap values, representing the object with the largest gap from the ideal state (i.e., the worst solution). The minimum index gap value is the minimum value among all the comprehensive gap values, representing the object with the smallest gap from the ideal state (i.e., the optimal solution).

[0229] The optimal and worst-case distances are calculated for each standardized index against the maximum positive index value and the minimum negative index value, respectively. The maximum positive index value is the highest positive index value among all evaluated objects; the minimum negative index value is the lowest negative index value among all evaluated objects. Euclidean distance can be used to calculate these distances. The optimal distance represents the distance from solution i to the ideal solution; the worst-case distance represents the distance from solution i to the negative ideal solution. The smaller the optimal distance, the closer the solution is to the ideal solution (the better); the larger the worst-case distance, the further the solution is from the negative ideal solution (the better).

[0230] Subsequently, the proximity score is calculated using the distances to the optimal and worst solutions. An evaluation index is then calculated based on the proximity score, the maximum index difference, the minimum index difference, and the overall index difference, transforming the multi-index decision-making problem into a single evaluation index to quantify the overall merits of the evaluated object. The proximity score is calculated by dividing the distance to the worst solution by the sum of the distances to the optimal and worst solutions; a closer proximity score to 1 indicates a better solution. Next, the evaluation index is calculated based on the proximity score, the maximum index difference, the minimum index difference, and the overall index difference. The proximity score reflects the degree of closeness between a solution and the optimal solution, and can be calculated based on the distances to the optimal and worst solutions. Its value range is generally [0,1], with larger values ​​indicating closer proximity to the optimal solution. The maximum / minimum index difference is the maximum and minimum value obtained after iterating through all the overall index difference values, representing the range of differences in the entire dataset, used for subsequent standardization or weight adjustment. The overall index difference is a comprehensive measure of difference combining index weights, positive and negative attributes, and standardized values, reflecting the degree of difference between a single matrix element and the ideal state. For example, if the evaluation index is Ei, the calculation method for the i-th scheme is as follows:

[0231]

[0232] Where α and β are weighting coefficients (α+β=1), used to balance the influence of proximity and difference, E i It is an evaluation index.

[0233] The marketing task allocation is determined by ranking all departments based on an evaluation index, which is a quantitative score of a department's overall capabilities. All departments are ranked in descending order of their evaluation indices, with higher indices indicating stronger overall capabilities. This quantitative differentiation of departmental strengths and weaknesses avoids subjective biases arising from experience-based allocation. Next, tasks are directly assigned to the department with the highest evaluation index (e.g., high-difficulty brand promotion tasks are assigned to the marketing department with the highest index), thus determining the marketing task allocation department.

[0234] By determining the evaluation index for all departments, not only is the scientific and objective nature of the evaluation ensured, but the comprehensiveness and integration of the evaluation are also improved. This allows for adjustments and optimization based on different business scenarios and needs, which helps promote continuous improvement and enhancement of departments.

[0235] In one embodiment of this example, determining the marketing task allocation department using a preset screening model includes:

[0236] S810. Using a preset screening model, the departmental resource information and marketing task indicators are matched to obtain basic matching results, quantity matching results and energy efficiency matching results.

[0237] S820. Based on the basic matching results, quantity matching results, and energy efficiency matching results, determine the departments responsible for allocating marketing tasks.

[0238] Using a pre-defined screening model, departmental resource information and marketing task indicators are matched to obtain basic matching results, quantity matching results, and energy efficiency matching results. In this embodiment, the pre-defined screening model can be a neural network model. The basic matching result assesses the basic compatibility between departmental resources and marketing tasks, i.e., whether the resources meet the most basic requirements for task execution. In this embodiment, the evaluation content of the basic matching result is resource type matching, that is, checking whether the types of resources possessed by the department match the types of resources required by the marketing task. For example, if the marketing task is a lending business, it determines whether the department's resource types can handle the lending business. The quantity matching result assesses whether the quantity of departmental resources is sufficient to meet the needs of the marketing task in terms of scale, scope, or intensity. In this embodiment, the evaluation content of the quantity matching result can compare the quantity of resources available to the department with the quantity of resources required by the task to assess whether there is a quantity gap or redundancy. The energy efficiency matching result assesses the efficiency of departmental resource utilization, i.e., the energy efficiency performance of resources in completing the task, including cost-effectiveness and time efficiency. In this embodiment, the evaluation content of the energy efficiency matching result is to calculate the ratio between resource input and task output to assess the economy of resource use.

[0239] Based on the basic matching results, quantity matching results, and energy efficiency matching results, the departments responsible for marketing task allocation are determined. These results are then incorporated into a comprehensive evaluation framework to assess the suitability of departmental resources for marketing tasks from multiple dimensions. First, it is ensured that the basic conditions for task execution are met, meaning that departmental resources and marketing tasks meet the acceptable standards in terms of basic matching results. If basic matching is insufficient, task allocation should be carefully considered even if other matching results are good. With basic matching satisfied, quantity matching results are considered to ensure that departmental resources are sufficient to support the smooth execution of tasks. If resources are insufficient, it is necessary to assess whether they can be supplemented through internal allocation or external procurement. Finally, based on energy efficiency matching results, the task allocation plan is optimized, selecting the department with the highest resource utilization efficiency to execute tasks, thereby reducing costs and improving execution efficiency.

[0240] By designating departments to assign marketing tasks, we can ensure the smooth execution of marketing activities, optimize resource allocation, and improve overall operational efficiency.

[0241] This application also provides a machine-readable storage medium storing instructions that cause a machine to execute the novel digital marketing collaboration method applicable to small and medium-sized banks described above.

[0242] This application also provides a novel digital marketing collaboration system suitable for small and medium-sized banks, comprising:

[0243] The memory is configured to store instructions; and

[0244] The processor is configured to retrieve instructions from memory and, when executing those instructions, to implement the aforementioned novel digital marketing collaboration method applicable to small and medium-sized banks.

[0245] 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.

[0246] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as 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. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0247] 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.

[0248] 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.

[0249] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0250] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0251] 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 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.

[0252] 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 process, method, article, or apparatus. Unless otherwise specified, 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 that element.

[0253] The above are merely embodiments of this application and are 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 novel digital marketing collaboration method suitable for small and medium-sized banks, characterized in that, The method includes the following steps: In response to receiving the target user's demand information, the system obtains the target user's user attributes, including permanent address, user age, and product consumption behavior. Use GIS technology to obtain the regional consumption activity value corresponding to the permanent address; Input the target user's age into the preset spending power prediction model to determine the target user's spending power assessment value; Determine the regional preferred products corresponding to the permanent address from the preset product preference matrix; By combining regional product preferences and product consumption behavior, a pre-set decision tree model is used to obtain the product demand value of the target users; A feature vector is constructed from regional consumption activity value, consumption capacity assessment value, and product demand value; Density clustering algorithm is used to perform cluster analysis on feature vectors to determine the consumption characteristic categories of target users; Build user profiles for target users based on consumption characteristic categories; When a target user changes their permanent address, GIS technology is used to obtain the migration consumption activity value corresponding to the target user's new address. The migration index is calculated using a preset consumption migration index formula by combining regional consumption activity value and migrating consumption activity value; Input the target user's current age into the preset spending power cycle curve to determine the target user's current spending power value; Determine the region-preferred products corresponding to the migration address from the preset product preference matrix; In the preset region-product weight matrix, determine the first weight value of the regional preference product corresponding to the migration address and the second weight value of the regional preference product corresponding to the permanent residence address; By combining the regional product preferences and product consumption behavior corresponding to the migration address, the migration product demand value of the target user is obtained using a pre-set decision tree model; The product demand transition value is calculated by subtracting the product of the migrated product demand value and the second weight value from the product of the product demand value and the first weight value. A migration feature vector is constructed based on the migration index, migration product demand value, and product demand transition value. Based on the migration feature vector, determine the migration consumption feature category of the target user and update the user profile of the target user; The target user's intent is determined by analyzing the demand information using a pre-set CNN-BiLSTM model. Verify the matching degree of user intent based on user profile; If the demand intent passes the matching verification, the demand intent will be converted into a marketing task indicator; Obtain resource information for all departments; The evaluation index for all departments is determined by combining departmental resource information and marketing task indicators with game theory weighting and multi-criteria compromise solutions. The evaluation index is compared with the preset evaluation threshold. If the evaluation index is greater than the preset evaluation threshold, the department responsible for assigning marketing tasks is determined. If the evaluation index is less than the preset evaluation threshold, the marketing task allocation department is determined using the preset screening model.

2. The method according to claim 1, characterized in that, The CNN-BiLSTM model includes a first convolutional layer, a second convolutional layer, a max pooling layer, a first multilayer perceptron, and a second multilayer perceptron. The step of analyzing demand information using the pre-defined CNN-BiLSTM model to determine the target user's intent includes the following steps: Transform demand information into word embedding vectors; The word embedding vectors are input into the first convolutional layer and the second convolutional layer respectively to output the first feature map and the second feature map. The first feature map and the second feature map are then filtered by the max pooling layer to determine the first key feature and the second key feature. The convolution kernel of the first convolutional layer is smaller than the convolution kernel of the second convolutional layer. The first embedded feature and the second embedded feature are generated by nonlinear transformation of the first key feature and the second key feature using a first multilayer perceptron. The BiLSTM module is used to capture the temporal relationships between sentences in demand information. The sentence temporal correlation, the first embedding feature, and the second embedding feature are used to construct a multi-layer semantic feature; The multi-layer semantic features are input into the second multi-layer perceptron to determine the target user's needs and intentions. The first multilayer perceptron is used to process key features, and the second multilayer perceptron is used to process multilayer semantic features.

3. The method according to claim 1, characterized in that, The process of verifying the matching degree of user intent based on user profile includes the following steps: Select the corresponding matching expression based on the user profile; The named entities of the user profile are determined based on the matching expression, and the entity category of the named entities is determined. Based on entity categories, matching expressions, and named entities, each user profile is traversed to determine the target attributes of the user profile; The target attribute features and demand intention features are extracted using a pre-defined language training model. Calculate the feature matching degree between target attribute features and demand intent features, and verify the matching degree of demand intent based on the feature matching degree.

4. The method according to claim 3, characterized in that, The step of selecting the corresponding matching expression based on the user profile includes the following steps: Select target information of the target user from the user profile, and select the first matching expression from the candidate matching expressions in the preset database based on the target information; Use the first matching expression to iterate through each user tag in the user profile to determine the number of matches for each user tag; Determine the number of rule tags corresponding to the first matching expression; The first match similarity is calculated by comparing the number of matches with the number of rule tags. Iterate through each candidate matching expression in the preset database and calculate the matching similarity of each candidate matching expression; The candidate matching expression with the highest similarity is used as the matching expression.

5. The method according to claim 1, characterized in that, The process of determining the evaluation index for all departments based on departmental resource information and marketing task indicators, combined with game theory weighting and multi-criteria compromise solutions, includes the following steps: Principal component analysis was used to select departmental resource indicators based on departmental resource information. A decision matrix is ​​constructed by combining departmental resource indicators and marketing task indicators, where the rows of the decision matrix are evaluation objects and the columns are evaluation indicators. Standardized indicators are obtained by standardizing each element in the decision matrix, and positive and negative indicators are determined based on preset business rules. The variability and conflict of indicators are calculated based on the preset variability formula and the preset conflict formula, respectively. The information content of the indicators is calculated by combining the variability and conflict of the indicators. The first weight vector is calculated based on the amount of indicator information using a preset weight formula. Calculate the proportion of each evaluation indicator in the evaluation object in the decision matrix, and calculate the information entropy of the indicator based on the indicator proportion using a preset information entropy formula. The difference coefficient of each evaluation indicator is calculated based on the indicator information entropy of each evaluation indicator, and the second weight vector of each evaluation indicator is calculated based on the difference coefficient. The optimal weights are determined by combining the first and second weight vectors using game theory. The evaluation index for all departments is determined by ranking them using a multi-criteria compromise solution based on the optimal weights.

6. The method according to claim 5, characterized in that, The process of determining the evaluation index of all departments by using a multi-criteria compromise solution based on optimal weights includes the following steps: Iterate through each positive and negative indicator to determine the maximum positive and minimum negative indicators; The comprehensive index gap value of each matrix element is calculated by using a preset index gap formula, combined with the maximum positive index, the minimum negative index, each standardized index, and the optimal weight. Iterate through the comprehensive gap values ​​of each indicator to determine the maximum and minimum indicator gap values; The distances to the best and worst solutions are calculated based on each standardized index, with respect to the maximum positive index value and the minimum negative index value, respectively. The proximity is calculated by the distance between the best and worst solutions, and the evaluation index is calculated based on the proximity, the maximum index gap, the minimum index gap, and the comprehensive index gap. All departments are ranked based on evaluation indices to determine which departments are responsible for marketing task allocation.

7. The method according to claim 1, characterized in that, The process of determining the marketing task allocation department using a preset screening model includes: Using a pre-set screening model, departmental resource information and marketing task indicators are matched to obtain basic matching results, quantity matching results, and energy efficiency matching results; The marketing task allocation department is determined based on the basic matching results, quantity matching results, and energy efficiency matching results.

8. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to execute a novel digital marketing collaboration method applicable to small and medium-sized banks according to any one of claims 1 to 7.

9. A novel digital marketing collaboration system suitable for small and medium-sized banks, characterized in that: include: The memory is configured to store instructions; as well as A processor is configured to retrieve instructions from memory and, when executing the instructions, to implement a novel digital marketing collaboration method applicable to small and medium-sized banks according to any one of claims 1 to 7.

Images