Product recommendation method, device, apparatus, and storage medium

Abstract

The present application belongs to the technical field of data processing, and discloses a product recommendation method, device, equipment and storage medium, the recommendation method comprising: constructing a strategy evolution sub-population corresponding to a recommendation strategy of different customers in each customer sub-group; constructing a strategy fitness function fused with real-time external risk signals; based on the strategy fitness function, driving multiple strategy evolution sub-populations to execute iterative optimization in parallel, and performing strategy knowledge migration operation across strategy evolution sub-populations in the iterative process; outputting a set of recommended strategies from the optimized strategy evolution sub-population, and generating explanation information associated with the strategy evolution trajectory of the set of recommended strategies. The present application can be applied to financial technology, medical health and other business management systems, and solves the technical problem of large workload in recommending products to customers according to the real needs of users when the mapping relationship between user characteristic values and product recommendations changes in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and is applied in the fields of fintech and healthcare. In particular, it relates to a product recommendation method, apparatus, device, and storage medium. Background Technology

[0002] In various customer service sectors, it's necessary to provide diverse services. To ensure customer satisfaction, it's crucial to recommend products that truly meet their needs. For example, in the insurance sector, if a customer needs car insurance, their actual needs must be determined based on factors such as their salary level and vehicle value. For instance, if a customer's salary is low, cheaper car insurance should be recommended; if their vehicle is valuable, insurance with higher claim amounts should be recommended. Similarly, in the financial sector, if various medical procurement groups need to purchase medical equipment, their actual purchasing needs must be determined based on their purchasing power and diagnostic / usage requirements.

[0003] However, currently, when determining a customer's true needs and recommending relevant products based on those needs, the recommendation is determined solely by the mapping relationship between different customer characteristics and corresponding products. But this mapping relationship is fixed. When the mapping relationship between characteristics and products changes, this mapping relationship needs to be re-established one by one, which is a huge undertaking. Summary of the Invention

[0004] This invention provides a product recommendation method, apparatus, device, and storage medium, which can solve the technical problem in the prior art of recommending products to customers based on the user's actual needs when the mapping relationship between user feature values ​​and product recommendations changes.

[0005] In a first aspect, the present invention provides a product recommendation method, comprising: Acquire at least one customer segment; Construct strategy evolution subpopulations corresponding to different customers in each of the aforementioned customer segments, wherein different strategy evolution subpopulations correspond to differentiated strategy search spaces; Construct a policy fitness function that integrates real-time external risk signals; Based on the policy fitness function, the multiple policy evolution subpopulations are driven to perform iterative optimization in parallel, and policy knowledge transfer operations are performed across the policy evolution subpopulations during the iteration process; Output a recommended strategy set from the optimized strategy evolution subpopulation, and generate explanatory information related to the strategy evolution trajectory of the recommended strategy set.

[0006] Secondly, the present invention provides a product recommendation device, comprising: The acquisition module is used to acquire at least one customer segment. The first construction module is used to construct a strategy evolution subpopulation formed by recommendation strategies for different customers in each of the customer segments, wherein different strategy evolution subpopulations correspond to differentiated strategy search spaces; The second construction module is used to construct a policy fitness function that integrates real-time external risk signals; The population optimization module is used to drive the multiple policy evolution subpopulations to perform iterative optimization in parallel based on the policy fitness function, and to perform policy knowledge transfer operations across the policy evolution subpopulations during the iteration process; The information generation module is used to output a recommended strategy set from the optimized strategy evolution subpopulation and generate explanatory information related to the strategy evolution trajectory of the recommended strategy set.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the product recommendation method described above.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the product recommendation method described above.

[0009] Understandably, the first step of this invention is to acquire at least one customer segment, dividing customers into multiple segments based on their core characteristics. This upgrades the fragmented mapping of "single feature - single product" to a unified management of a group-level strategy pool. In this approach, when the mapping relationship changes, there's no need for manual adjustments to the matching relationship for each individual feature value; only the relevant rules need to be adjusted for each customer segment. This significantly reduces the workload of basic maintenance and solves the problem of high workload at the management granularity level.

[0010] Secondly, the second step of the solution, "constructing strategy evolution sub-populations corresponding to different customers within each customer segment, with different strategy evolution sub-populations corresponding to differentiated strategy search spaces," builds a dedicated recommendation strategy pool for each customer segment and sets a differentiated strategy search space for each sub-population. When the mapping relationship changes, there is no need to manually modify the product corresponding to each feature value one by one. It is only necessary to simply adjust the boundary of the strategy search space of the corresponding sub-population. All recommendation strategies within the sub-population will autonomously explore and evolve within the new search space boundary, automatically adapting to the new mapping relationship. There is no need for manual intervention to match them one by one, thus eliminating a large amount of manual workload in terms of adaptation method.

[0011] Furthermore, the third step of the solution constructs a strategy fitness function that integrates real-time external risk signals, providing an automatically updatable evaluation criterion for the autonomous evolution of the strategy. This fitness function is not fixed but can integrate external change signals (external risk signals) in real time. These external changes are essentially the core reasons that cause changes in the "user feature value-product" mapping relationship. The function automatically transforms these external changes into evaluation criteria for strategy evolution, eliminating the need for manually redefining the "feature-product" matching rules. This allows the strategy to automatically evolve in the direction of adapting to the new mapping relationship, avoiding the enormous workload of manually reconstructing the rules in terms of evaluation criteria.

[0012] Then, the fourth step of the solution, "based on the policy fitness function, drives multiple policy evolution subpopulations to perform iterative optimization in parallel, and performs policy knowledge transfer operations across policy evolution subpopulations during the iteration process," achieves fully automated optimization of recommended policies and reuse of high-quality policies. Driven by the fitness function, each policy evolution subpopulation autonomously performs iterative optimization in parallel: automatically calculating the fitness of each policy, selecting high-quality policies that fit the new mapping relationship, and eliminating invalid policies that do not conform to the new mapping relationship; at the same time, the cross-population policy knowledge transfer operation can reuse the high-quality policies of one group to other groups with similar needs, without the need for manual establishment of similar mapping relationships for different groups, further reducing the workload of duplication. The entire iterative optimization and knowledge transfer process is fully automated, without the need for manual intervention in screening, matching, and reuse, completely eliminating the large amount of workload of manually re-screening and matching policies when the mapping relationship changes, which is the core link in solving this technical problem.

[0013] Finally, the fifth step of the solution, "outputting a recommended strategy set from the optimized strategy evolution subpopulation and generating explanatory information about the strategy evolution trajectory association of the recommended strategy set," completes the automatic output and explanation of high-quality strategies under the new mapping relationship. The final output of the solution is a high-quality recommended strategy set that has undergone automated iterative optimization, rather than a fixed mapping table. This set can be directly used to recommend products to customers without the need for manual screening and organization of strategies. Simultaneously, the solution automatically generates explanatory information about the strategy evolution trajectory, further reducing subsequent manual workload.

[0014] In summary, when the recommendation mapping relationship between "user feature value - product" changes, existing technologies require manual intervention throughout the entire process, resulting in a huge workload. However, this solution, through the synergistic effect of the above five steps, forms a complete closed loop of "group overall management - search space adaptation - automatic update of evaluation criteria - automatic iterative optimization of strategies - automatic output of high-quality strategies." Only a few core parameters (such as search space boundaries and fitness function weights) need to be adjusted manually. All other processes for adapting to the new mapping relationship are completed automatically by the solution. There is no need to manually re-establish the recommendation mapping relationship between feature values ​​and products one by one, thus fundamentally solving the technical problem of the large workload of recommending products when the mapping relationship changes in existing technologies. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a product recommendation method in one embodiment of the present invention.

[0017] Figure 2 yes Figure 1 A flowchart of step S120.

[0018] Figure 3 yes Figure 1 Another flowchart of step S120.

[0019] Figure 4 yes Figure 1 A flowchart of step S140.

[0020] Figure 5 This is a schematic diagram of a product recommendation device in one embodiment of the present invention.

[0021] Figure 6 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention.

[0022] Figure 7 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Figure 1 A flowchart of the product recommendation method provided in the embodiments of the present invention is shown below. Figure 1 As shown, the product recommendation method provided in this embodiment of the invention includes the following steps.

[0025] Step S110: Obtain at least one customer segment; Specifically, in this step, the core characteristic dimensions of customer segmentation can be determined and can be divided into static characteristics (such as demographics, assets / purchasing capacity, industry attributes, basic health / medical needs), dynamic characteristics (such as recent browsing / consultation / purchasing behavior, interaction frequency, and changes in needs), preference characteristics (such as product selection preferences, acceptance of marketing channels, risk tolerance, and service demand preferences), and scenario characteristics (such as usage / consumption scenarios, geographical attributes, and time attributes), forming a standardized characteristic library.

[0026] More specifically, a combination of hierarchical clustering and DBSCAN density clustering can be used. First, hierarchical clustering is used to determine the initial number of clusters, and then DBSCAN is used to accurately segment high-density customer groups. At the same time, the K-means algorithm is used to verify the results. The evaluation index of the clustering algorithm can be the silhouette coefficient (SC greater than or equal to 0.6 is considered effective clustering) and the Calinski-Harabasz index (the higher the index, the better the clustering effect), etc.

[0027] As a concrete example, in the insurance scenario, we can take the individual customers of life insurance companies as the target group and select age, annual income, premium budget ratio, type of protection needs, risk tolerance and channel preference as core characteristics. After clustering, we can obtain three core customer segments: the middle-aged and elderly critical illness protection group, the young and light protection group, and the family comprehensive protection group. The middle-aged and elderly critical illness protection group can be those aged 55-70, retired, with an annual income of 50,000-150,000 yuan, a premium budget of 5%-10% of their annual income, whose core needs are critical illness / cancer prevention / million-yuan medical insurance, conservative risk tolerance, and who prefer offline agent consultation. The young and youth lightweight protection group can be those aged 20-35, new graduates / white-collar workers, with an annual income of 60,000-200,000 yuan, a premium budget of less than 5% of their annual income, whose core needs are million-yuan medical insurance and accident insurance, neutral risk tolerance, and who prefer online self-service insurance purchase. The family comprehensive protection group can be those aged 30-45, married with children, with an annual income of 200,000-500,000 yuan, a premium budget of 8%-15% of their annual income, whose core needs are critical illness insurance for the whole family, million-yuan medical insurance, and children's education funds, stable risk tolerance, and who accept a combination of online and offline marketing methods.

[0028] As another concrete example, in the medical technology scenario, we can take the customers of medical device companies as the target, select customer identity, purchasing power, diagnosis / use needs, equipment operation capabilities, purchasing channel preferences, and after-sales needs as core features, and obtain four core customer segments after clustering. These four customer segments can be categorized as follows: Top-tier hospitals purchasing high-end equipment; community hospitals purchasing basic medical devices; home-based patients consuming portable equipment; and private institutions purchasing mid-range equipment. Top-tier hospitals purchasing high-end equipment can be general / specialized top-tier hospitals with ample budgets, whose core need is high-end diagnostic / surgical equipment, with customized after-sales requirements, and who prefer offline industry exhibitions and dedicated sales connections. Community hospitals purchasing basic medical devices can be community health service centers / township hospitals with limited budgets, whose core need is basic rehabilitation / monitoring equipment, requiring easy-to-operate devices, and who prefer regional distributors and online purchasing platforms. Home-based patients consuming portable equipment can be chronic disease / post-operative home-based patients without professional operating skills, whose core need is portable monitoring / rehabilitation equipment, seeking high cost-effectiveness, and who prefer e-commerce platforms and online customer service connections. Private institutions purchasing mid-range equipment can be private clinics / medical aesthetic institutions with moderate budgets, whose core need is mid-range diagnostic / treatment equipment, requiring low maintenance costs, and who prefer direct supply from manufacturers and bulk purchase discounts.

[0029] Step S120: Construct a strategy evolution subpopulation for each customer segment corresponding to a different customer in each of the aforementioned customer segments.

[0030] Different policy evolution subpopulations correspond to differentiated policy search spaces.

[0031] Specifically, in this step, the strategy evolution subpopulation is mapped one-to-one with the customer segment, so that a dedicated strategy evolution subpopulation can be configured for each customer segment. To this end, the identifier of the strategy evolution subpopulation can be bound to the name of the customer segment, ensuring that the optimization direction of the strategy evolution subpopulation is highly matched with the customer needs of the corresponding customer segment.

[0032] Specifically, each individual in the strategy evolution subpopulation represents a complete product recommendation strategy for the corresponding customer group. The core attributes of each individual may include product portfolio schemes, marketing strategy schemes, and adaptability constraint parameters. More specifically, the core of designing a differentiated strategy search space can be to design differentiated strategy search spaces for different subgroups from four dimensions: search dimensions, search scope, search constraints, and weight allocation, to avoid local optima caused by the convergence of optimization directions among multiple subgroups. The specific design rules are as follows: At the search dimension level, different subgroups can emphasize different product / marketing dimensions (e.g., low-risk groups emphasize product risk level, while high-purchasing-capacity groups emphasize product customization); at the search scope level, the range of products and marketing channels that subgroups can choose can be limited (e.g., excluding high-premium short-term products for middle-aged and elderly groups, and excluding high-end imported equipment for community hospitals); at the search constraint level, hard constraints on product combinations / marketing strategies can be set (e.g., monthly premiums for young people are less than 500 yuan, and the number of steps for operating equipment for home-based patients is less than 3); at the weight allocation level, differentiated weights can be set for the core indicators of products / marketing (e.g., in insurance scenarios, the weight for "comprehensive coverage" is 0.4 for middle-aged and elderly groups, while the weight for "cost-effectiveness" is 0.5 for young people). More specifically, when initializing the strategy evolution subpopulation, initial individuals can be selected based on historical recommendation data and industry expert experience. The number of initial individuals is set according to the search space complexity (e.g., 50-200 per subpopulation). At the same time, individuals can be supplemented by random generation and data augmentation to ensure the diversity of initial individuals (e.g., the similarity of initial individuals is less than 0.5 to avoid local optima).

[0033] As a concrete example, in the insurance scenario, strategy evolution subpopulations can be constructed for the three customer groups obtained in step S110 to achieve differentiation of the search space.

[0034] For example, targeting the critical illness protection subgroup for middle-aged and elderly people, the search scope can be limited to a combination of whole life critical illness insurance, million-dollar medical insurance, and cancer prevention insurance. The marketing channels can be limited to offline agents, community science popularization, and children's communities. The hard constraints can be that the annual premium is less than 10% of the customer's annual income and the coverage includes common diseases of the elderly such as hypertension and diabetes. The core dimension weights can be comprehensive coverage 0.4, convenient claims 0.3, and premium stability 0.2. The initial number of individuals can be 80. For example, targeting the lightweight protection subgroup of young people, the search scope can be limited to lightweight combinations of one-year million medical insurance, comprehensive accident insurance, and small-amount critical illness insurance. The marketing channels are limited to short video platforms, mini-programs, and online communities. The hard constraints are that the monthly premium is less than 500 yuan and the threshold for first-time investment discounts is low. The core dimensions are weighted as follows: cost-effectiveness 0.5, ease of application 0.3, and protection flexibility 0.1. The initial number of individuals can be 60. For the family comprehensive protection subgroup, the search scope can be limited to a combination of family-shared million-dollar medical insurance, critical illness insurance for parents, critical illness insurance for children, and family accident insurance. The marketing channels can be online mini-programs and offline family financial advisors. The hard constraints can be that the family's annual premium is less than 15% of the family's annual income and includes coverage for children's special diseases / parents' chronic diseases. The core dimensions can be weighted as family suitability 0.4, cost-effectiveness 0.3, and repurchase rate 0.2. The initial number of individuals can be 100.

[0035] As another concrete example, in the medical technology scenario, strategy evolution subpopulations can be constructed for the four customer groups obtained by S110 to achieve differentiation of the search space.

[0036] Specifically, for the high-end equipment subgroup of tertiary hospitals, the search scope can be limited to imported / high-end domestic CT / MRI / mass spectrometers, supporting post-processing systems, and customized after-sales services. The marketing channels can be limited to offline industry exhibitions, manufacturers' dedicated sales teams, and academic seminars. The hard constraints can be that the equipment is compatible with the annual treatment volume of tertiary hospitals and supports customized function development. The core dimension weights can be equipment performance 0.4, after-sales customization 0.3, and brand reputation 0.2. The initial number of individuals can be 120. For the basic medical equipment subgroup of community hospitals, the search scope can be limited to domestically produced basic rehabilitation equipment, simple monitoring equipment, and bulk purchase discounts. The marketing channels can be limited to regional distributors and online medical equipment procurement platforms. The hard constraints can be that the equipment operation steps are less than 3 steps and the annual maintenance cost is less than 5% of the total equipment price. The core dimension weights can be ease of operation 0.4, cost-effectiveness 0.3, and maintenance convenience 0.2. The initial number of individuals can be 90. For the portable device subgroup for home-based patients, the search scope can be limited to portable monitoring / rehabilitation devices, supporting consumables, and free calibration services. The marketing channels can be limited to e-commerce platforms, community group buying, and online customer service. The hard constraints can be that the device weight is less than 2kg, supports one-click operation, and the consumables are cost-effective. The core dimension weights can be portability 0.4, cost-effectiveness 0.3, and after-sales convenience 0.2. The initial number of individuals can be 70. For the subgroup of mid-range equipment in private institutions, the search scope can be limited to mid-range domestic diagnostic / treatment equipment, basic maintenance and bulk purchase discounts. The marketing channels can be limited to direct supply from manufacturers and offline regional promotion meetings. The hard constraints can be that the equipment investment return cycle is less than 2 years and the maintenance cost is low. The core dimension weights can be cost-effectiveness 0.4, return on investment 0.3 and equipment stability 0.2. The initial number of individuals can be 80.

[0037] In some embodiments of the present invention, such as Figure 2 As shown, step S120 includes the following steps.

[0038] Step S121: For each customer in each of the aforementioned customer segments, match the corresponding product portfolio and marketing strategy for each customer.

[0039] Specifically, in this step, a feature extraction algorithm can be used to reduce the dimensionality of each customer's standardized features (static, dynamic, preference, scenario) to a 5-8 dimensional feature vector, while retaining core feature information (the variance contribution rate of core feature information is greater than or equal to 90%). A vector normalization algorithm is used to map the feature vector to the [0,1] interval, which facilitates similarity calculation.

[0040] More specifically, when matching product combinations for each customer, features of products in the product library can be extracted and encoded according to dimensions such as category, price, core function, and applicable scenarios to build a product feature library; inverted index technology is used to associate product features with customer features for easy and fast matching. Furthermore, a fusion algorithm of collaborative filtering and logistic regression can be adopted. On the one hand, through user collaborative filtering, the cosine similarity between the current customer and other customers in the same segment is calculated, and customers with a similarity greater than or equal to 0.7 are selected, and their preferred product combinations are used as candidate combinations. On the other hand, the logistic regression algorithm is used, with the customer feature vector as input and the product combination suitability as output (0-1 interval), to calculate the suitability score of each candidate product combination, and product combinations with a score greater than or equal to 0.7 are selected as the initial product combination for the customer. Furthermore, a greedy algorithm can be used to optimize the initial product portfolio, eliminating redundant products (products with a functional overlap with other products greater than or equal to 0.8) and supplementing missing products (products that do not cover the core needs of customers).

[0041] More specifically, when matching marketing strategies to customers, features of all marketing strategies can be extracted and encoded according to dimensions such as channel, frequency, and discount method to build a marketing strategy feature library; and classification algorithms can be used to classify marketing strategies according to customer preference features (such as middle-aged and elderly customers - offline channels, young customers - online channels). Furthermore, a decision tree algorithm can be used to construct a marketing strategy matching model with customer preference features and scenario features as input and marketing strategy type as output. Cross-validation (10-fold cross-validation) is used to ensure that the model accuracy is greater than or equal to 85%. The feature vector of each customer is input into the model, and the appropriate marketing strategy type is output. Then, a random forest algorithm is used to optimize the specific parameters of the marketing strategy (such as reach frequency and discount level) to generate the initial marketing strategy for the customer.

[0042] As a concrete example, in an insurance scenario, the PCA algorithm can be used to reduce the dimensionality of the characteristics of middle-aged and elderly critical illness insurance customers to 6 dimensions (variance contribution rate of 92%), and then normalize them for matching. An inverted index technique is used to construct an insurance product feature library, linking customer critical illness insurance, offline consultation, and other features. Furthermore, a collaborative filtering algorithm is used to select 20 customers with a similarity greater than or equal to 0.7, extracting their preferred combinations of "lifetime critical illness, million-dollar medical insurance, and cancer prevention insurance" as candidates. A logistic regression algorithm calculates the suitability score, selecting combinations with scores greater than or equal to 0.7 as the initial product combination. A greedy algorithm eliminates short-term medical insurance with overlapping functions and supplements cancer prevention insurance clauses. A decision tree model is used to match marketing strategies from offline agents and community outreach, and a random forest algorithm optimizes the reach frequency to twice a month and the discount to 9.5% of the premium, generating the initial marketing strategy for the customer.

[0043] As another concrete example, in the medical technology scenario, the PCA algorithm can be used to reduce the dimensionality of community hospital customer groups to 5 dimensions (variance contribution rate of 91%); an inverted index is used to build a medical equipment feature library, associating basic instruments and ease of operation features; further, collaborative filtering is used to select 15 community hospital customers with a similarity greater than or equal to 0.7, extracting their preferred combinations of basic rehabilitation instruments and simple monitoring equipment as candidates, logistic regression is used to calculate the suitability score, and combinations with a score greater than or equal to 0.7 are selected; a greedy algorithm is used to remove redundant high-end monitoring functions and supplement bulk purchase terms; a decision tree model is used to match the marketing strategies of regional distributors and online purchasing platforms, and a random forest is used to optimize the reach frequency to once per quarter and the bulk purchase discount to 10%, generating an initial marketing strategy.

[0044] Step S122: Chromosome encoding is performed on the product portfolio and marketing strategy of each customer to obtain the corresponding product portfolio gene segment and marketing strategy gene segment for each customer.

[0045] Specifically, in this step, when encoding the product components, product selection and product priority can be encoded separately.

[0046] When selecting and coding products, binary coding can be used, so that the code length equals the number of suitable products in the corresponding customer segment. Each bit of the code can correspond to one product, with 1 indicating that the product is included in the combination and 0 indicating that it is not included. Bitwise operation technology is used to quickly locate the products included in the combination, and the coding efficiency is greater than or equal to 100 customers / second. When prioritizing products, real-number encoding can be used, ensuring that the encoding length matches the product selection encoding. Each real number (0-1 range) corresponds to the priority of a product included in the combination. The closer the real number is to 1, the higher the priority. A linear mapping algorithm is used to convert the product suitability score into a priority encoding value. When the suitability score is greater than or equal to 0.9, the encoding value is 0.9-1.0; when the suitability score is 0.7-0.9, the encoding value is 0.7-0.9.

[0047] More specifically, when coding marketing strategies, channel investment and reach frequency / discount intensity can be coded.

[0048] When coding channel investment, real number coding can be used. The coding dimension is the number of marketing channels that are suitable for this customer group. Each real number (0-1 range) corresponds to the investment ratio of one channel, and the sum of the investment ratios of all channels is 1. A normalization algorithm is used to ensure that the coding value meets the constraints. When encoding the frequency of reach / discount, integer encoding can be used. The frequency of reach encoding value corresponds to the number of reach per month (1-5 times), and the discount encoding value corresponds to the discount rate (80-100, corresponding to 80% off - original price). An interval mapping algorithm is used to convert the actual parameter values ​​into integer encoding, which facilitates genetic operations.

[0049] As a concrete example, in an insurance scenario, when there are 12 products in the product library suitable for the youth light-weight protection group, the product combination gene segment code length is 12. The binary code "101001000100" can represent the inclusion of the 1st, 3rd, 6th, and 10th products (one-year million medical insurance, comprehensive accident insurance, etc.); the real number code "0.85 0 0.9 0 0 0.8 0 0 0 0.750 0" can correspond to the priority of each product, with the 3rd product having the highest priority (0.9); the marketing strategy gene segment includes 3 channels (short video, mini-program, and community), and the real number code "0.5 0.3 0.2" corresponds to the channel investment ratio; the reach frequency code is "3" (3 times per month), and the discount code is "90" (10% discount); during code verification, one real number code that exceeds the range (1.05-1.0) is corrected, and hash verification ensures the integrity of the code.

[0050] As another concrete example, in the medical technology scenario, when there are 15 products in the product library suitable for portable devices for home-based patients, the product combination gene segment code length is 15. The binary code "100100100001000" indicates that the 1st, 4th, 7th, and 12th products (portable monitors, rehabilitation equipment, etc.) are included; the real number code "0.8 0 0 0.88 0 0 0.78 0 0 00 0.72 0 0 0" corresponds to the product priority; the marketing strategy gene segment includes 3 channels (e-commerce, social media, online customer service), and the real number code "0.6 0.2 0.2" corresponds to the channel investment ratio; the reach frequency code is "2" (twice a month), and the discount code is "85" (8.5% off); the code verification has no illegal values, and the hash verification passes.

[0051] Step S123: Combine the product portfolio gene segment and the marketing strategy gene segment corresponding to each customer to form a chromosome corresponding to each customer, which serves as an individual in the strategy evolution subpopulation.

[0052] Specifically, when combining product portfolio gene segments and marketing strategy gene segments, a gene segment splicing algorithm can be used to splice the product portfolio gene segment (containing binary and real numbers) and the marketing strategy gene segment (containing real and integer numbers) in the order of "product portfolio first, marketing strategy second" to form a complete chromosome. During the splicing process, delimiter encoding can be used to distinguish different types of gene segments, which facilitates gene segment cutting during subsequent genetic operations. Vector splicing technology can be used to convert gene segments into a unified chromosome vector, with the vector length equal to the length of the product portfolio gene segment plus the length of the marketing strategy gene segment.

[0053] As a concrete example, in an insurance scenario, the product portfolio gene segment (length 24) and marketing strategy gene segment (length 8) of a customer in the middle-aged and elderly critical illness protection group are concatenated using a splicing algorithm, separated by delimiters, to form a chromosome vector of length 32. Regular expression validation is used to confirm that the gene segment order is correct and the delimiter position is correct. Correlation analysis shows that the correlation between the product portfolio gene segment (critical illness insurance and cancer insurance) and the marketing strategy gene segment (offline agents and community science popularization) is 0.72, which is greater than 0.6, making it suitable for concatenation. The similarity with existing individuals in the subpopulation is calculated to be 0.45, which is less than 0.5, so it is directly included in the subpopulation as the initial individual. Shannon entropy calculation shows that the diversity entropy value of the subpopulation is 0.68, which is qualified.

[0054] As another concrete example, in the medical technology scenario, the product portfolio gene segment (length 30) and the marketing strategy gene segment (length 10) of a certain customer in the high-end equipment group of a top-tier hospital are spliced ​​together to form a chromosome vector of length 40, separated by delimiters; the structure is verified by regular expressions; correlation analysis shows that the correlation between the product portfolio gene segment (high-end CT plus customized after-sales service) and the marketing strategy gene segment (offline exhibitions plus exclusive sales) is 0.75, which is greater than 0.6, and is suitable for matching; the similarity with existing individuals is 0.48, which is less than 0.5, and it is included in the subpopulation; the diversity entropy value of the subpopulation is 0.66, which is qualified.

[0055] Understandably, the above embodiments transform product recommendation strategies into heritable, evolvable, and quantifiable genetic chromosome structures, constructing a standardized individual foundation for iterative optimization of subpopulations. This addresses the technical pain points of traditional insurance recommendation strategies, such as unstructured nature, inability to perform genetic operations, and the inability to effectively reuse strategies across populations. The above embodiments, through a hybrid encoding technology of product portfolio gene segments and marketing strategy gene segments, transform the concrete strategies of business scenarios into structured gene sequences, achieving a quantitative expression of recommendation strategies. This provides a feasible technical prerequisite for operations such as genetic crossover, mutation, and cross-population migration—a core breakthrough that traditional recommendation algorithms cannot achieve. Through a full-process processing of encoding verification, chromosome splicing, and fitness verification, it ensures a high degree of fit between the product portfolio and marketing strategy of each initial individual.

[0056] In some embodiments of the present invention, such as Figure 3 As shown, after step S123, the following steps are also included.

[0057] Step S124: Configure different initialization strategy distributions, genetic operation operator parameters, or local exploration targets for different strategy evolution subpopulations to form each strategy evolution subpopulation with a corresponding differentiated strategy search space.

[0058] Specifically, in this step, when configuring different initial strategy distributions for different strategy evolution subpopulations, probability distribution modeling and maximum likelihood estimation techniques can be used to configure initial strategy probability distributions (uniform distribution, normal distribution, skewed distribution) that are adapted to the needs of their customer groups for different subpopulations. For low-risk customer groups with concentrated needs (such as the middle-aged and elderly critical illness protection group), a normal distribution can be used to concentrate the initial strategies on the strategy interval corresponding to the core needs (the distribution mean is the core strategy feature vector, and the variance is less than or equal to 0.2), reducing ineffective exploration. For customer groups with diverse needs and high customization requirements (such as the family comprehensive protection group), a uniform distribution is used to ensure that the initial strategies uniformly cover the entire search interval (the distribution interval is [0,1], with no obvious mean bias), improving strategy diversity. The maximum likelihood estimation algorithm is used to solve the distribution parameters, and the distribution adaptability is verified by combining the historical strategy data of the customer group to ensure that the matching degree between the distribution and customer needs is greater than or equal to 0.8.

[0059] Specifically, in this step, when configuring the genetic operation operator parameters differently for different strategy evolutionary subpopulations, the search space entropy quantification technique can be used first to calculate the entropy value of the search space for each subpopulation's strategy. A higher entropy value indicates higher search space complexity. Then, based on the entropy results, differentiated crossover probability, mutation probability, and selection pressure parameters can be configured for different subpopulations: for subpopulations with low search space complexity (e.g., entropy less than 0.6), a high crossover probability (e.g., 0.75-0.8) and a low mutation probability (e.g., 0.0) can be set. (5-0.07) to accelerate the integration of high-quality strategies; for subpopulations with high search space complexity (such as entropy greater than or equal to 0.6), low crossover probability (such as 0.7-0.75) and high mutation probability (such as 0.08-0.1) can be set to improve the breadth of strategy exploration; at the same time, the selection pressure is increased for subpopulations corresponding to core needs (such as selecting the top 30% of high-quality individuals) to strengthen the evolutionary direction of core strategies, and the selection pressure is reduced for subpopulations with marginal needs (such as selecting the top 50% of high-quality individuals) to retain more potential strategies.

[0060] Specifically, in this step, when configuring different local exploration objectives for different strategy evolution subpopulations, the customer demand feature entropy weight ranking technique can be used to extract the core demand indicators of each customer segment (such as the cost-effectiveness of youth light-insurance groups and the ease of operation for community hospital groups), and the core demand indicators can be defined as the local exploration objectives of the corresponding subpopulations. Through objective quantification modeling, the local exploration objectives are transformed into calculable numerical objective thresholds (such as cost-effectiveness greater than or equal to 0.8, operation steps less than or equal to 3), and the objective thresholds are embedded into the search space constraints of the subpopulations. The local exploration boundary limitation technique is used to define the exploration range for each subpopulation, limiting the exploration weight of non-core objectives to less than or equal to 0.3, avoiding meaningless exploration by the subpopulations that deviate from the core needs, and ensuring the consistency between the local exploration objectives and the needs of the customer groups.

[0061] Further, a search space overlap calculation technique can be used. The similarity of the search space feature vectors of any two subpopulations after configuration can be calculated using cosine similarity. If the similarity is greater than or equal to 0.6, it is determined that the configurations are convergent. The distribution parameters / operator parameters / exploration targets of one of the subpopulations are immediately adjusted until the search space overlap between all subpopulations is less than 0.6. At the same time, a strategy adaptability score is used to calculate the adaptability of the initial strategy of each subpopulation to the needs of the corresponding customer group. If the adaptability score is greater than or equal to 0.7, the configuration is considered effective. Otherwise, the configuration parameters are re-optimized to ensure the effectiveness of differentiated configurations.

[0062] As a concrete example, in an insurance scenario, differentiated configurations can be applied to three strategy evolution subpopulations: critical illness protection for middle-aged and elderly individuals, lightweight protection for young adults, and comprehensive family protection. For instance, regarding the initial strategy distribution, the middle-aged and elderly subpopulation can be configured with a normal distribution (mean = critical illness protection and offline channel feature vectors, variance 0.18), the young adult subpopulation with a uniform distribution (covering the entire search range of lightweight products), and the family subpopulation with a skewed distribution (mean = family protection and online / offline combined feature vectors, variance 0.2). Regarding the genetic operator parameters, the middle-aged and elderly subpopulation (…) can be configured with… The search space entropy value is 0.55, with a crossover probability of 0.8 and a mutation probability of 0.06. The youth subpopulation (entropy value 0.72) is configured with a crossover probability of 0.7 and a mutation probability of 0.09, and the family subpopulation (entropy value 0.68) is configured with a crossover probability of 0.73 and a mutation probability of 0.08. For local exploration targets, the comprehensiveness of the middle-aged and elderly subpopulation can be set to be greater than or equal to 0.8, the cost-effectiveness of the youth subpopulation can be set to be greater than or equal to 0.85, and the family subpopulation can be set to be greater than or equal to 0.8. All of these are limited to a non-core target exploration weight of less than or equal to 0.2.

[0063] As another concrete example, in the medical technology scenario, differentiated configurations can be applied to four subgroups: high-end equipment in tertiary hospitals, basic equipment in community hospitals, portable equipment for home patients, and mid-range equipment in private institutions. Regarding the initialization strategy distribution, the tertiary hospital subgroup can be configured with a skewed distribution (mean: high-end equipment and customized after-sales feature vectors, variance: 0.19), the community subgroup with a normal distribution (mean: simple equipment and cost-effective feature vectors, variance: 0.17), the home subgroup with a normal distribution (mean: portable equipment and one-click operation feature vectors, variance: 0.18), and the private subgroup with a uniform distribution (covering the entire search range for mid-range equipment). Regarding the genetic operator parameters, the community / home subgroups (both with entropy values ​​less than 0.6) can be configured with crossover probabilities. The crossover probability is 0.78 and the mutation probability is 0.05. The crossover probability is 0.72 and the mutation probability is 0.08 for the three-family subpopulation (entropy value 0.70), and the crossover probability is 0.7 and the mutation probability is 0.1 for the private subpopulation (entropy value 0.75). For local exploration targets, the equipment performance of the three-family subpopulation can be set to be greater than or equal to 0.9, the ease of operation of the community subpopulation can be set to be greater than or equal to 0.85, the portability of the home subpopulation can be set to be greater than or equal to 0.85, the return on investment of the private subpopulation can be set to be greater than or equal to 0.8, and the exploration weight of non-core targets can be less than or equal to 0.3.

[0064] It is understandable that the above embodiments create exclusive optimization boundaries for subpopulations by initializing the strategy distribution, genetic operation operator parameters, and local exploration targets in a three-dimensional differentiated configuration, so that the iterative optimization of each subpopulation always revolves around its own core needs, thus avoiding wildcard strategies from the root.

[0065] Step S130: Construct a policy fitness function that integrates real-time external risk signals; Specifically, in this step, a combination of risk classification and fuzzy comprehensive evaluation techniques can be used to classify external risk signals into three main categories: industry, market, and customer. Each category contains several sub-risks (e.g., industry category includes policy adjustments and compliance requirements). Further, a fuzzy evaluation matrix can be constructed to determine the set of evaluation factors (risk types) and the set of comments (risk levels: high, medium, and low). The weight of each risk type is calculated using the analytic hierarchy process. A fuzzy transformation algorithm is used to map the risk feature vectors to the set of comments to obtain the risk level of each risk type. Then, through risk coefficient mapping (e.g., high risk = 0.8, medium risk = 0.4, low risk = 0.1), it is transformed into a calculable dynamic risk indicator for the strategy (risk indicator range [0,1]).

[0066] More specifically, when constructing the fitness function, the benefit indicators can be quantified first. Three core benefit indicators are selected: expected return of the recommendation strategy, execution efficiency, and customer matching degree. The Z-score standardization algorithm is used to map each benefit indicator to the [0,1] interval. The weights of each benefit indicator are then calculated using the entropy weight method (the sum of the weights is 1), and a linear weighting method is used to calculate the comprehensive benefit score (benefit score range [0,1]). Furthermore, when constructing the fitness function, function fusion can be performed. A nonlinear normalization fusion algorithm is used to integrate the comprehensive benefit score with the strategy's dynamic risk indicators to construct the fitness function, where the function formula is: Fitness value = Comprehensive benefit score. (1-(Various Risk Indicators) The sum of risk weights, where risk weights are the weights of each risk type, and the fitness value ranges from [0,1]. The higher the value, the better the strategy. A penalty mechanism is set up. When extreme risks occur (such as policy-prohibited risks), a step function is used to directly deduct the fitness value by 0.1-0.3 to ensure the compliance of the strategy.

[0067] More specifically, the fitness function can be dynamically adjusted. Specifically, a sliding window algorithm can be used to monitor changes in external risk signals in real time. When the change in risk indicators is greater than or equal to 0.2, the fitness function parameters can be automatically adjusted (risk weights and benefit weights can be recalculated). A function verification mechanism can also be set up to verify the prediction accuracy of the fitness function (accuracy greater than or equal to 85%) using historical recommendation data. If it does not meet the requirements, the fuzzy evaluation matrix parameters can be adjusted and the model can be rebuilt.

[0068] As a concrete example, in an insurance scenario, risk signals such as insurance regulatory policies, market interest rates, and customer health status can be collected. Further, a fuzzy comprehensive evaluation method can be used to quantify policy adjustments (weight 0.5), market interest rate fluctuations (weight 0.2), and customer performance risk (weight 0.3) into risk indicators, with a risk indicator value of 0.8 for high-risk situations. Further, the entropy weight method can be used to calculate the weights of the benefit indicators: premium conversion rate 0.4, repurchase rate 0.3, and matching degree 0.3, linearly weighted to obtain the comprehensive benefit score. The fitness function is: Fitness value = (Premium conversion rate...) / (Premium conversion rate...) 0.4+ repurchase rate 0.3+ match 0.3)×(1-(Policy Risk) 0.5+ market risk 0.2+ Customer Risk 0.3) When there is an adjustment to the scope of critical illness insurance coverage (extreme policy risk), deduct 0.2 fitness value; monitor risk signals through a sliding window and verify once every 30 minutes.

[0069] As another concrete example, in the medical technology scenario, risk signals such as centralized procurement policies, equipment registration certificates, and supply chain status can be collected. Furthermore, fuzzy comprehensive evaluation can be used to quantify centralized procurement policies (weight 0.6), supply chain fluctuations (weight 0.2), and customer operational risks (weight 0.2) into risk indicators. Further, the weights of benefit indicators can be calculated using the entropy weight method: transaction rate 0.45, renewal rate 0.35, and profit margin 0.2, linearly weighted to obtain a comprehensive benefit score. The fitness function can be: Fitness value = (Transaction rate...) 0.45+ renewal rate 0.35+ profit margin 0.2)×(1- (Centralized Procurement Risk) 0.6+ supply chain risk 0.2+ Customer Risk 0.2) When the equipment quality fails the spot check, deduct 0.3 fitness value; the sliding window is calibrated every 1 hour.

[0070] Step S140: Based on the policy fitness function, drive the multiple policy evolution subpopulations to perform iterative optimization in parallel, and perform policy knowledge transfer operations across the policy evolution subpopulations during the iteration process.

[0071] Specifically, in this step, when driving the multiple policy evolution subpopulations to perform iterative optimization in parallel based on the policy fitness function, parallel iterative optimization techniques and genetic operation techniques within the subpopulations can be used. Among them, the parallel iterative optimization technology can adopt a distributed scheduling framework to allocate multiple policy evolution subpopulations to different computing nodes. The time-slice round-robin scheduling algorithm is used to allocate equal computing resources (CPU, memory) to each computing node to ensure that the iteration progress of each subpopulation is synchronized (the difference in iteration rounds is less than or equal to 1). An iteration monitoring module can also be set up to collect the iteration progress and fitness average of each subpopulation in real time. When a node is stuck in iteration (progress lags behind by more than or equal to 3 rounds), resource scheduling is automatically triggered to increase the computing resources of that node.

[0072] Specifically, genetic manipulation techniques within a subpopulation can include fitness calculation, crossover, mutation, and generation of a new generation of populations.

[0073] Fitness calculation can be performed using a vector computing engine to batch calculate the fitness values ​​of all individuals in the subpopulation, with a computational efficiency of greater than or equal to 500 individuals / second. Furthermore, a sorting algorithm (quick sort) can be used to sort individuals from high to low fitness values, retaining the top 50% of high-fitness individuals as parent individuals.

[0074] Further crossover operations can be performed using a two-point crossover algorithm. Two individuals from the parent generation are randomly selected, and two crossover points (corresponding to different positions of gene segments) are randomly generated. The gene segments between the two individuals' crossover points are exchanged to generate offspring individuals. The crossover probability can be set to 0.7-0.8 (adjusted according to the diversity of the offspring population; 0.8 is used when the diversity is low, and 0.7 is used when the diversity is high) to avoid the loss of high-quality genes due to excessive crossover.

[0075] Further mutation operations can employ site mutation algorithms, randomly selecting gene loci in offspring individuals and changing the coding value of that locus (e.g., binary coding 0-1, real number coding 0.5-0.6) according to the mutation probability (0.05-0.1). For core gene segments (e.g., key sites in product portfolios), mutation protection mechanisms can be set up to ensure that high-quality core genes are not destroyed.

[0076] Further generation of the next generation population can employ a fusion technique of high-quality parent individuals and offspring individuals. The top 30% of high-quality parent individuals are merged with offspring individuals, and individuals with a similarity greater than or equal to 0.5 are removed, while offspring individuals with a similarity less than 0.5 to the parent individuals are retained, ensuring the diversity of the next generation population. The iteration termination condition can be that the mean change rate of fitness is less than 0.05 for three consecutive rounds or that a preset number of iteration rounds is reached (e.g., 50-100 rounds).

[0077] More specifically, when performing strategy knowledge transfer operations across strategy-evolving subpopulations, this can be achieved through strategy distribution distance calculation technology, high-quality gene segment extraction technology, and gene segment introduction and replacement technology. Among these, the strategy distribution distance calculation technology can use a dual measurement method of cosine similarity and Euclidean distance to calculate the strategy feature distribution distance of each subpopulation. All individual gene segments of each subpopulation are reduced to 10-dimensional feature vectors using principal component analysis algorithm. The cosine similarity (similarity greater than or equal to 0.7) and Euclidean distance (distance less than or equal to 0.3) of the feature vectors of the two subpopulations are calculated. If both conditions are met, the transfer is deemed possible.

[0078] Furthermore, by employing high-quality gene segment extraction technology and a feature importance ranking algorithm (random forest), the core gene segments (gene segments with feature importance greater than or equal to 0.8) of the top N fitness individuals in the source subpopulation (e.g., N=5-10, set according to the subpopulation size) can be extracted. Then, a gene segment cutting algorithm can be used to accurately cut the core gene segments, avoiding the inclusion of redundant gene fragments.

[0079] Furthermore, through gene segment introduction and replacement technology, high-quality gene segments extracted can be introduced into the target subpopulation using a gene segment insertion algorithm. A low-fitness individual replacement strategy can be adopted to remove the M-position (M=N) individuals with the lowest fitness in the target subpopulation. The introduced high-quality gene segments can be combined with the non-core gene segments of the individuals in the target subpopulation to generate new individuals, ensuring the stability of the number of individuals in the target subpopulation.

[0080] As a concrete example, three subpopulations can be assigned to three computing nodes, with time-slice round-robin scheduling to ensure synchronized iteration progress. Further, the iteration cycle can be preset to 80 rounds, with knowledge transfer performed every 8 rounds. During genetic operations within the subpopulations, the crossover probability can be set to 0.75, the mutation probability to 0.08, and the mutation probability of the core gene segment to 0.02. Fitness values ​​are calculated in batches, and after quick sorting, the top 50% of parent individuals are retained. Further, the PCA algorithm can be used to reduce the dimensionality of the subpopulation gene segments to 10 dimensions, calculating a cosine similarity of 0.75 and an equivalence similarity of 10 to 10 between the family subpopulation and the youth subpopulation. The distance between the two populations is 0.28, which meets the migration condition. Further, a random forest can be used to extract the core gene segments (feature importance greater than or equal to 0.8) of the top 8 high-quality individuals in the family subpopulation. These segments are then introduced into the young subpopulation using a gene segment insertion algorithm, and 8 low-fit individuals are removed. The introduced gene segments undergo a small mutation (mutation probability 0.08). Using Shannon entropy verification, the diversity entropy value of the young subpopulation is 0.65, which is acceptable. The distance between the middle-aged and older subpopulations and other subpopulations does not meet the condition, so the iteration proceeds independently. After 65 iterations, the mean fitness change rate is 0.04, which is less than 0.05, so the iteration is terminated early.

[0081] As another concrete example, in a medical technology scenario, four subpopulations can be assigned to four computing nodes, with a scheduling synchronization error of less than or equal to one round; the preset iteration rounds are 100 rounds, with knowledge transfer performed every 10 rounds; furthermore, when performing genetic operations within the subpopulations, the crossover probability can be set to 0.7, the mutation probability to 0.1, the mutation probability of the core gene segment to 0.02, and the fitness calculation efficiency to 550 individuals / second; after PCA dimensionality reduction, the private subpopulation and The cosine similarity of the Sanjia subpopulation was 0.72, the Euclidean distance was 0.29, and the Euclidean distance with the community subpopulation was 0.28, satisfying the migration conditions. The core gene segments of the top 10 high-quality individuals in the private subpopulation were extracted and introduced into the Sanjia and community subpopulations respectively. Ten individuals with low fitness in each subpopulation were removed, resulting in a small mutation (mutation probability 0.09). Shannon entropy verification showed that the diversity entropy values ​​of the two target subpopulations were 0.63 and 0.64 respectively, which were acceptable. The home subpopulation was iterated independently until 100 rounds, at which point the mean fitness value tended to stabilize.

[0082] In some embodiments of the present invention, the policy fitness function is constructed as follows: The value of the strategy fitness function increases in response to the expected return index of the recommended strategy, and decreases in response to the dynamic risk index of the strategy calculated based on the recommended strategy.

[0083] Specifically, in this embodiment, a core benefit indicator screening technique can be first used. From the total benefit indicators of the recommended strategy, the expected benefit core indicators that are highly correlated with the actual implementation value of the strategy can be screened using the Pearson correlation coefficient (the correlation coefficient can be set to be greater than or equal to 0.4). For example, in the insurance scenario, premium conversion rate, customer repurchase rate, and product portfolio profit margin can be selected; in the medical technology scenario, equipment procurement transaction rate, customer renewal rate, and ancillary service profit margin can be selected. Further, the Z-score standardization algorithm can be used to map each core benefit indicator to the [0,1] interval to eliminate dimensional differences. The weight of each indicator is calculated using the entropy weight method, and a comprehensive benefit scoring model is constructed. The comprehensive benefit score = (standardized indicator value) / (standardized indicator value) ... The sum of indicator weights; ultimately, a linear positive mapping technique can be used to establish a positive correlation between the overall return score and the fitness function value, so that for every 0.1 increase in the overall return score, the function value increases by 0.1 (mapping coefficient = 1), ensuring that the higher the return, the larger the function value.

[0084] More specifically, a strategy risk chain decomposition technique can be used to break down the risk points throughout the entire chain from the implementation of the recommended strategy to its execution, and to screen dynamic risk core indicators that are strongly correlated with the strategy's execution effect (such as policy risk coefficient, market risk coefficient, and customer performance risk coefficient in insurance scenarios, and centralized procurement policy risk coefficient, supply chain risk coefficient, and customer operation risk coefficient in medical technology scenarios). Using fuzzy comprehensive evaluation, each core risk indicator can be quantified into a risk coefficient in the range [0,1], with a higher risk coefficient indicating a greater risk in strategy execution. Furthermore, a comprehensive risk scoring model can be constructed, such that the comprehensive risk score = (risk coefficient) / (risk coefficient). The sum of risk weights, where risk weights are determined by the analytic hierarchy process and pass the consistency test; further, a linear negative mapping technique can be used to establish a negative correlation between the comprehensive risk score and the fitness function value. For every 0.1 increase in the comprehensive risk score, the function value decreases by 0.1 (mapping coefficient = -1), ensuring that the higher the risk, the smaller the function value.

[0085] More specifically, a weighted fusion modeling technique can be used to construct the core calculation model for the fitness function value: Fitness function value = Overall return score (1-Risk Comprehensive Score) enables a two-way linkage between returns and risks, allowing the function value to simultaneously reflect the strategy's profitability and risk control capabilities; a return guarantee mechanism is set, for example, if the strategy's return comprehensive score is less than 0.3, even if the risk comprehensive score is 0, the function value is still limited to less than 0.3, preventing ineffective strategies with low returns and no risks from obtaining high scores; a risk circuit breaker mechanism is set, if the coefficient of a certain core risk indicator is greater than or equal to 0.8 (extreme risk), the risk comprehensive score is directly set to 1 and the function value is set to 0, completely eliminating high-risk strategies.

[0086] As a concrete example, in an insurance scenario, the core expected return indicators for a recommended strategy for a certain middle-aged and elderly critical illness insurance group are premium conversion rate (0.7), customer repurchase rate (0.6), and product portfolio profit margin (0.8). Calculated using the entropy weight method, their weights are 0.4, 0.3, and 0.3 respectively, resulting in a comprehensive return score of 0.7. 0.4 + 0.6 0.3 + 0.8 0.3 = 0.7; The core dynamic risk indicators are policy risk coefficient (0.2), market risk coefficient (0.1), and customer performance risk coefficient (0.1), with weights of 0.5, 0.2, and 0.3 respectively calculated by AHP, resulting in a comprehensive risk score of 0.2. 0.5 + 0.1 0.2 + 0.1 0.3 = 0.15; fitness function value = 0.7 (1-0.15)=0.595. If the premium conversion rate of this strategy increases to 0.8, the overall return score increases to 0.77, and the function value increases accordingly to 0.77×0.85=0.6545 (the increase in return leads to an increase in the function value); if the policy risk coefficient increases to 0.5, the overall risk score increases to 0.32, and the function value decreases to 0.7. (1-0.32)=0.476 (Increased risk leads to a decrease in function value), which perfectly matches the construction rule that "increased returns lead to increased function value, and increased risk leads to decreased function value".

[0087] As another concrete example, in the medical technology scenario, a recommendation strategy for mid-range equipment from a private institution has the following expected return key indicators: equipment purchase conversion rate (0.75), customer renewal rate (0.7), and ancillary service profit margin (0.65), with entropy weights of 0.45, 0.35, and 0.2 respectively, resulting in a comprehensive return score of 0.75. 0.45 + 0.7 0.35 + 0.65 0.2 = 0.7175; The core dynamic risk indicators are the centralized procurement policy risk coefficient (0.1), supply chain risk coefficient (0.05), and customer operation risk coefficient (0.1), with AHP weights of 0.6, 0.2, and 0.2 respectively, resulting in a comprehensive risk score of 0.1. 0.6 + 0.05 0.2 + 0.1 0.2 = 0.09; fitness function value = 0.7175 (1-0.09)=0.6529. If the success rate of this strategy increases to 0.85, the overall return score rises to 0.79, and the function value rises to 0.79 accordingly. 0.91 = 0.7189; if the risk coefficient of the centralized procurement policy rises to 0.4, the comprehensive risk score rises to 0.27, and the function value drops to 0.7175. (1-0.27)=0.5238, accurately realizing the two-way linkage control of returns and risks on the function value.

[0088] Understandably, the above embodiments achieve real-time linkage between strategy evaluation and business scenario risks. The function value is positively correlated with the return indicator and negatively correlated with the risk indicator, ensuring that the individual evaluation in each iteration is consistent with the real-time state of the business and preventing the retention of strategies with high returns but high violation risks. The above embodiments have built-in risk circuit breaker and return guarantee mechanisms, which directly set the fitness of extreme risk strategies to zero and directly limit the score of low return strategies. From the evaluation level, high-risk and valueless strategies in business scenarios are completely eliminated, so that the final output recommended strategies have no high-risk items.

[0089] In some embodiments of the present invention, such as Figure 4 As shown, step S140 includes the following steps.

[0090] Step S1411: Based on the policy fitness function, drive the multiple policy evolution subpopulations to perform iterative optimization in parallel.

[0091] Specifically, in this step, a distributed scheduling framework can be used to construct a master-slave node architecture. The master node is responsible for global parameter management, iteration progress monitoring, and resource allocation, while the slave nodes correspond to each strategy evolution subpopulation and independently execute iterative optimization operations. A time-slice round-robin scheduling algorithm is used to allocate equal computing resources to each slave node. The time slice length can be set to 10-30 seconds (this time slice length can be adjusted according to the subpopulation size, such as 30 seconds when the number of individuals is greater than 100), ensuring that the iteration progress of each subpopulation is synchronized. The difference in iteration rounds can be less than or equal to one round, avoiding inefficiency in global optimization caused by the lag of a certain subpopulation.

[0092] More specifically, an iteration progress monitoring module can be added to collect the iteration rounds, average population fitness, and computing resource utilization of each slave node in real time. A threshold alarm mechanism is adopted. When a slave node is stuck in iteration (progress lags by more than or equal to 3 rounds) or its resource utilization is greater than or equal to 85%, the master node automatically triggers dynamic resource scheduling to allocate the computing resources of idle nodes to that slave node, ensuring the continuity of iteration.

[0093] More specifically, the policy fitness function constructed earlier can be called, and a vector computation engine can be used to perform batch fitness calculations on all individuals in the current subpopulation. The computation efficiency can be set to greater than or equal to 500 individuals / second to avoid the inefficiency caused by calculating one individual at a time. After the calculation is completed, a quicksort algorithm can be used to sort the individuals from high to low fitness values, retain the top 50% of high-fitness individuals as parent individuals, and remove the bottom 50% of low-fitness individuals to reduce invalid iterations.

[0094] More specifically, a combination of roulette wheel selection and elite retention strategy can be adopted. Roulette wheel selection allocates selection probability according to the fitness ratio of individuals. The higher the fitness, the greater the probability of being selected as the parent generation for reproduction. At the same time, the top 10% of the best individuals (elite individuals) are retained and directly enter the next generation population to avoid the loss of high-quality genes.

[0095] More specifically, based on the configured differentiated crossover probability, a two-point crossover algorithm can be used to randomly select two parent individuals, generate two non-repeating crossover points (corresponding to the gene segment positions on the chromosome), exchange the gene segments between the two individuals' crossover points, and generate offspring individuals; after crossover, an encoding verification algorithm is used to remove offspring individuals with disordered gene segments or those that do not meet the constraints, to ensure the effectiveness of crossover.

[0096] More specifically, based on the configured differential mutation probability, a site mutation algorithm can be used to randomly select non-core gene sites in offspring individuals (the mutation probability of core gene sites is reduced to 0.02), and change the encoding value of the site (binary 0-1 / 1-0, real number 0.5-0.6, etc.); after mutation, a fitness check is performed to ensure that the mutated individuals are still suitable for the needs of the corresponding customer group (fitness score greater than or equal to 0.6).

[0097] Finally, elite individuals and selected offspring individuals can be merged to form a new generation of population, with the population size remaining consistent with the initial size (50-200 individuals / subpopulation). A convergence judgment algorithm is used to calculate the rate of change of the average fitness of the population over three consecutive rounds. If the rate of change is less than 0.05, the population is considered to have converged, and the iteration of the subpopulation is terminated. If it has not converged, the above operation is repeated until the preset number of iterations (50-100 rounds) is reached or the convergence condition is met.

[0098] Step S1412: For the source subpopulation and target subpopulation in the multiple strategy evolution subpopulations, when the strategy distribution distance between the source subpopulation and the target subpopulation is less than a threshold, the individual corresponding to the recommended strategy whose fitness ranking is in the top N positions in the source subpopulation is introduced into the target subpopulation.

[0099] Specifically, in this step, a subpopulation clustering algorithm can be used to cluster all policy evolution subpopulations according to policy characteristics. Subpopulations in the same cluster are candidate source subpopulations and target subpopulations for each other (i.e., subpopulations with high demand similarity can migrate between each other); subpopulations in a single cluster (without suitable migration objects) are excluded to avoid blind migration.

[0100] More specifically, Principal Component Analysis (PCA) can be used to reduce the dimensionality of all individual chromosomes (gene segments) in each subpopulation to a 10-dimensional feature vector (retaining more than 90% of the feature information), eliminating high-dimensional data redundancy. Then, the cosine similarity and Euclidean distance between the feature vectors of the source and target subpopulations are calculated to double-verify the differences in strategy distribution. If the cosine similarity is greater than or equal to 0.7 (indicating high similarity in strategy distribution) and the Euclidean distance is less than or equal to 0.3 (indicating small differences in strategy distribution), the distance is determined to be less than the threshold, satisfying the migration condition. If either indicator is not satisfied, the migration condition is determined not to be met, and the migration operation is not performed.

[0101] More specifically, the threshold can be determined by verifying historical migration data to ensure that subpopulations that meet the threshold conditions have a strategy fit score greater than or equal to 0.7 after migration.

[0102] More specifically, after determining the source subpopulation, the top N individuals in fitness ranking within that subpopulation can be selected (N is set according to the subpopulation size; N=5 when the number of individuals is 50-100, and N=10 when the number of individuals is 100-200). Using the random forest algorithm, the core gene segments of these high-quality individuals (such as gene segments with feature importance greater than or equal to 0.8, such as core configurations of product portfolios and key parameters of marketing strategies) are extracted, redundant gene segments are removed, reducing the amount of migration data and improving migration efficiency.

[0103] More specifically, a combination of gene segment insertion and replacement techniques can be used. The extracted high-quality core gene segments are replaced with the non-core gene segments of the low-fitness individuals ranked Nth in the fitness ranking of the target subpopulation, generating new individuals to replace the original low-fitness individuals. After replacement, the similarity between the new individuals and the original individuals in the target subpopulation is calculated using cosine similarity. For example, if the similarity is greater than or equal to 0.5, a small mutation (mutation probability 0.05) can be performed to avoid homogenization of the subpopulation.

[0104] More specifically, after the migration is completed, the mean fitness of the target subpopulation can be recalculated to verify the migration effect. If the mean fitness increases by more than or equal to 0.05, the migration is considered effective; if it does not increase, the migration operation is canceled, the N value is adjusted, or the source subpopulation is re-selected.

[0105] As a concrete example, in an insurance scenario, the Spark Streaming framework can be used. A master node manages three slave nodes, corresponding to three sub-populations: critical illness protection for middle-aged and elderly individuals, lightweight protection for young adults, and comprehensive family protection. Each sub-population has a 20-second time slice and equal resource allocation. Each sub-population iterates independently. The middle-aged and elderly sub-population (initially 80 individuals, crossover probability 0.8, mutation probability 0.06): fitness is calculated in batches using NumPy, and after quicksort, 40 parent individuals are retained. A roulette wheel selection and elite retention (8 elite individuals) are used, with crossover at two points generating offspring. After mutation at a point, a new generation of individuals is formed. Iteration terminates at round 65 when the mean fitness change rate (0.04) is less than 0.05 for three consecutive rounds. The young adult sub-population (initially 60 individuals, crossover probability 0.7, mutation probability 0.09) and the family sub-population (initially 100 individuals, crossover probability 0.73, mutation probability 0.08) iterate synchronously, reaching convergence at rounds 70 and 75, respectively.

[0106] Furthermore, migration can be performed every 8 iterations. K-means clustering shows that the family subpopulation and the youth subpopulation are in the same cluster. After PCA dimensionality reduction, the cosine similarity between the family subpopulation and the youth subpopulation is 0.75, and the Euclidean distance is 0.28, which meets the migration conditions. Further, the top 8 fittest individuals from the family subpopulation (source subpopulation) can be extracted, and core gene segments (family insurance product combination, online and offline channel parameters) can be extracted using random forest. The bottom 8 fittest individuals from the youth subpopulation (target subpopulation) have their gene segments replaced, generating 8 new individuals. After a small mutation (mutation probability of 0.05), these new individuals are included in the youth subpopulation. After migration, the mean fitness of the youth subpopulation increases from 0.72 to 0.78, indicating effective migration. The distance between the middle-aged and elderly subpopulation and the other two subpopulations does not meet the threshold, so migration was not performed.

[0107] As another concrete example, in the medical technology scenario, the Flink framework can be used, with 4 slave nodes corresponding to 4 subpopulations, and the time slice is set to 30 seconds; the three subpopulations (initially 120 individuals, crossover probability 0.72, mutation probability 0.08) and the community subpopulation (initially 90 individuals, crossover probability 0.78, mutation probability 0.05) are iterated independently, and convergence is achieved in 80 rounds and 68 rounds respectively, with the average fitness value being greater than or equal to 0.75. Furthermore, the migration can be performed every 10 iterations. K-means clustering shows that the private subpopulation is in the same cluster as the top three and community subpopulations. After PCA dimensionality reduction, the cosine similarity between the private subpopulation and the top three subpopulation is 0.72 and the Euclidean distance is 0.29; the Euclidean distance between the private subpopulation and the community subpopulation is 0.28 and the cosine similarity is 0.73, both of which meet the migration conditions. Furthermore, the core gene segments (bulk purchase discount, maintenance cost control parameters) of the top 10 fittest individuals in the private subpopulation can be extracted and used to replace the non-core gene segments of the bottom 10 low-fit individuals in the top three and community subpopulations, respectively, and then incorporated after slight mutation. After migration, the mean fitness of the two target subpopulations increased by 0.06 and 0.07, respectively, indicating that the migration was effective.

[0108] It is understandable that the above embodiments implement parallel iteration of multiple subpopulations based on a distributed scheduling framework, and use time-slice round-robin scheduling to synchronize the iteration progress of different subpopulations, thus solving the problem of low efficiency in traditional serial iteration. Furthermore, the above embodiments construct a precise migration determination mechanism based on strategy distribution distance. By performing migration only on populations with similar needs, the decrease in strategy adaptability caused by blind migration is avoided. It can be understood that cross-population migration injects high-quality core genes into the target subpopulation. After migration, through minor mutation and deduplication, the core needs of lightweight and cost-effective young subpopulations are preserved, achieving the dual preservation of advantages and unique characteristics.

[0109] In some embodiments of the present invention, driving the plurality of policy evolution subpopulations to perform iterative optimization in parallel based on the policy fitness function includes: Step S1421, cyclically execute within each subpopulation: calculate the fitness of each individual within the subpopulation based on the policy fitness function; Specifically, in this step, the constructed policy fitness function that integrates real-time external risk signals is invoked. The NumPy vector computing engine is used to perform batch fitness calculations on all individuals within the current subpopulation. The computational efficiency can be set to greater than or equal to 500 individuals / second to avoid the efficiency bottleneck of calculating each individual and to suit the 50-200 individual size of each subpopulation. Furthermore, the latest external risk factors (such as policy risk coefficients in insurance scenarios and centralized procurement risk coefficients in medical technology scenarios) can be synchronized from the streaming risk signal database before calculation to update the comprehensive risk score of the fitness function, ensuring that the fitness value reflects the real-time payoff-risk status without lag. Further, the mean fitness of the current subpopulation can be calculated in real-time and recorded in the iteration log for subsequent convergence assessment. The log also records key information such as risk factors and calculation time for each round, ensuring traceability of the trajectory.

[0110] Step S1422: Select target individuals whose fitness value is higher than a preset value; Specifically, in this step, the preset value can be set as a stage-by-stage screening threshold during the iteration process, such as uniformly setting it to 0.6 (lower than the final strategy output threshold of 0.7, leaving enough room for optimization during iteration). This value is a hard screening standard, directly eliminating individuals with fitness values ​​less than 0.6 to reduce invalid genetic operations.

[0111] More specifically, for individuals with fitness scores greater than or equal to 0.6, a second screening is conducted using a customer group fit score (greater than or equal to 0.6) to remove individuals with high fitness scores but who do not match the needs of the corresponding customer group (such as strategies with high fitness scores but excessively high premiums in the youth sub-population), ensuring that the target individuals meet both the quality and fit criteria.

[0112] More specifically, an elite retention strategy can be adopted, in which the top 10% of the individuals in the current subpopulation with the best fitness are directly identified as core target individuals. They do not need to participate in secondary selection and can directly enter the subsequent genetic operations, thus avoiding the accidental deletion of high-quality core genes due to selection rules. Elite individuals account for 20%-30% of the total number of target individuals.

[0113] More specifically, the number of target individuals after screening can be controlled to 50% of the current subpopulation size (e.g., retaining 40 target individuals in a subpopulation of 80 individuals), which ensures the efficiency of genetic operations while avoiding the dilution of high-quality genes caused by too many target individuals.

[0114] Step S1423: Perform crossover and mutation on the target individual to generate a new individual.

[0115] Specifically, in this step, the crossover operation can employ a two-point crossover algorithm and differentiated crossover probabilities. More specifically, the selected target individuals can be randomly paired (elite individuals are mixed with ordinary target individuals to increase the probability of high-quality gene transmission), and two-point crossover can be performed based on differentiated crossover probabilities, such as for middle-aged / elderly / community / home-based subpopulations (crossover probability 0.78-0.8) and youth / family / top-tier / private sector subpopulations (0.7-0.73). Further, two non-repeating crossover points can be randomly generated (corresponding to gene segment positions on chromosomes, such as the separation point between product portfolio gene segments and marketing strategy gene segments, or non-core regions outside core gene segments), and the gene segments between the paired individuals' crossover points can be exchanged to generate the first-generation offspring. Finally, after crossover, a coding verification algorithm is used to remove first-generation offspring individuals with disordered gene segments or those violating hard constraints of the customer group (such as monthly premiums exceeding 500 yuan for the youth subpopulation), ensuring the effectiveness of the crossover.

[0116] Specifically, in this step, the mutation operation can employ site mutation algorithms and core gene protection. More specifically, site mutation can be performed on the verified first-generation offspring individuals based on differentiated mutation probabilities, such as youth / private sector offspring populations (mutation probability 0.09-0.1), middle-aged / elderly / family / 30-year-old offspring populations (0.06-0.08), and community / home-based offspring populations (0.05).

[0117] More specifically, the mutation probability of an individual's core gene segments (such as core configurations of the product portfolio and key parameters of marketing strategies, with a feature importance greater than or equal to 0.8) can be reduced to 0.02, while non-core gene segments mutate according to their original probability, thus avoiding the destruction of high-quality core genes.

[0118] Step S1424: The new individual is merged with the target individual of the previous generation to form a new generation of subpopulation.

[0119] Specifically, in this step, the previous generation of core target individuals (the top 10% of elite individuals) and the verified new individuals can be merged. The elite individuals are directly included without participating in any deduplication operations, ensuring that 100% of the high-quality genes are preserved.

[0120] More specifically, the cosine similarity algorithm can be used to calculate the similarity between non-elite new individuals after merging, and duplicate individuals with a similarity greater than or equal to 0.5 can be removed to avoid homogenization of the new generation population strategy and ensure population diversity.

[0121] More specifically, if the number of individuals after deduplication is lower than the initial size of the subpopulation (e.g., 80 / 60 / 100), a small number of new individuals that meet the customer group constraints can be added using a random generation method (accounting for less than or equal to 5%) to ensure that the size of each subpopulation remains constant during the iteration process and does not affect the iteration efficiency.

[0122] More specifically, after completion, a new generation of subpopulation is formed, which serves as the base population for the next iteration. At the same time, the mean fitness of the new generation of subpopulation is calculated and compared with the mean of the previous two rounds. If the mean change rate is less than 0.05 for three consecutive rounds, the subpopulation is considered to have converged, and the iteration is terminated. If it has not converged, the closed-loop process of S1421-S1424 is repeated until the convergence condition or the preset iteration round is reached.

[0123] As a concrete example, in the insurance scenario, the three sub-populations of the insurance scenario independently execute the above closed-loop iteration, adapting differentiated parameters, a preset value of 0.6, and a convergence condition where the average change rate of the three consecutive rounds is less than 0.05. The iteration parameters and execution results of each sub-population are completely consistent with the settings mentioned above, ultimately providing a high-quality population foundation for the subsequent output of 45 recommendation strategies. The specific execution process can be implemented in the middle-aged and elderly critical illness protection sub-population (initial 80 individuals, crossover 0.8, variation 0.06, preset 80 iterations). Specifically, in each round, NumPy is used to batch calculate the fitness of 80 individuals, synchronizing with risk factors such as CBIRC policies and customer health status, and logging the average value for each round. Individuals with fitness greater than or equal to 0.6 are selected, and 40 target individuals are retained, of which 8 elite individuals are directly retained. The 40 target individuals are randomly paired, and two-point crossover is performed with a crossover probability of 0.8, and site mutation is performed with a mutation probability of 0.06. The mutation probability of the core gene (ensuring comprehensiveness) is 0.02, generating 40 new individuals. Three non-compliant individuals are verified and removed. The 8 elite individuals and 37 new individuals are merged, and two duplicate individuals are removed. Three random individuals are added to restore the scale to 80. In the 65th round, the average fitness change rate of the three consecutive rounds is less than 0.05 (0.04), indicating early convergence. After convergence, the average fitness is 0.76, and there are 22 individuals with fitness greater than or equal to 0.7, providing candidates for subsequent strategy output. The youth light-weight support subpopulation (initial 60 individuals, crossover probability 0.7, mutation probability 0.09, preset 80 iterations) executes the closed-loop process described above, with a crossover probability of 0.7, a mutation probability of 0.09, and a core gene (cost-effectiveness) mutation probability of 0.02. The iteration result can converge at the 70th round, with a mean of 0.78 after convergence, and 18 individuals with a fitness greater than or equal to 0.7, providing candidates for subsequent strategy output. The family comprehensive support subpopulation (initial 100 individuals, crossover probability 0.73, mutation probability 0.08, preset 80 iterations) executes the closed-loop process described above, with a crossover probability of 0.73, a mutation probability of 0.08, and a core gene (family fit) mutation probability of 0.02. The iteration result can converge at the 75th round, with a mean of 0.77 after convergence, and 30 individuals with a fitness greater than or equal to 0.7, providing candidates for subsequent strategy output. After convergence, a total of 22+18+30=70 candidate strategies with fitness greater than or equal to 0.7 were selected. After subsequent S151 proportional screening (top 20%), feasibility verification, and deduplication, 45 high-quality recommendation strategies were finally output, which is completely consistent with the previous settings.

[0124] As another concrete example, in the medical technology scenario, the four sub-populations also independently executed closed-loop iterations, adapting differentiated parameters, a preset value of 0.6, and a convergence condition where the mean change rate was less than 0.05 for three consecutive rounds. The iteration parameters and execution results of each sub-population were consistent with the settings mentioned above, ultimately providing a high-quality population foundation for the subsequent output of 67 recommendation strategies. The specific core iteration results are as follows: The high-end equipment sub-population of tertiary hospitals (120 items, crossover 0.72, mutation 0.08): converged in the 80th round, with a mean of 0.79, greater than or equal to 0.7. 35 individuals were selected; the community hospital basic equipment sub-population (90 individuals, crossover 0.78, variation 0.05) converged in the 68th round, with a mean of 0.77 and 28 individuals greater than or equal to 0.7; the home patient portable equipment sub-population (70 individuals, crossover 0.78, variation 0.05) converged in the 65th round, with a mean of 0.76 and 20 individuals greater than or equal to 0.7; the private institution mid-range equipment sub-population (80 individuals, crossover 0.7, variation 0.1) converged in the 72nd round, with a mean of 0.80 and 25 individuals greater than or equal to 0.7. After convergence, a total of 35+28+20+25=108 candidate strategies with fitness greater than or equal to 0.7 were selected. After subsequent S151 proportional screening (top 20%), feasibility verification, and deduplication, 67 high-quality recommended strategies were finally output, which is completely consistent with the previous settings.

[0125] It is understandable that the above embodiments achieve the complete preservation of elite genes. By using the elite preservation strategy, individuals with the highest fitness in each subpopulation are directly included in the next generation, avoiding the loss of high-quality strategy genes, ensuring the positive direction of iterative optimization, and making the average fitness of each subpopulation high after convergence.

[0126] Step S150: Output a recommended strategy set from the optimized strategy evolution subpopulation, and generate explanatory information related to the strategy evolution trajectory of the recommended strategy set.

[0127] Specifically, in this step, when outputting the recommended strategy set from the optimized strategy evolution subpopulation, a preliminary screening can be performed using a threshold screening method. The fitness threshold is set to 0.7 (determined based on industry standards and historical data). Individuals with a fitness value greater than or equal to 0.7 in each subpopulation are extracted as candidate high-quality strategies. Further, a ranking algorithm is used to sort the candidate strategies from high to low fitness values, retaining the top 20% of strategies as the preliminary screening results.

[0128] Furthermore, a feasibility verification model can be constructed, setting verification indicators from three dimensions: compliance, resource adaptability, and execution difficulty (e.g., compliance: whether it conforms to industry policies; resource adaptability: whether it matches the product library and marketing resources). A fuzzy comprehensive evaluation method is used to calculate the feasibility score of candidate strategies (greater than or equal to 0.8 is considered feasible), and strategies with a feasibility score less than 0.8 are eliminated. An association rule mining algorithm is used to eliminate duplicate strategies with a similarity greater than or equal to 0.7 between strategies, ensuring the diversity and feasibility of the recommended strategy set.

[0129] Furthermore, classification algorithms (such as K-means) can be used to categorize and archive the selected high-quality strategies according to customer segments, generating sub-strategy sets. Each sub-strategy set contains structured information such as strategy identifier, fitness value, suitable customer group, and core advantages, which are stored in a relational database and support fast query and retrieval.

[0130] Furthermore, in this step, the strategy evolution trajectory tracing technology can be used to obtain the strategy evolution trajectory of the recommended strategy set. The strategy evolution trajectory tracing technology can include trajectory log collection and trajectory tracing algorithm. The trajectory log collection can adopt a log collection framework to record the core information of each individual in each iteration in real time, including fitness value, gene segment crossover and mutation, knowledge transfer records, external risk signal influence, parent / offspring relationship, to ensure the integrity and traceability of the log.

[0131] The trajectory tracing algorithm can employ a reverse tracing algorithm, starting from the final optimal strategy and tracing back to the initial individual layer by layer through parent identifiers, extracting key nodes in each iteration (such as the rounds where fitness is greatly improved, the location of gene segment crossover mutations, and the source of knowledge transfer); and using a time-series analysis algorithm to sort out the temporal order and causal relationship of key nodes to form a complete evolutionary trajectory.

[0132] Furthermore, in this step, when generating explanatory information related to the evolutionary trajectory of the recommended strategy set, template generation technology can be used to pre-set an explanatory information template. This template can include five modules: core strategy advantages, key nodes of the evolutionary trajectory, optimization logic, risk prevention and control points, and execution guidelines. Natural language processing algorithms are used to convert the structured data in the evolutionary trajectory (such as gene segment encoding and fitness value changes) into natural language descriptions (such as in the 15th iteration, the fitness value increased from 0.68 to 0.75 due to the introduction of high-quality gene segments from the family subpopulation).

[0133] As a concrete example in the insurance scenario, a fitness threshold of 0.7 can be set to extract 52 candidate high-quality strategies from three subpopulations. After sorting, 10 strategies (top 20%) are retained. A feasibility verification model is then constructed, using fuzzy comprehensive evaluation to eliminate three strategies with feasibility scores less than 0.8 based on compliance, resource adaptability, and execution difficulty. Association rule mining is then used to eliminate two duplicate strategies, ultimately generating a recommended strategy set of 45 strategies, which are then archived in a database according to customer groups. Iteration logs for each individual are collected, containing information such as fitness values ​​and gene mutation locations. A reverse tracing algorithm is used to trace back from the optimal strategy of the family subpopulation to the initial individual, extracting eight key nodes. Finally, an NLP algorithm is used to convert the structured logs into natural language descriptions.

[0134] As another concrete example, in the medical technology scenario, a fitness threshold of 0.7 can be set to extract 78 candidate high-quality strategies from 4 subpopulations. After sorting, 16 strategies (top 20%) are retained. The feasibility verification model eliminates 5 infeasible strategies and 2 duplicate strategies, ultimately generating a recommended strategy set of 67 strategies. Iteration logs are collected, and a reverse tracing algorithm is used to trace the evolutionary trajectory of each high-quality strategy, extracting 10 key nodes. An NLP algorithm is used to convert the data into natural language to generate a gene segment variation diagram, generating an explanatory report with visual charts to clarify the core advantages and risk control points of the strategy, facilitating its implementation.

[0135] In some embodiments of the present invention, step S150 includes the following steps.

[0136] Step S151: Output a recommended strategy set from the optimized strategy evolution subpopulation.

[0137] Specifically, in this step, for each policy evolution subpopulation that has completed iterative optimization (convergence or reaching a preset number of rounds), a dual mechanism of threshold screening and proportional screening can be adopted. Threshold screening can set a fitness threshold (determined based on industry standards and historical data, uniformly set to 0.7) to extract individuals in the subpopulation with a fitness value greater than or equal to 0.7 as candidate high-quality policies. Proportional screening can sort the candidate high-quality policies from high to low fitness values, retaining the top 20% of policies (to avoid an excessive number of policies and improve practicality). If the fitness value of the top 20% of policies is less than 0.75, the proportion can be appropriately increased to 30% to ensure quality.

[0138] Furthermore, during the screening process, a vector deduplication algorithm can be used to calculate the similarity between candidate strategies, eliminate duplicate strategies with a similarity greater than or equal to a preset value (such as 0.7), and retain strategy diversity.

[0139] More specifically, a classification algorithm (K-means) can be used to classify and archive all high-quality strategies that have passed the verification according to their corresponding customer segments, forming a hierarchical structure of sub-strategy sets and total strategy sets: each customer segment corresponds to a sub-strategy set, and all sub-strategies are combined to form a global recommendation strategy set; furthermore, the strategy set can be standardized, and each strategy can be assigned a unique identifier (format: sub-segment identifier and strategy sequence number, such as "middle-aged and elderly critical illness-01"), and the core parameters of the strategy can be marked (product mix, marketing strategy, fitness value, and suitable customer characteristics).

[0140] Step S152: Trace the subpopulation affiliation changes and migration events that each recommendation strategy in the recommendation strategy set has experienced during the evolution process.

[0141] Specifically, during the iterative optimization process, a log collection framework can be used to collect the full lifecycle information of each individual (strategy) in real time. Each log entry can include a unique identifier for the individual, the identifier of its subpopulation, the fitness value of each iteration, gene segment crossover / mutation records, cross-subpopulation migration records (migration round, source subpopulation, target subpopulation, gene segment migration content), and subpopulation affiliation change records (if any). Furthermore, the logs can be stored in a distributed database with partitioned storage (partitioned by subpopulation and iteration round), supporting fast backtracking queries. The log retention time can be set to greater than or equal to one year to ensure traceability.

[0142] More specifically, for each strategy (individual) in the recommended strategy set, a reverse tracing algorithm can be used. Starting from the final individual identifier of the strategy, the algorithm traces back to the initial individual layer by layer through the parent individual identifiers in the log, forming a complete evolutionary trajectory chain. During the tracing process, a temporal sorting algorithm can be used to sort the data by iteration rounds, extract key nodes in the trajectory, and focus on screening nodes that change subpopulation affiliation, nodes that migrate across subpopulations, nodes that significantly increase fitness values, and nodes with major gene segment mutations.

[0143] More specifically, an event extraction algorithm (based on regular expressions) can be used to extract two types of key events from the traced trajectory chain, forming a structured event report. These two types of key events can include subpopulation affiliation change events and cross-subpopulation migration events. For subpopulation affiliation change events, the following can be extracted: change round, original subpopulation, new subpopulation, reason for change (e.g., migration, subpopulation merging), changes in gene segments after the change, and changes in fitness values. For cross-subpopulation migration events, the following can be extracted: migration round, source subpopulation identifier, target subpopulation identifier, core gene segments of the migration, changes in fitness values ​​before and after the migration, and migration effect (whether it was effective).

[0144] More specifically, an association mapping algorithm can be used to bind the structured event report and complete evolution trajectory of each recommendation strategy to the strategy itself (associated through the strategy's unique identifier), embedding it into the interpretation module of the recommendation strategy set, forming an integrated output of strategy, trajectory, and event, which is convenient for business personnel to query and understand.

[0145] As a concrete example, in the insurance scenario, 52 candidate strategies with a fitness greater than or equal to 0.7 can be extracted from the three optimized sub-populations. The top 20% are then selected to obtain 10 preliminary high-quality strategies. After vector deduplication, two duplicate strategies are removed, leaving 8. A fuzzy comprehensive evaluation model can be constructed to verify the 8 strategies, removing one strategy that does not meet the premium cap requirements of the China Banking and Insurance Regulatory Commission (compliance score 0.75 < 0.8) and one strategy that lacks sufficient marketing resources (resource suitability score 0.78 < 0.8), ultimately retaining 6 high-quality strategies. Further, the strategies can be divided into three sub-strategy sets according to three customer groups (2 for middle-aged and elderly, 2 for young people, and 2 for families), and then merged into a global recommendation strategy set. Each strategy is assigned a unique identifier and its core parameters are marked.

[0146] Furthermore, taking the "Youth Lightening-01" strategy in the youth subpopulation as an example (fitness 0.82), the initial individual can be traced back through logs to trace its complete trajectory of 70 iterations. One cross-subpopulation migration event was further extracted: in the 40th iteration, the individual migrated from the family comprehensive support subpopulation (source subpopulation) to the youth lightening support subpopulation (target subpopulation). The core gene segment of the migration was the "online and offline combined reach parameter." The fitness before migration was 0.71, which increased to 0.78 after migration, indicating the migration was effective. There were no affiliation change events. Furthermore, the migration event report and the complete evolutionary trajectory (fitness change curve, gene segment mutation record) can be bound to this strategy and embedded into the strategy interpretation module to form an integrated output.

[0147] As another concrete example, in the medical technology scenario, 78 candidate strategies with a fitness greater than or equal to 0.7 can be extracted from 4 optimized subpopulations. The top 20% are screened to obtain 16 strategies, and after vector deduplication, 14 strategies are retained. Further, 3 strategies without medical device registration certificates (compliance score less than 0.8) and 2 strategies with excessive execution difficulty (execution difficulty score greater than 0.3) are removed, and finally 9 high-quality strategies are retained. Further, the strategies are classified into 4 sub-strategy sets according to 4 customer groups, which are then merged into a global recommendation strategy set, standardized and labeled, and stored. Furthermore, taking the "Top-Tier High-End-01" strategy in the tertiary hospital subpopulation as an example (fitness 0.85), we can trace back to the initial individual (fitness 0.60) and trace its 80-round iteration trajectory. We can extract one cross-subpopulation migration event: in the 60th iteration, it migrated from the mid-range equipment subpopulation of private institutions. The core gene segment of the migration was "bulk purchase discount configuration". The fitness before the migration was 0.78, and it increased to 0.83 after the migration, indicating that the migration was effective. There were no ownership change events. By binding the migration event and trajectory log with this strategy, we can form a queryable and interpretable integrated strategy output, providing trajectory support for the implementation of equipment procurement strategies in tertiary hospitals.

[0148] It is understandable that this embodiment enables full-link evolution trajectory to be traceable, realizes the integrated binding output of strategy and evolution trajectory, completely solves the black box problem of product recommendation, and enables business personnel to understand the source of the core advantages of the strategy through the trajectory.

[0149] Understandably, the first step of this invention is to acquire at least one customer segment, dividing customers into multiple segments based on their core characteristics. This upgrades the fragmented mapping of "single feature - single product" to a unified management of a group-level strategy pool. In this approach, when the mapping relationship changes, there's no need for manual adjustments to the matching relationship for each individual feature value; only the relevant rules need to be adjusted for each customer segment. This significantly reduces the workload of basic maintenance and solves the problem of high workload at the management granularity level.

[0150] Secondly, the second step of the solution, "constructing strategy evolution sub-populations corresponding to different customers within each customer segment, with different strategy evolution sub-populations corresponding to differentiated strategy search spaces," builds a dedicated recommendation strategy pool for each customer segment and sets a differentiated strategy search space for each sub-population. When the mapping relationship changes, there is no need to manually modify the product corresponding to each feature value one by one. It is only necessary to simply adjust the boundary of the strategy search space of the corresponding sub-population. All recommendation strategies within the sub-population will autonomously explore and evolve within the new search space boundary, automatically adapting to the new mapping relationship. There is no need for manual intervention to match them one by one, thus eliminating a large amount of manual workload in terms of adaptation method.

[0151] Furthermore, the third step of the solution constructs a strategy fitness function that integrates real-time external risk signals, providing an automatically updatable evaluation criterion for the autonomous evolution of the strategy. This fitness function is not fixed but can integrate external change signals (external risk signals) in real time. These external changes are essentially the core reasons that cause changes in the "user feature value-product" mapping relationship. The function automatically transforms these external changes into evaluation criteria for strategy evolution, eliminating the need for manually redefining the "feature-product" matching rules. This allows the strategy to automatically evolve in the direction of adapting to the new mapping relationship, avoiding the enormous workload of manually reconstructing the rules in terms of evaluation criteria.

[0152] Then, the fourth step of the solution, "based on the policy fitness function, drives multiple policy evolution subpopulations to perform iterative optimization in parallel, and performs policy knowledge transfer operations across policy evolution subpopulations during the iteration process," achieves fully automated optimization of recommended policies and reuse of high-quality policies. Driven by the fitness function, each policy evolution subpopulation autonomously performs iterative optimization in parallel: automatically calculating the fitness of each policy, selecting high-quality policies that fit the new mapping relationship, and eliminating invalid policies that do not conform to the new mapping relationship; at the same time, the cross-population policy knowledge transfer operation can reuse the high-quality policies of one group to other groups with similar needs, without the need for manual establishment of similar mapping relationships for different groups, further reducing the workload of duplication. The entire iterative optimization and knowledge transfer process is fully automated, without the need for manual intervention in screening, matching, and reuse, completely eliminating the large amount of workload of manually re-screening and matching policies when the mapping relationship changes, which is the core link in solving this technical problem.

[0153] Finally, the fifth step of the solution, "outputting a recommended strategy set from the optimized strategy evolution subpopulation and generating explanatory information about the strategy evolution trajectory association of the recommended strategy set," completes the automatic output and explanation of high-quality strategies under the new mapping relationship. The final output of the solution is a high-quality recommended strategy set that has undergone automated iterative optimization, rather than a fixed mapping table. This set can be directly used to recommend products to customers without the need for manual screening and organization of strategies. Simultaneously, the solution automatically generates explanatory information about the strategy evolution trajectory, further reducing subsequent manual workload.

[0154] In summary, when the recommendation mapping relationship between "user feature value - product" changes, existing technologies require manual intervention throughout the entire process, resulting in a huge workload. However, this solution, through the synergistic effect of the above five steps, forms a complete closed loop of "group overall management - search space adaptation - automatic update of evaluation criteria - automatic iterative optimization of strategies - automatic output of high-quality strategies." Only a few core parameters (such as search space boundaries and fitness function weights) need to be adjusted manually. All other processes for adapting to the new mapping relationship are completed automatically by the solution. There is no need to manually re-establish the recommendation mapping relationship between feature values ​​and products one by one, thus fundamentally solving the technical problem of the large workload of recommending products when the mapping relationship changes in existing technologies.

[0155] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention. Software tools or components not belonging to this company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.

[0156] In one embodiment, a product recommendation device is provided, which corresponds one-to-one with the product recommendation method described in the above embodiments. For example... Figure 5 As shown, the product recommendation device includes a data acquisition module 510, a first construction module 520, a second construction module 530, a population optimization module 540, and an information generation module 550. Detailed descriptions of each functional module are as follows: Module 510 is used to acquire at least one customer segment; The first construction module 520 is used to construct a strategy evolution subpopulation formed by recommendation strategies for different customers in each of the customer segments, wherein different strategy evolution subpopulations correspond to differentiated strategy search spaces. The second construction module 530 is used to construct a policy fitness function that integrates real-time external risk signals; The population optimization module 540 is used to drive the multiple policy evolution subpopulations to perform iterative optimization in parallel based on the policy fitness function, and to perform policy knowledge transfer operations across the policy evolution subpopulations during the iteration process. The information generation module 550 is used to output a recommended strategy set from the optimized strategy evolution subpopulation and generate explanatory information related to the strategy evolution trajectory of the recommended strategy set.

[0157] In one embodiment, the first construction module 520 is specifically used for: For each customer in each of the aforementioned customer segments, match the corresponding product portfolio and marketing strategy for each customer; Chromosome encoding is performed on the product portfolio and marketing strategy of each customer to obtain the corresponding product portfolio gene segment and marketing strategy gene segment for each customer. The product portfolio gene segment and the marketing strategy gene segment corresponding to each customer are combined to form a chromosome corresponding to each customer, which serves as an individual in the strategy evolution subpopulation.

[0158] In one embodiment, the first construction module 520 is further configured to: Different initialization strategy distributions, genetic operation operator parameters, or local exploration targets are configured for different policy evolution subpopulations to form each policy evolution subpopulation with a corresponding differentiated policy search space.

[0159] In one embodiment, the policy fitness function is constructed as follows: The value of the strategy fitness function increases in response to the expected return index of the recommended strategy, and decreases in response to the dynamic risk index of the strategy calculated based on the recommended strategy.

[0160] In one embodiment, the population optimization module 540 is specifically used for: Based on the policy fitness function, the multiple policy evolution subpopulations are driven to perform iterative optimization in parallel; For the source subpopulation and target subpopulation in the multiple strategy evolution subpopulations, when the strategy distribution distance between the source subpopulation and the target subpopulation is less than a threshold, the individual corresponding to the recommended strategy whose fitness ranking is in the top N in the source subpopulation is introduced into the target subpopulation.

[0161] In one embodiment, the information generation module 550 is specifically used for: Output a set of recommended strategies from the optimized strategy evolution subpopulation; The association changes and migration events of the subpopulations experienced by each recommendation strategy in the recommendation strategy set during the evolution process are traced.

[0162] In one embodiment, the population optimization module 540 is further configured to: Within each subpopulation, the following process is repeated: calculating the fitness of each individual within the subpopulation based on the policy fitness function; Select target individuals whose fitness value is higher than a preset value; Perform crossover and mutation on the target individual to generate a new individual; The new individuals are merged with the target individuals of the previous generation to form a new generation of subpopulation.

[0163] Understandably, the first step of this invention is to acquire at least one customer segment, dividing customers into multiple segments based on their core characteristics. This upgrades the fragmented mapping of "single feature - single product" to a unified management of a group-level strategy pool. In this approach, when the mapping relationship changes, there's no need for manual adjustments to the matching relationship for each individual feature value; only the relevant rules need to be adjusted for each customer segment. This significantly reduces the workload of basic maintenance and solves the problem of high workload at the management granularity level.

[0164] Secondly, the second step of the solution, "constructing strategy evolution sub-populations corresponding to different customers within each customer segment, with different strategy evolution sub-populations corresponding to differentiated strategy search spaces," builds a dedicated recommendation strategy pool for each customer segment and sets a differentiated strategy search space for each sub-population. When the mapping relationship changes, there is no need to manually modify the product corresponding to each feature value one by one. It is only necessary to simply adjust the boundary of the strategy search space of the corresponding sub-population. All recommendation strategies within the sub-population will autonomously explore and evolve within the new search space boundary, automatically adapting to the new mapping relationship. There is no need for manual intervention to match them one by one, thus eliminating a large amount of manual workload in terms of adaptation method.

[0165] Furthermore, the third step of the solution constructs a strategy fitness function that integrates real-time external risk signals, providing an automatically updatable evaluation criterion for the autonomous evolution of the strategy. This fitness function is not fixed but can integrate external change signals (external risk signals) in real time. These external changes are essentially the core reasons that cause changes in the "user feature value-product" mapping relationship. The function automatically transforms these external changes into evaluation criteria for strategy evolution, eliminating the need for manually redefining the "feature-product" matching rules. This allows the strategy to automatically evolve in the direction of adapting to the new mapping relationship, avoiding the enormous workload of manually reconstructing the rules in terms of evaluation criteria.

[0166] Then, the fourth step of the solution, "based on the policy fitness function, drives multiple policy evolution subpopulations to perform iterative optimization in parallel, and performs policy knowledge transfer operations across policy evolution subpopulations during the iteration process," achieves fully automated optimization of recommended policies and reuse of high-quality policies. Driven by the fitness function, each policy evolution subpopulation autonomously performs iterative optimization in parallel: automatically calculating the fitness of each policy, selecting high-quality policies that fit the new mapping relationship, and eliminating invalid policies that do not conform to the new mapping relationship; at the same time, the cross-population policy knowledge transfer operation can reuse the high-quality policies of one group to other groups with similar needs, without the need for manual establishment of similar mapping relationships for different groups, further reducing the workload of duplication. The entire iterative optimization and knowledge transfer process is fully automated, without the need for manual intervention in screening, matching, and reuse, completely eliminating the large amount of workload of manually re-screening and matching policies when the mapping relationship changes, which is the core link in solving this technical problem.

[0167] Finally, the fifth step of the solution, "outputting a recommended strategy set from the optimized strategy evolution subpopulation and generating explanatory information about the strategy evolution trajectory association of the recommended strategy set," completes the automatic output and explanation of high-quality strategies under the new mapping relationship. The final output of the solution is a high-quality recommended strategy set that has undergone automated iterative optimization, rather than a fixed mapping table. This set can be directly used to recommend products to customers without the need for manual screening and organization of strategies. Simultaneously, the solution automatically generates explanatory information about the strategy evolution trajectory, further reducing subsequent manual workload.

[0168] In summary, when the recommendation mapping relationship between "user feature value - product" changes, existing technologies require manual intervention throughout the entire process, resulting in a huge workload. However, this solution, through the synergistic effect of the above five steps, forms a complete closed loop of "group overall management - search space adaptation - automatic update of evaluation criteria - automatic iterative optimization of strategies - automatic output of high-quality strategies." Only a few core parameters (such as search space boundaries and fitness function weights) need to be adjusted manually. All other processes for adapting to the new mapping relationship are completed automatically by the solution. There is no need to manually re-establish the recommendation mapping relationship between feature values ​​and products one by one, thus fundamentally solving the technical problem of the large workload of recommending products when the mapping relationship changes in existing technologies.

[0169] Based on the above product recommendation methods, such as Figure 6 As shown in the diagram, this embodiment of the invention also provides a structural schematic of an apparatus for a product recommendation method, the apparatus including a processor 61 and a memory 62 coupled to the processor 61. The memory 62 stores a computer program, which, when executed by the processor 61, causes the processor 61 to perform the steps of the product recommendation method described above.

[0170] For further details regarding the implementation of the above technical solution by the processor 61 in the device for the above product recommendation method steps, please refer to the description of the product recommendation method provided in the above embodiments of the invention, which will not be repeated here.

[0171] The processor 61 can also be called a CPU (Central Processing Unit). The processor 61 may be an integrated circuit chip with signal processing capabilities. The processor 61 can also be a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor, or the processor 61 can be any conventional processor.

[0172] like Figure 7 As shown in the diagram, this embodiment of the invention also provides a schematic diagram of a computer-readable storage medium, on which a readable computer program 71 is stored. The computer program 71 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in various embodiments of the invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks or optical disks, ROM (Read-Only Memory), RAM (Random Access Memory), or terminal devices such as computers, servers, mobile phones, and tablets.

[0173] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, apparatuses, or modules, and may be electrical, mechanical, or other forms.

[0174] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0175] Furthermore, the functional modules in the various embodiments of this invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0176] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0177] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium, or a semiconductor medium (e.g., SSD (solid state disk)).

[0178] The technical solution provided by the present invention has been described in detail above. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

[0179] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0180] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0181] 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.

[0182] 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.

[0183] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A product recommendation method, characterized in that, include: Acquire at least one customer segment; Construct strategy evolution subpopulations corresponding to different customers in each of the aforementioned customer segments, wherein different strategy evolution subpopulations correspond to differentiated strategy search spaces; Construct a policy fitness function that integrates real-time external risk signals; Based on the policy fitness function, the multiple policy evolution subpopulations are driven to perform iterative optimization in parallel, and policy knowledge transfer operations are performed across the policy evolution subpopulations during the iteration process; Output a recommended strategy set from the optimized strategy evolution subpopulation, and generate explanatory information related to the strategy evolution trajectory of the recommended strategy set.

2. The product recommendation method according to claim 1, characterized in that, The strategy evolution subpopulation formed by constructing each customer recommendation strategy corresponding to each of the aforementioned customer segments includes: For each customer in each of the aforementioned customer segments, match the corresponding product portfolio and marketing strategy for each customer; Chromosome encoding is performed on the product portfolio and marketing strategy of each customer to obtain the corresponding product portfolio gene segment and marketing strategy gene segment for each customer. The product portfolio gene segment and the marketing strategy gene segment corresponding to each customer are combined to form a chromosome corresponding to each customer, which serves as an individual in the strategy evolution subpopulation.

3. The product recommendation method according to claim 2, characterized in that, After combining the product portfolio gene segment and the marketing strategy gene segment corresponding to each customer to form a chromosome corresponding to each customer, which serves as an individual in the strategy evolution subpopulation, the method further includes: Different initialization strategy distributions, genetic operation operator parameters, or local exploration targets are configured for different policy evolution subpopulations to form each policy evolution subpopulation with a corresponding differentiated policy search space.

4. The product recommendation method according to any one of claims 1-3, characterized in that, The policy fitness function is constructed as follows: The value of the strategy fitness function increases in response to the expected return index of the recommended strategy, and decreases in response to the dynamic risk index of the strategy calculated based on the recommended strategy.

5. The product recommendation method according to claim 1, characterized in that, The process of driving the multiple policy evolution subpopulations to perform iterative optimization in parallel based on the policy fitness function, and performing policy knowledge transfer operations across the policy evolution subpopulations during the iteration process, includes: Based on the policy fitness function, the multiple policy evolution subpopulations are driven to perform iterative optimization in parallel; For the source subpopulation and target subpopulation in the multiple strategy evolution subpopulations, when the strategy distribution distance between the source subpopulation and the target subpopulation is less than a threshold, the individual corresponding to the recommended strategy whose fitness ranking is in the top N in the source subpopulation is introduced into the target subpopulation.

6. The product recommendation method according to claim 1, characterized in that, The step of outputting a recommended strategy set from the optimized strategy evolution subpopulation and generating explanatory information related to the strategy evolution trajectory of the recommended strategy set includes: Output a set of recommended strategies from the optimized strategy evolution subpopulation; The evolutionary process of each recommendation strategy in the recommendation strategy set involves tracing the subpopulation affiliation changes and migration events experienced by each recommendation strategy during its evolution.

7. The product recommendation method according to claim 1, characterized in that, The step of driving the multiple policy evolution subpopulations to perform iterative optimization in parallel based on the policy fitness function includes: Within each subpopulation, the following process is repeated: calculating the fitness of each individual within the subpopulation based on the policy fitness function; Select target individuals whose fitness value is higher than a preset value; Perform crossover and mutation on the target individual to generate a new individual; The new individuals are merged with the target individuals of the previous generation to form a new generation of subpopulation.

8. A product recommendation device, characterized in that, include: The acquisition module is used to acquire at least one customer segment. The first construction module is used to construct a strategy evolution subpopulation formed by recommendation strategies for different customers in each of the customer segments, wherein different strategy evolution subpopulations correspond to differentiated strategy search spaces; The second construction module is used to construct a policy fitness function that integrates real-time external risk signals; The population optimization module is used to drive the multiple policy evolution subpopulations to perform iterative optimization in parallel based on the policy fitness function, and to perform policy knowledge transfer operations across the policy evolution subpopulations during the iteration process; The information generation module is used to output a recommended strategy set from the optimized strategy evolution subpopulation and generate explanatory information related to the strategy evolution trajectory of the recommended strategy set.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the product recommendation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the product recommendation method as described in any one of claims 1 to 7.

Images