A customer target product conversion strategy generation method, device and equipment based on multi-source heterogeneous data and a medium
By processing multi-source heterogeneous data and updating the model feedback, the problems of data dispersion and disconnection of scoring results in the marketing of financial products and e-commerce products have been solved. This has enabled dynamic reflection of changes in customer interests and scenario adaptation, thereby improving the applicability and responsiveness of marketing strategies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI SHUHUA INFORMATION TECHNOLOGY CO LTD
- Filing Date
- 2026-04-14
- Publication Date
- 2026-07-14
Smart Images
Figure CN122390777A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, specifically to a method, apparatus, equipment, and medium for generating customer target product conversion strategies based on multi-source heterogeneous data. Background Technology
[0002] In financial product marketing and online merchandise operations, businesses typically need to identify target customers with a high probability of purchasing a particular product from a large customer base, and then arrange recommendation channels, timing of outreach, and marketing frequency accordingly. Existing solutions mostly revolve around scoring based on basic customer information, historical transaction records, or a single recommendation model, and then sort, filter, or recommend products based on the scoring results.
[0003] While existing technologies can identify potential customers to some extent, they still face the following challenges in multi-scenario business environments. First, customer data sources are fragmented, including basic information, asset information, product holding information, as well as behavioral information such as browsing, clicking, inquiries, adding to cart, subscribing, and payment. Furthermore, data may be affected by external environmental factors such as interest rate fluctuations, promotional activities, holidays, and market conditions. The time granularity and field structure of data from different sources vary significantly, leading to feature distortion when used directly. Second, many solutions rely on static profiles or long-term historical behavior for prediction, failing to reflect recent changes in customer interests and conversion intentions. Third, many solutions, after obtaining predicted scores, only use them for ranking and display, lacking linkage with recommendation actions, marketing resources, and feedback updates, resulting in a disconnect between scoring results and subsequent execution. Fourth, in scenarios where financial products and e-commerce goods coexist, product risk levels, price ranges, inventory status, activity tags, and channel constraints all affect conversion results. Existing single scoring methods struggle to express the scenario compatibility between customers and target products. In response, we propose a method, apparatus, equipment, and medium for generating customer target product conversion strategies based on multi-source heterogeneous data. Summary of the Invention
[0004] To address the aforementioned technical issues, this paper provides a method, apparatus, equipment, and medium for generating customer target product conversion strategies based on multi-source heterogeneous data. This technical solution solves the problems of scattered customer data sources, static scoring failing to reflect short-term intention changes, scoring results being disconnected from the marketing execution chain, and insufficient adaptability to different product scenarios.
[0005] To achieve the above objectives, the technical solution adopted by this invention is: a method for generating customer target product conversion strategies based on multi-source heterogeneous data, comprising: S1. Obtain multi-source basic data, target product data, real-time behavioral data, and external environment data of the customer to be evaluated; S2. Clean, standardize, repair missing data, align time, and associate entities with the multi-source basic data, target product data, real-time behavior data, and external environment data to obtain a unified sample dataset. S3. Construct customer static profile feature vector, customer dynamic intention feature vector, and customer-product scenario adaptation feature vector based on the unified sample dataset; S4. Input the customer static profile feature vector into the offline static scoring model to obtain the basic conversion score; S5. Input the customer dynamic intention feature vector into the online dynamic correction model to obtain the intention correction score; S6. Determine the scenario adaptation coefficient based on the customer-product scenario adaptation feature vector; S7. The basic conversion score, intention correction score and scenario adaptation coefficient are combined to obtain the comprehensive conversion probability of the customer to be evaluated for the target product. S8. Generate marketing strategy output results based on comprehensive conversion probability and preset marketing resource constraints; S9. Update sample labels, sample weights, model parameters, and marketing strategy thresholds based on subsequent customer feedback.
[0006] Preferably, the customer static profile feature vector in step S3 is constructed based on at least two or more of the following fields, including: customer basic information, account level, asset size, historical holding structure, credit information, historical transaction frequency, historical purchase amount, historical product category distribution, and historical marketing response rate.
[0007] Preferably, the customer dynamic intention feature vector in step S3 is constructed based on browsing, clicking, searching, consulting, adding to favorites, adding to cart, claiming coupons, subscribing, placing orders, paying, and activating behaviors within a preset time window, and aggregated according to behavior type weights and time decay rules.
[0008] Preferably, the time decay rule satisfies:
[0009] in, Indicates the first The contribution value of each behavioral event to dynamic intention characteristics. Indicates the weight of behavior type. Indicates the intensity of behavior. Indicates the time when the action occurred. Indicates the current scoring time. This represents the time decay coefficient.
[0010] Preferably, the customer-product scenario adaptation feature vector in step S3 includes one or more of the following: risk level matching degree, price range matching degree, term matching degree, channel accessibility, activity participation degree, and inventory or quota availability.
[0011] Preferably, the specific formula for fusing the basic conversion score, intention correction score, and scene adaptation coefficient in step S7 is as follows:
[0012]
[0013] in, Indicates the basic conversion score. Indicates intention to adjust score, Indicates the scene adaptation coefficient. Indicates the cross-action term. This represents the scene constraint gating term, where α, β, γ, and μ are the fusion weights. For the overall score; for function; For the final predicted probability, This is the Sigmoid function.
[0014] Preferably, the marketing strategy output in step S8 includes one or more of the following: target customer set, target product identifier, reach channel, reach time period, reach frequency, and priority.
[0015] Preferably, the step S9, which involves updating the sample labels, sample weights, model parameters, and marketing strategy thresholds based on subsequent customer feedback behavior, includes: Receive subsequent customer feedback within a preset observation window; Based on whether the customer has achieved the target conversion, the samples are labeled as positive samples, weakly positive samples, or negative samples; Incremental training or periodic retraining is performed based on the updated sample labels and sample weights; The marketing strategy thresholds are re-estimated based on the updated forecasts.
[0016] Preferably, the sample weights satisfy:
[0017] in, This represents the overall weighted preference value. , and These are three learnable weight coefficients; For time decay term, Here is the current timestamp, and t is the current rating time. The time-related decay coefficient of the sample. Prioritize the target product. The importance of the customer's channel.
[0018] Preferably, a customer target product conversion strategy generation device based on multi-source heterogeneous data includes: The data acquisition module is used to acquire multi-source basic customer data, target product data, real-time customer behavior data, and external environment data. The data preprocessing module is used to perform cleaning, standardization, missing data repair, time alignment, and entity association. The static profile building module is used to build customer static profile feature vectors; The dynamic intent building module is used to construct dynamic intent feature vectors for customers; The scenario adaptation calculation module is used to construct customer-product scenario adaptation feature vectors and output scenario adaptation coefficients; The offline static scoring module is used to output basic conversion scores; The online dynamic correction module is used to output the intended correction score; The integrated scoring module is used to determine the overall conversion probability; The marketing strategy generation module is used to generate marketing action results based on the overall conversion probability and resource constraints; The feedback update module is used to update sample labels, sample weights, model parameters, and policy thresholds based on subsequent feedback behavior.
[0019] A customer target product conversion strategy generation device based on multi-source heterogeneous data, the device includes a processor and a memory, the memory stores a computer program, and the processor executes the computer program.
[0020] A customer target product conversion strategy generation medium based on multi-source heterogeneous data is provided. The generation medium stores a computer program, which, when executed by a processor, is used to implement a customer target product conversion strategy generation method based on multi-source heterogeneous data.
[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention separates and models low-frequency stable attributes and high-frequency behavioral changes, thus avoiding static scoring results from masking short-term changes in customer interest. By introducing a customer-product scenario adaptation coefficient, it can output conversion probabilities that better fit actual execution conditions under different product types, channels, and business constraints. By directly mapping the comprehensive conversion probability to marketing action levels, it can shorten the link between scoring results and marketing execution. By introducing a feedback-driven sample relabeling and model incremental update mechanism, it can maintain the consistency of model parameters with changes in customer behavior when the business environment changes. This solution can be applied to financial products such as bank wealth management, funds, and credit cards, as well as to physical and virtual goods scenarios in enterprise online stores, demonstrating strong transferability. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the structure of the offline static scoring, online dynamic correction, and scene adaptation fusion of the present invention; Figure 3 This is a schematic diagram of the closed loop of marketing action generation and feedback update in this invention; Figure 4 This is a schematic diagram of the time decay aggregation process of customer dynamic behavior in this invention; Figure 5 This is a schematic diagram illustrating the customer-product adaptation constraint calculation for multi-product scenarios according to the present invention. Detailed Implementation
[0023] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0024] Example 1: Overall Processing Flow like Figure 1 As shown, this embodiment provides a method for generating customer target product conversion strategies based on multi-source heterogeneous data. This method is for constructing conversion prediction tasks between customers and target products, including an offline static scoring sub-link, an online dynamic correction sub-link, a scenario adaptation calculation sub-link, a marketing action generation sub-link, and a feedback update sub-link.
[0025] Specifically, the data acquisition module first retrieves raw data from the customer master data system, transaction system, product center system, online behavior collection system, and external environment data system. The customer master data system provides basic customer attributes and account attributes. The transaction system provides historical transaction records, holding records, payment records, refund records, and asset change records. The product center system provides product attributes, risk levels, terms, prices, activity tags, and available sales channels. The online behavior collection system provides page views, product clicks, search keywords, consultation records, favorites records, add-to-cart records, coupon redemption records, and subscription records. The external environment data system provides market interest rates, industry trends, holiday tags, and activity context tags.
[0026] In the data preprocessing stage, outlier identification, field standardization, and primary key unification are performed on the raw data. Then, entity and time alignment are performed based on customer identifier, product identifier, and timestamp. For numeric fields, range pruning and standardization can be used; for categorical fields, encoding mapping can be used; for missing fields, default value filling, historical value imputation, or statistical value interpolation can be used depending on the field type. For data sources with different frequencies, they can be uniformly mapped to daily, hourly, or minute-level time windows.
[0027] During the feature construction phase, fields such as customer age, region, occupation type, risk preference level, asset size, historical holdings structure, historical purchase frequency, average purchase amount, and historical marketing response rate are input into the static profile construction unit to form a customer static profile feature vector. Meanwhile, behaviors such as the number of views, consecutive clicks, search keyword matching degree, dwell time, number of times added to cart, number of times coupons were claimed, number of inquiries, and the interval between the last payment within the most recent preset time window are input into the dynamic intention construction unit, and the dynamic intention feature vector is calculated in combination with the time decay coefficient. .
[0028] In the scenario adaptation calculation phase, for the target product Calculating customer-product matching relationships involves considering the following key characteristics: If the target product is a wealth management product or fund product, the matching characteristics should focus on the relationship between the customer's risk level and the product's risk level, the relationship between available funds and the product's minimum investment threshold, and the relationship between historical investment periods and the product's term. If the target product is a credit card, consumer loan, or insurance product, the matching characteristics should focus on credit conditions, debt ratio, historical card usage behavior, premium affordability, and protection preferences. If the target product is an e-commerce physical or virtual product, the matching characteristics should focus on price range, category preference, inventory status, logistics accessibility, participation in promotional activities, and channel reach availability.
[0029] Next, the static profile feature vector and product attribute feature vector are input into the offline static scoring model to obtain the basic conversion score. The dynamic intention feature vector and historical state summary are input into the online dynamic correction model to obtain the intention correction score. The scene adaptation feature vector is input into the scene adaptation calculation module to obtain the scene adaptation coefficient. The fusion module calculates the overall conversion probability according to the aforementioned formula. .
[0030] The marketing strategy generation module generates marketing actions based on comprehensive conversion probability, channel resource constraints, and frequency control rules. For high-tier customers, actions such as SMS, App push notifications, contacting account managers, or prioritizing homepage display can be generated; for mid-tier customers, general recommendation placement, promotional coupon push notifications, or light contact actions can be generated; for low-tier customers, contact can be delayed, reduced in frequency, or not at all. The action results include not only the target customer set but also the corresponding product identifier, contact channel, priority, contact time period, and contact frequency.
[0031] During the feedback update phase, the system continuously receives feedback events such as customer actions including clicks, inquiries, favorites, subscriptions, orders, payments, activations, unsubscriptions, and rejections. If a customer completes the target conversion within a preset observation window, the corresponding sample is marked as a positive sample; if only superficial behavior occurs and no conversion is completed, it is marked as a weak positive sample or a sample awaiting confirmation; if no effective conversion occurs within a preset time limit, it is marked as a negative sample. The model parameters are then updated based on the sample labels and recentity weights, or incremental training and threshold reestimation are performed when update trigger conditions are met.
[0032] Example 2: Offline Static Scoring Model like Figure 2 As shown in this embodiment, the offline static scoring model is explained. The offline static scoring model reflects a customer's relatively stable purchasing base and long-term preferences, and does not directly respond to real-time behavioral fluctuations.
[0033] Specifically, first, the customer static profile feature vector With product attribute vector Concatenate the data to form an offline scoring input vector. :
[0034] Under the logistic regression approach, the basic transformation score is obtained using the following formula:
[0035] When implemented using a tree model, the basic conversion score can be represented as:
[0036] in, Indicates the first The output of each tree, This represents the corresponding weight. To ensure that the output is within the probability range, a normalization mapping function can be added after the tree model output.
[0037] The training samples for the offline static scoring model can be constructed from historical customer-product interaction records. If a customer completes a purchase, subscription, activation, or contract signing within a preset conversion period, it is marked as a positive sample; otherwise, it is marked as a negative sample. To reduce the impact of sample class imbalance on the model, positive samples can be given higher class weights, or a stratified sampling method can be used to construct the training set.
[0038] Example 3: Online Dynamic Correction Model like Figure 2 and Figure 4 As shown, the online dynamic correction model is used to characterize changes in customer interest and conversion tendencies around a specific target product within a short period of time.
[0039] In this embodiment, a sequence of behavioral events is used as input. Let the set of behavioral events within the preset window be:
[0040] in, Indicates the type of behavior. Indicates the intensity of behavior. This indicates the time when the action occurred. Different event types can correspond to different base weights. For example, browsing behavior has a lower weight than inquiry behavior, inquiry behavior has a lower weight than adding to cart behavior, and adding to cart behavior has a lower weight than confirming payment behavior.
[0041] After applying time decay processing to the set of behavioral events, the dynamic cumulative score of the behavior can be obtained:
[0042] Furthermore, extended features can be constructed based on the diversity, continuity, and interruption of the behavioral sequence, such as the number of consecutive access days, the interval between the last access and the current time, the average interval of the last three behaviors, and whether there is a follow-up purchase link after consultation.
[0043] In a deployable implementation, the online dynamic correction model takes the following form:
[0044] in, The behavior state value retained from the previous cycle can be updated as follows:
[0045] in, This is a state preservation coefficient used to balance changes in historical states and current behavior.
[0046] Through the above processing, the online dynamic correction model can make clear adjustments to the basic conversion score in situations such as continuous clicks, repeated searches, inquiries and conversions, failure to pay after adding items to the cart, and interruptions after submitting applications.
[0047] Example 4: Scene Adaptation Calculation Module like Figure 2 and Figure 5 As shown, the scenario adaptation calculation module is used to map customer attributes, target product attributes, and business constraints to a degree of adaptation, so as to avoid outputting recommendation results that do not conform to scenario constraints based solely on interest scores.
[0048] In this embodiment, the scene adaptation feature vector include:
[0049] in, Indicates the degree of matching of risk levels. Indicates the degree of matching between price ranges. Indicates the degree of time-matching. Indicates channel accessibility, Indicates the relevance of participation in the activity. Indicates regional suitability. Indicates inventory or credit availability. Indicates the time period suitability.
[0050] For financial product scenarios, risk level matching can be used. Defined as:
[0051] in, Indicates the customer's risk level. Indicates the product risk level. This represents the maximum difference in risk levels.
[0052] For price range matching degree , can be defined as:
[0053] in, This indicates the customer's disposable budget or investable amount. Indicates the product price or minimum purchase amount. To prevent smooth terms with a denominator of zero.
[0054] In the comprehensive adaptation calculation, the following can be used:
[0055] When there are mandatory business conditions that must be met, a gating constraint function can be introduced. For financial products with mismatched risk levels, directly order... For goods with zero inventory, the same applies. This process can filter out customer-product pairs that do not meet the hard criteria before fusion.
[0056] Example 5: Integrating Scoring and Marketing Strategy Output like Figure 2 and Figure 3 As shown, in this embodiment, the overall conversion probability is determined by the base conversion score, the intention correction score, and the scenario adaptation coefficient. Log-probability space weighted fusion is preferably used to reduce the impact of single-path extreme values on the overall score. Its expression is as follows:
[0057]
[0058] Among them, the cross-action term This is used to characterize whether a customer's long-term preferences are consistent with their short-term behavior. If a customer has a long-term preference for a certain type of stable product and has been consistently browsing this type of product recently, then... Positive values are acceptable; however, if a customer's long-term preferences are contrary to their short-term behavior, then... A suitable suppression value can be selected. One possible implementation is as follows:
[0059] in, This is a mapping matrix from static space to dynamic space. Once the overall conversion probability is determined, it can be used based on the marketing resource constraint matrix. Generate action strategies based on preset frequency control rules. Let the set of selectable marketing actions be... The target action can then be selected based on the following objective function:
[0060] in, Indicates action For customers and products Channel compatibility value, Indicates the cost of the action. This indicates a frequency control penalty item.
[0061] When the overall conversion probability is higher than the threshold When the account manager channel is available, you can choose between manual outbound calls or recommendations from dedicated consultants; when the overall conversion rate is between and During this period, you can choose App push, SMS, or in-app message; when the overall conversion rate is lower than At that time, it may be temporarily not necessary to reach it.
[0062] Example 6: Feedback Update and Model Adaptation like Figure 3 As shown in this embodiment, the feedback update module receives subsequent customer actions and constructs a closed-loop training mechanism.
[0063] If a customer completes the target behavior within the observation window, such as completing a wealth management subscription, fund purchase, credit card application submission, ordering goods, or paying for virtual products, the corresponding sample will be recorded as a positive sample. If a customer engages in superficial behaviors such as clicking, browsing, saving, or consulting but does not complete the final purchase, the sample will be recorded as a weak positive sample according to business rules. If a customer does not convert effectively after the observation window ends, the sample will be recorded as a negative sample.
[0064] To make model updates more relevant to recent behavioral changes, samples can be assigned dynamic weights:
[0065] in, The time-related decay coefficient of the sample. Indicates product priority. This indicates the importance of the channel to which the customer belongs.
[0066] The aforementioned loss function can be minimized during model updates. And re-estimate the threshold after each update. , and This ensures that the reach, budget expenditure, and expected conversions at different marketing levels meet the business-side resource requirements.
[0067] In one implementation, the update triggering condition includes at least one of the following: reaching a preset sample quantity threshold, reaching a preset time period, a change in the target product's shelf status, the start or end of a target activity, or fluctuations in market environment variables exceeding a preset threshold.
[0068] Example 7: Based on the above method, the present invention also provides a customer target product conversion strategy generation device based on multi-source heterogeneous data, the device comprising: The data acquisition module is used to acquire multi-source basic customer data, target product data, real-time customer behavior data, and external environment data. The data preprocessing module is used to perform cleaning, standardization, missing data repair, time alignment, and entity association. The static profile building module is used to build customer static profile feature vectors; The dynamic intent building module is used to construct dynamic intent feature vectors for customers; The scenario adaptation calculation module is used to construct customer-product scenario adaptation feature vectors and output scenario adaptation coefficients; The offline static scoring module is used to output basic conversion scores; The online dynamic correction module is used to output the intended correction score; The integrated scoring module is used to determine the overall conversion probability; The marketing strategy generation module is used to generate marketing action results based on the overall conversion probability and resource constraints; The feedback update module is used to update sample labels, sample weights, model parameters, and policy thresholds based on subsequent feedback behavior.
[0069] Example 8: Application Case Taking the recommendation of bank wealth management products as an example, customers The risk level is Level 3, the investable asset size is 200,000 yuan, and there are three low-risk wealth management product subscription records in the past six months. In the last two days, the customer browsed a medium-to-low risk wealth management product page five times consecutively, viewing the return and term descriptions on the page. After obtaining the customer's basic information, asset information, holding information, browsing behavior information, and market interest rate information, the system constructs a static profile feature vector and a dynamic intention feature vector. For this wealth management product, the scenario adaptation calculation module obtains a high degree of matching in risk level and capital threshold. The offline static scoring model outputs a basic conversion score, and the online dynamic correction model outputs a higher intention correction score based on continuous browsing and page dwell behavior. The fusion yields a high overall conversion probability. Based on this, the marketing strategy generation module selects to reach the customer through mobile banking homepage recommendations and customer manager message reminders. If the customer subsequently completes a subscription, the sample is marked as a positive sample and participates in subsequent incremental updates.
[0070] Taking the scenario of recommending virtual products in an online store as an example, a customer has browsed a certain type of membership benefit product multiple times in the past month and spent a considerable amount of time on the activity page, but has few previous purchase records. The system calculates relevant characteristics based on historical spending levels, device activity, coupon redemption records, page dwell times, activity period tags, and product price ranges. Because this customer has continuous access and coupon redemption behavior during the activity period, the dynamic intention correction score increases significantly; at the same time, the virtual product has sufficient inventory, a short payment process, and high accessibility to marketing channels, resulting in a high scenario suitability coefficient. Ultimately, the system generates in-site pop-ups and limited-time offer reminders.
[0071] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. A method for generating customer target product conversion strategies based on multi-source heterogeneous data, characterized in that, include: S1. Obtain multi-source basic data, target product data, real-time behavioral data, and external environment data of the customer to be evaluated; S2. Clean, standardize, repair missing data, align time, and associate entities with the multi-source basic data, target product data, real-time behavior data, and external environment data to obtain a unified sample dataset. S3. Construct customer static profile feature vector, customer dynamic intention feature vector, and customer-product scenario adaptation feature vector based on the unified sample dataset; S4. Input the customer static profile feature vector into the offline static scoring model to obtain the basic conversion score; S5. Input the customer dynamic intention feature vector into the online dynamic correction model to obtain the intention correction score; S6. Determine the scenario adaptation coefficient based on the customer-product scenario adaptation feature vector; S7. The basic conversion score, intention correction score and scenario adaptation coefficient are combined to obtain the comprehensive conversion probability of the customer to be evaluated for the target product. S8. Generate marketing strategy output results based on comprehensive conversion probability and preset marketing resource constraints; S9. Update sample labels, sample weights, model parameters, and marketing strategy thresholds based on subsequent customer feedback.
2. The method for generating customer target product conversion strategies based on multi-source heterogeneous data according to claim 1, characterized in that: The customer static profile feature vector described in step S3 is constructed based on at least two or more of the following fields: customer basic information, account level, asset size, historical holdings structure, credit information, historical transaction frequency, historical purchase amount, historical product category distribution, and historical marketing response rate.
3. The method for generating customer target product conversion strategies based on multi-source heterogeneous data according to claim 1, characterized in that: The customer dynamic intention feature vector mentioned in step S3 is constructed based on browsing, clicking, searching, consulting, adding to favorites, adding to cart, claiming coupons, subscribing, placing orders, paying, and activating behaviors within a preset time window, and is aggregated according to behavior type weights and time decay rules.
4. The method for generating customer target product conversion strategies based on multi-source heterogeneous data according to claim 3, characterized in that, The time decay rule satisfies: in, Indicates the first The contribution value of each behavioral event to the dynamic intention feature. Indicates the weight of behavior type. Indicates the intensity of behavior. Indicates the time when the action occurred. Indicates the current scoring time. This represents the time decay coefficient.
5. The method for generating customer target product conversion strategies based on multi-source heterogeneous data according to claim 1, characterized in that: The customer-product scenario matching feature vector mentioned in step S3 includes one or more of the following: risk level matching degree, price range matching degree, term matching degree, channel accessibility, activity participation degree, and inventory or quota availability.
6. The method for generating customer target product conversion strategies based on multi-source heterogeneous data according to claim 1, characterized in that: The specific formula for integrating the basic conversion score, intention correction score, and scenario adaptation coefficient in step S7 is as follows: in, Indicates the basic conversion score. Indicates intention to adjust score, Indicates the scene adaptation coefficient. Indicates the cross-action term. This represents the scene constraint gating term, where α, β, γ, and μ are the fusion weights. For the overall score; for function; For the final predicted probability, The output of the marketing strategy described in step S8 includes one or more of the following: target customer set, target product identifier, reach channel, reach time period, reach frequency, and priority.
7. The method for generating customer target product conversion strategies based on multi-source heterogeneous data according to claim 1, characterized in that: Step S9, which involves updating sample labels, sample weights, model parameters, and marketing strategy thresholds based on subsequent customer feedback behavior, includes: Receive subsequent customer feedback within a preset observation window; Based on whether the customer has achieved the target conversion, the samples are labeled as positive samples, weakly positive samples, or negative samples; Incremental training or periodic retraining is performed based on the updated sample labels and sample weights; Re-estimate the marketing strategy thresholds based on the updated forecast results; The sample weights satisfy: in, This represents the overall weighted preference value. , and These are three learnable weight coefficients; For time decay term, Here is the current timestamp, and t is the current rating time. The time-related decay coefficient of the sample. Prioritize the target product. The importance of the customer's channel.
8. A customer target product conversion strategy generation device based on multi-source heterogeneous data, applied to the customer target product conversion strategy generation method based on multi-source heterogeneous data as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire multi-source basic customer data, target product data, real-time customer behavior data, and external environment data. The data preprocessing module is used to perform cleaning, standardization, missing data repair, time alignment, and entity association. The static profile building module is used to build customer static profile feature vectors; The dynamic intent building module is used to construct dynamic intent feature vectors for customers; The scenario adaptation calculation module is used to construct customer-product scenario adaptation feature vectors and output scenario adaptation coefficients; The offline static scoring module is used to output basic conversion scores; The online dynamic correction module is used to output the intended correction score; The integrated scoring module is used to determine the overall conversion probability; The marketing strategy generation module is used to generate marketing action results based on the overall conversion probability and resource constraints; The feedback update module is used to update sample labels, sample weights, model parameters, and policy thresholds based on subsequent feedback behavior.
9. A customer target product conversion strategy generation device based on multi-source heterogeneous data, applied to the customer target product conversion strategy generation method based on multi-source heterogeneous data as described in any one of claims 1 to 7, characterized in that, The device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program.
10. A customer target product conversion strategy generation medium based on multi-source heterogeneous data, applied to the target product conversion strategy generation method based on multi-source heterogeneous data as described in any one of claims 1 to 7, characterized in that: The generating medium stores a computer program, which, when executed by a processor, is used to implement the method described in any one of claims 1 to 7.