Product recommendation method and device based on artificial intelligence, equipment and storage medium

Abstract

The application relates to the technical field of artificial intelligence, and discloses a product recommendation method and device based on artificial intelligence, equipment and a storage medium, which comprises the following steps: extracting target customer data and determining a potential feature vector thereof; clustering a plurality of potential feature vectors to determine a target segmented demand group to which each target customer belongs; using a reinforcement learning model to analyze the target segmented demand group and an original product feature vector to generate a plurality of recommended product feature vectors corresponding to each target customer; matching the potential feature vector with the recommended product feature vector to determine a target product corresponding to a recommended product feature vector with the highest matching degree, and recommending the target product to the target customer. The method is suitable for personalized recommendation in scenarios such as financial digital products, medical drugs and medical guidance, and can effectively improve the recommendation accuracy by dynamically optimizing the product features through reinforcement learning.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, finance, and healthcare, and in particular to a product recommendation method, apparatus, device, and storage medium based on artificial intelligence. Background Technology

[0002] Traditional product design and recommendation rely on conventional statistical models such as linear regression and statistical table analysis, combined with historical data and limited static market research results. Product parameters are manually set, and basic customer information is collected through questionnaires and telephone interviews. Recommendations are then made based on fixed rules or simple decision trees. However, these methods cannot automatically generate product features that match customers' potential needs based on multi-source dynamic customer data, resulting in recommendations with low matching degree to customers' actual needs and insufficient personalization.

[0003] For example, in the fintech field, systems often recommend low-yield, long-term locked-in wealth management products based solely on a conservative risk questionnaire completed by a customer six months prior and limited historical transaction records. However, if the customer has recently been frequently browsing highly liquid, medium-to-high-risk products and has received large sums of money, their actual risk appetite has changed, and traditional solutions cannot dynamically adjust accordingly. Similarly, in the healthcare field, traditional medical guidance systems recommend standardized medications and departments based solely on the initial type 2 diabetes label and fixed guidelines. However, if the patient's wearable devices have recorded continuous abnormal fluctuations in blood sugar, existing solutions cannot integrate this dynamic vital sign data, resulting in delayed recommendations.

[0004] Therefore, how to automatically generate product features that match customers' potential needs based on multi-source dynamic data, and improve the matching degree and personalization of recommendation results with actual needs, is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] This invention provides a product recommendation method, apparatus, device, and storage medium based on artificial intelligence, aiming to solve the problems of poor dynamic matching between products and customer needs and insufficient personalization in the prior art.

[0006] Firstly, an artificial intelligence-based product recommendation method is provided, including: Feature extraction is performed on the customer data corresponding to the target customer to determine the potential feature vector corresponding to the target customer; Clustering is performed on the potential feature vectors corresponding to multiple target customers to obtain at least one cluster, and the target segment demand group to which each target customer belongs is determined, wherein the target segment demand group corresponds to a cluster. A reinforcement learning model is used to analyze the target segment demand group and N original product feature vectors to which each target customer belongs, and to determine M recommended product feature vectors corresponding to each target customer, where N and M are both integers greater than 1; For each target customer, the potential feature vector and M recommended product feature vectors are matched, and the recommended product feature vector with the highest matching degree is determined as the target product feature vector. The target product corresponding to the target product feature vector is recommended to the target customer.

[0007] Secondly, an artificial intelligence-based product recommendation device is provided, including: The customer feature extraction module is used to extract features from the customer data corresponding to the target customer and determine the potential feature vector corresponding to the target customer. The segmented demand determination module is used to cluster the potential feature vectors corresponding to multiple target customers to obtain at least one cluster, and to determine the target segmented demand group to which each target customer belongs, wherein the target segmented demand group corresponds to a cluster. The recommended product determination module is used to analyze the target segment demand group and N original product feature vectors to which each target customer belongs using a reinforcement learning model, and determine M recommended product feature vectors corresponding to each target customer, where N and M are both integers greater than 1; The target product determination module is used to match the potential feature vector and M recommended product feature vectors corresponding to each target customer, and determine the recommended product feature vector with the highest matching degree as the target product feature vector. The product recommendation module is used to recommend the target product corresponding to the target product feature vector to the target customer.

[0008] 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 above-described artificial intelligence-based product recommendation method.

[0009] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned artificial intelligence-based product recommendation method.

[0010] The aforementioned AI-based product recommendation method, device, equipment, and storage medium extracts features from target customer data and maps them to latent feature vectors, enabling a deep characterization of customers' multidimensional needs. It then clusters multiple latent feature vectors to segment target demand groups, achieving precise customer segmentation and differentiated positioning. Based on this, a reinforcement learning model dynamically analyzes the target demand groups and the original product feature vectors, automatically generating multiple recommended product feature vectors for each customer. This overcomes the limitations of traditional static rule-based recommendations, allowing product features to continuously optimize based on customer feedback. Finally, through matching latent feature vectors with recommended product feature vectors, the target product feature vector with the highest matching degree is determined, and the corresponding product is recommended, achieving a precise mapping between customer needs and product features. This solution is applicable to complex scenarios such as financial digital products, medical products, and medical guidance, effectively solving the problems of poor dynamic matching between products and customer needs and insufficient personalization in existing technologies, significantly improving recommendation accuracy and user experience. Attached Figure Description

[0011] 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.

[0012] Figure 1 This is a schematic diagram of an application environment for an artificial intelligence-based product recommendation method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a product recommendation method based on artificial intelligence according to an embodiment of the present invention; Figure 3 yes Figure 2 A schematic diagram of a specific implementation method for step S203; Figure 4 yes Figure 2 A flowchart illustrating a specific implementation of step S204; Figure 5 This is another flowchart illustrating the product recommendation method based on artificial intelligence in one embodiment of the present invention; Figure 6 yes Figure 5 A schematic diagram of a specific implementation method for step S502; Figure 7 This is a schematic diagram of a product recommendation device based on artificial intelligence according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 9 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0013] 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.

[0014] The AI-based product recommendation method provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can receive customer data corresponding to the target customer through the client, extract features from the customer data, and determine the potential feature vectors corresponding to the target customers. It then clusters the potential feature vectors corresponding to multiple target customers to obtain at least one cluster, determining the target segment demand group to which each target customer belongs. A reinforcement learning model is used to analyze the target segment demand group and the original product feature vectors, generating multiple recommended product feature vectors corresponding to each target customer. The potential feature vectors are matched with the recommended product feature vectors to determine the target product feature vector with the highest matching degree. Finally, the target product corresponding to the target product feature vector is recommended to the target customer, and the recommendation result is fed back to the client. In this invention, for complex recommendation scenarios such as financial digital products, medical drugs, and medical guidance, a reinforcement learning model can be used to dynamically optimize product features. First, clustering is used to achieve accurate segmentation of customer needs. Then, reinforcement learning generates suitable recommended product feature vectors and performs matching and filtering, selecting the target product with the highest matching degree to recommend to the customer. This effectively breaks through the limitations of traditional static rule-based recommendations, achieving continuous automatic optimization of product features and significantly improving the accuracy of personalized recommendations and user experience. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.

[0015] Please see Figure 2 As shown, Figure 2 A flowchart illustrating an artificial intelligence-based product recommendation method provided in an embodiment of the present invention includes the following steps: S201: Extract features from the customer data corresponding to the target customer to determine the potential feature vector corresponding to the target customer; S202: Cluster the potential feature vectors corresponding to multiple target customers to obtain at least one cluster, and determine the target segment demand group to which each target customer belongs, wherein the target segment demand group corresponds to a cluster. S203: Use a reinforcement learning model to analyze the target segment demand group and N original product feature vectors to which each target customer belongs, and determine M recommended product feature vectors corresponding to each target customer, where N and M are both integers greater than 1; S204: Perform a matching process on the potential feature vector and M recommended product feature vectors corresponding to each target customer, and determine the recommended product feature vector with the highest matching degree as the target product feature vector; S205: Recommend the target product corresponding to the target product feature vector to the target customer.

[0016] The product recommendation method provided by this invention can be applied to product recommendation scenarios in fields such as finance and healthcare, and is typically implemented through a server-side implementation. For example, in the insurance field, insurance companies need to recommend suitable insurance products based on customers' personalized needs; in the healthcare field, medical institutions need to recommend personalized health management plans based on patients' health conditions to improve service efficiency and user experience.

[0017] Here, target customers refer to individuals or groups of customers who require product recommendations; customer data refers to customer information obtained from multidimensional data sources. The latent feature vector is a low-dimensional latent representation obtained by extracting features from customer data and reducing its dimensionality. It represents the digital mapping of the target customer's core needs attributes in the latent space, with each dimension corresponding to the target customer's preference intensity at different levels, used for subsequent clustering and matching.

[0018] For example, in insurance recommendation scenarios, the target customers are individual customers or corporate groups with the intention to purchase insurance. Customer data includes basic information, behavioral data, claims history, health data, and social behavior. Basic information includes age, occupation, and annual income; behavioral data includes product browsing history and consultation records; claims history includes the number of past claims and the amount claimed; health data includes physical examination reports, chronic disease diagnosis records, and BMI index; and social behavior includes interactions in insurance communities and product reviews. In medical recommendation scenarios, the target customers are individual patients or patient groups who need health management or medical guidance. Customer data includes basic information, behavioral data, medical history, health data, and social behavior. Basic information includes age, gender, and occupation; behavioral data includes department browsing history, registration records, and consultation records; medical history includes past diagnosis records, medication records, and surgical records; health data includes physical examination reports, chronic disease diagnosis records, and physiological indicators; and social behavior includes interactions in health communities and doctor reviews.

[0019] As an example, in step S201, the server establishes a communication connection with the database and reads customer data corresponding to the target customer from the database. Then, the server can extract features from the customer data to obtain the latent feature vectors corresponding to the target customer. For example, the server can use pre-set feature extraction tools such as deep autoencoders or variational autoencoders to encode and reduce the dimensionality of the customer data to obtain the latent feature vectors.

[0020] Among them, the target segmentation demand group refers to the customer group obtained by clustering the potential feature vectors of multiple target customers. Customers in the same group have similar demand characteristics.

[0021] As an example, in step S202, the server clusters the latent feature vectors corresponding to multiple target customers to determine the target segment demand group to which each target customer belongs. The server calculates the similarity between each latent feature vector and uses a preset K-Means clustering algorithm or other clustering algorithms such as Gaussian Mixture Model (GMM) to group the latent feature vectors. Customers within the same group have higher similarity in latent features, while those between different groups have lower similarity. Each group is a segment demand group. The server calculates the similarity between the latent feature vector corresponding to the target customer and the group center vector of each segment demand group, and determines the segment demand group to which the target customer belongs based on the segment demand group corresponding to the maximum similarity.

[0022] For example, in insurance recommendation scenarios, the server clusters the latent feature vectors of multiple target customers to obtain several sub-demand groups, such as critical illness protection, retirement security, and family asset protection. In medical recommendation scenarios, the server clusters the latent feature vectors of multiple patients to obtain several sub-demand groups, such as chronic disease management, rehabilitation care, and health monitoring. For a given target customer or patient, the server calculates the similarity between their latent feature vector and the central vectors of each group, assigning them to the sub-demand group with the highest similarity; this sub-demand group corresponds to a cluster.

[0023] The reinforcement learning model refers to a machine learning model that learns optimal decision-making strategies through interaction with the environment. It includes core components such as a state space, action space, reward function, and policy network. In this embodiment, the policy network analyzes and processes the original product feature vectors based on the target segmentation demand group characteristics in the state space, determining M recommended product feature vectors. The original product feature vectors are pre-designed vector representations containing various product parameters, such as the coverage amount, premium, coverage scope, and term of insurance products, or the medication plan type, monitoring frequency, treatment cycle, and applicable population characteristics of medical and health management plans. The recommended product feature vectors are the M product feature vectors determined after analysis and processing by the reinforcement learning model, which meet the needs of the target customers.

[0024] As an example, in step S203, the server uses a reinforcement learning model to analyze the target segment demand group to which each target customer belongs and N original product feature vectors, determining M recommended product feature vectors corresponding to each target customer. The policy network of the reinforcement learning model adopts a multi-layer fully connected neural network, including an input layer, a hidden layer, and an output layer. The input layer receives the group characteristics of the target segment demand group, the hidden layer performs non-linear transformations on the state, and the output layer outputs M recommended product feature vectors. The server inputs the N original product feature vectors and the group characteristics of the target segment demand group into the policy network, and after forward propagation calculation, obtains the M recommended product feature vectors.

[0025] For example, in insurance recommendation scenarios, the strategy network takes the group characteristics of health-conscious groups as input and analyzes and processes the feature vectors of original critical illness insurance, medical insurance, etc. The original product feature vectors include designed insurance parameters such as sum insured, premium, coverage scope, and term. It determines and generates feature vectors for critical illness insurance with expanded coverage and moderately increased premiums, and feature vectors for medical insurance with reduced deductibles. In medical recommendation scenarios, the strategy network takes the group characteristics of chronic disease management groups as input and analyzes and processes the feature vectors of original antihypertensive drugs, blood glucose monitoring, etc. The original product feature vectors include designed health management parameters such as medication regimen type, monitoring frequency, treatment cycle, and applicable population characteristics. It determines and generates feature vectors for antihypertensive drugs after adjusting the medication regimen type and feature vectors for high-frequency blood glucose management regimens.

[0026] Among them, the target product feature vector refers to the recommended product feature vector that has the highest matching degree with the potential feature vector of the target customer.

[0027] As an example, in step S204, the server performs a matching process between the latent feature vector and the M recommended product feature vectors corresponding to each target customer. The server calculates the similarity between the latent feature vector and each recommended product feature vector, and determines the recommended product feature vector corresponding to the maximum similarity as the target product feature vector.

[0028] For example, in an insurance recommendation scenario, the server calculates the cosine similarity between the latent feature vector of the target customer and the feature vectors of each recommended insurance product, and determines the critical illness insurance feature vector with the highest similarity as the target product feature vector. In a medical recommendation scenario, the server calculates the cosine similarity between the latent feature vector of the patient and the feature vectors of each recommended health management plan, and determines the antihypertensive drug plan feature vector with the highest similarity as the target product feature vector.

[0029] The target product refers to the configuration scheme containing specific product parameters, determined by the target product feature vector.

[0030] As an example, in step S205, the server recommends the target product corresponding to the target product feature vector to the target customer. The server encapsulates the product parameters contained in the target product feature vector into a target product configuration scheme, and sends this product configuration scheme to the target customer's corresponding terminal device through a preset recommendation channel, thus completing the product recommendation. This recommendation channel includes methods such as application push notifications, SMS notifications, emails, or web page displays.

[0031] For example, in an insurance recommendation scenario, the server maps the identified critical illness insurance feature vector to the terms and rates of specific insurance products, generating an insurance product configuration plan that includes the sum insured, premium, coverage scope, and term, and pushes it to the target customer's mobile phone through an insurance application. In a medical recommendation scenario, the server maps the identified antihypertensive medication regimen feature vector to specific drugs and medical services, generating a health management configuration plan that includes medication regimen type, monitoring frequency, and treatment cycle, and pushes it to the patient's mobile phone through a medical application.

[0032] The AI-based product recommendation method provided in this embodiment first extracts features from customer data corresponding to the target customer and maps them into latent feature vectors, which can deeply characterize the multidimensional needs of customers. Then, it clusters multiple latent feature vectors to divide the target into segmented demand groups, achieving accurate segmentation and differentiated positioning of customer needs. Based on this, a reinforcement learning model is used to dynamically analyze the target segmented demand groups and the original product feature vectors to determine multiple recommended product feature vectors for each customer, breaking through the limitations of traditional static rule-based recommendations. Finally, through matching the latent feature vectors and recommended product feature vectors, the target product feature vector with the highest matching degree is determined and the corresponding product is recommended, achieving accurate mapping between customer needs and product features. This solution is applicable to complex scenarios such as financial digital products and healthcare management, effectively solving the problem of poor dynamic matching between products and customer needs in existing technologies, and significantly improving recommendation accuracy and user experience.

[0033] In one embodiment, such as Figure 3As shown, in step S203, a reinforcement learning model is used to analyze the target segment demand group and N original product feature vectors to which each target customer belongs, and to determine M recommended product feature vectors corresponding to each target customer, including: S301: Obtain market environment data, extract features from the market environment data to obtain an environment vector; use the target segment demand group to which the target customer belongs and the environment vector as a state vector; S302: Construct a product feature vector set, and select a product feature vector to be optimized from the product feature vector set as the action space. The product feature vector set initially includes all original product feature vector sets. S303: Input the state vector and the action space into the reinforcement learning model for learning, obtain the output value of the comprehensive reward function in the reinforcement learning model, and determine the output value as the reward score of the product feature vector to be optimized; S304: Based on the reward score, optimize the product feature vector to be optimized to obtain an optimized product feature vector, write the optimized product feature vector into the product feature vector set, and return to execute the step of selecting a product feature vector to be optimized from the product feature vector set as the action space until the reward score reaches the optimization target; S305: Output the M recommended product feature vectors that enable the reward score to reach the optimization objective.

[0034] Market environment data refers to external environmental factors that influence product recommendation decisions, including market policies, economic indicators, industry trends, and competitive landscape. The environment vector is a low-dimensional vector representation obtained after feature extraction from the market environment data, used to characterize the overall state of the current market environment.

[0035] As an example, in step S301, the server acquires market environment data, extracts features from the market environment data, and obtains an environment vector. The server obtains current market environment data from the market data interface, including market policy adjustment information, GDP growth rate, inflation rate, interest rate level, industry competition index, etc., cleans, encodes, and standardizes this data, concatenates it into a high-dimensional market environment feature vector, and then compresses it into a low-dimensional environment vector through a preset dimensionality reduction network. The server concatenates the group feature vector of the target customer's target segment demand group with the environment vector to form a state vector.

[0036] For example, in an insurance recommendation scenario, the server acquires current market environment data, including information such as policies on rising medical costs, GDP growth rate, benchmark interest rate cuts, and intensified competition in the insurance industry, and extracts an environment vector. The server then concatenates the group characteristic vector of the health-conscious group with this environment vector to form a state vector, which simultaneously contains customer demand characteristics and market environment characteristics. In a medical recommendation scenario, the server acquires current market environment data, including information such as medical insurance catalog adjustment policies, changes in medical resource supply, seasonal epidemic trends, and drug price fluctuations, and extracts an environment vector. The server then concatenates the group characteristic vector of the chronic disease management group with this environment vector to form a state vector, which simultaneously contains patient demand characteristics and medical environment characteristics.

[0037] Here, the product feature vector set refers to the set of product feature vectors used for iterative optimization. Initially, it contains all original product feature vectors, and as the optimization process progresses, optimized product feature vectors are continuously added. The action space refers to the set of actions that the reinforcement learning model can execute in the current state; that is, selecting a product feature vector from the product feature vector set for optimization.

[0038] As an example, in step S302, the server constructs a product feature vector set, initially including all original product feature vectors in this set. The server selects a product feature vector to be optimized from this set as the current action and inputs it into the reinforcement learning model for subsequent learning and optimization.

[0039] For example, in an insurance recommendation scenario, the server constructs a product feature vector set, initially containing feature vectors for original insurance products such as critical illness insurance, medical insurance, accident insurance, life insurance, and annuity insurance. The server selects the critical illness insurance feature vector from this set as the current action. This critical illness insurance feature vector includes parameters such as the sum insured, premium, coverage scope, and term. In a medical recommendation scenario, the product feature vector set initially contains feature vectors for original health management plans such as antihypertensive medication plans, blood glucose monitoring plans, rehabilitation training plans, and nutritional intervention plans. The server selects the antihypertensive medication plan feature vector from this set as the current action. This antihypertensive medication plan feature vector includes parameters such as medication plan type, monitoring frequency, treatment duration, and applicable population characteristics.

[0040] The comprehensive reward function is a reward calculation function weighted by market reaction and profitability indicators. It is used to evaluate the reward score of a reinforcement learning model after performing a specific action in a specific state. The reward score refers to the evaluation result of the comprehensive reward function on the combination of the current state and the executed action, and is used to guide the direction of subsequent optimization.

[0041] As an example, in step S303, the server inputs the state vector and action space into the reinforcement learning model for learning, and obtains the output value of the comprehensive reward function as the reward score. The formula for calculating the comprehensive reward function is: in, This represents the reward score for the combination of the current state and the action performed. Indicates market reaction weight. Indicates the target segmentation of demand groups Action The market reaction, Indicates the weight of profitability indicators. Indicates action The profitability metrics. Specifically, the server segments demand group characteristics and actions based on the target in the current state vector. The product feature vector is used to calculate the reward score through a comprehensive reward function. Market reaction is also included in the comprehensive reward function. Group preference characteristics and actions based on target segmentation of demand groups The degree of product parameter matching is determined. The server analyzes the group's sensitivity to and willingness to accept each product parameter, assesses the matching degree between the current product parameters and the group's preferences, and obtains a market response score. Profitability metrics. Based on action The product parameter configuration is determined through calculations based on a preset revenue structure and cost model. The server substitutes market response and profitability indicators into the above formula to calculate the reward score.

[0042] For example, in insurance recommendation scenarios, the overall reward function sets market response weights. and profitability indicator weights The server selects critical illness insurance for health-conscious groups. An evaluation is conducted. The server analyzes the historical preference patterns of the health-conscious population to assess the match between the current critical illness insurance coverage and premium parameters and the group's preferences, thus determining market response. Based on the premium income and claims cost model of this critical illness insurance, profitability indicators are determined. Substitute the values ​​into the formula to calculate the reward score. In the medical recommendation scenario, the market response weight is set in the comprehensive reward function. and profitability indicator weights The server's actions regarding selecting antihypertensive medication regimens for chronic disease management groups. An evaluation is conducted. The server analyzes the historical preference patterns of the chronic disease management group, assesses the match between parameters such as the type of medication, dosage, monitoring frequency, and treatment duration of the current antihypertensive regimen and the group's preferences, and determines market response. Based on the treatment revenue and drug cost model of this scheme, profitability indicators are determined. Substitute the values ​​into the formula to calculate the bonus score.

[0043] As an example, in step S304, the server optimizes the product feature vector to be optimized based on the reward score. The policy network of the reinforcement learning model adopts a deep Q-network framework, including an input layer, hidden layers, and an output layer. The input layer receives the state vector, the hidden layers perform nonlinear transformations on the state, and the output layer outputs the Q-value estimates for each action. The server selects an action based on the current Q-value, and after execution, the comprehensive reward function calculates the reward score based on the combination of the new state and the executed action, updating the network parameters according to the Q-value iteration formula. The calculation formula is as follows: Where Q(st,at) represents the Q-value estimate of the combination of the current state and the action to be performed, η represents the learning rate, Rt represents the reward score, and γ represents the discount factor. This represents the maximum Q-value estimate for the next state. The server iteratively executes the above updates until the reward score reaches the optimization target or the number of iterations reaches the preset upper limit, at which point the model is considered converged. The policy network adjusts the parameters of each dimension of the product feature vector based on the current Q-value and reward feedback, such as adjusting insurance parameters like coverage amount, premium, and coverage scope, or medical parameters like medication plan type and monitoring frequency. The server writes the optimized product feature vector into the product feature vector set and reselects the next product feature vector to be optimized for iteration until the reward score reaches the optimization target.

[0044] As an example, in step S505, the server selects M product feature vectors from the product feature vector set that optimize the reward score, and outputs them as recommended product feature vectors. These recommended product feature vectors contain iteratively optimized product parameter configurations and can be directly used in subsequent matching processing steps. For example, in an insurance recommendation scenario, the server outputs a feature vector for critical illness insurance with expanded coverage and moderately increased premiums, and a feature vector for medical insurance with reduced deductibles, as the recommended product feature vectors for that customer. In a medical recommendation scenario, the server outputs a feature vector for antihypertensive drugs with adjusted medication regimens and a feature vector for high-frequency blood glucose management regimens, as the recommended product feature vectors for that patient.

[0045] The AI-based product recommendation method provided in this embodiment configures a comprehensive reward function that takes into account market response and profitability indicators, enabling the reinforcement learning model to balance customer appeal and commercial benefits when optimizing product feature vectors. By fusing target segmentation demand group characteristics and market environment characteristics through state vector integration, the recommendation decision can adapt to market changes. Through an iterative optimization mechanism, product parameters are dynamically adjusted to generate personalized product feature schemes that meet customer needs and have market competitiveness, breaking through the limitations of traditional static rule-based recommendations and improving the accuracy, dynamism, and market response speed of product recommendations.

[0046] In one embodiment, such as Figure 4 As shown, in step S204, the potential feature vector and M recommended product feature vectors corresponding to each target customer are matched to determine the recommended product feature vector with the highest matching degree as the target product feature vector, including: S401: Map the potential feature vector of the target customer and the feature vectors of the M recommended products to the same feature space, and determine the customer mapping vector corresponding to the target customer and the product mapping vector corresponding to the feature vectors of the M recommended products respectively; S402: Based on the customer mapping vector and each product mapping vector, determine the matching degree between the target customer and each recommended product, and determine the feature vector of the recommended product corresponding to the product mapping vector with the highest matching degree as the feature vector of the target product.

[0047] In this context, the feature space refers to a low-dimensional vector space used to uniformly measure customer and product characteristics. The customer mapping vector is the vector representation obtained by mapping the latent feature vectors of target customers onto this feature space. The product mapping vector is the vector representation obtained by mapping the feature vectors of each recommended product onto this feature space.

[0048] As an example, in step S401, the server maps the latent feature vector of the target customer and the feature vectors of M recommended products to the same feature space, determining the customer mapping vector and the mapping vectors of each product. The server uses a pre-set mapping network, which includes an encoder layer and a projection layer. The encoder layer performs nonlinear transformations on the input features to extract high-level semantics, and the projection layer maps the high-level semantics to a feature space of a unified dimension. The server inputs the latent feature vector of the target customer into the mapping network and outputs the customer mapping vector; it inputs the feature vectors of each recommended product into the same mapping network and outputs the mapping vectors of each product, so that the customer mapping vector and the product mapping vector are in the same feature space, which facilitates subsequent similarity calculation. For example, in an insurance recommendation scenario, the server inputs the 128-dimensional latent feature vector of the target customer and the feature vectors of two recommended insurance products into the mapping network. The encoder layer extracts semantic features through multiple nonlinear transformations, and the projection layer maps the semantic features to a 256-dimensional unified feature space, outputting the customer mapping vector and the two product mapping vectors.

[0049] The matching degree refers to the similarity between the customer mapping vector and the product mapping vector, which is used to quantify the fit between the target customer and each recommended product.

[0050] As an example, in step S402, the server determines the matching degree between the target customer and each recommended product based on the customer mapping vector and the mapping vectors of each product. The server calculates the cosine similarity between the customer mapping vector and each product mapping vector, obtaining multiple similarity values. These similarity values ​​are then sorted from highest to lowest. Based on the product mapping vector corresponding to the highest similarity value in the sorting results, the recommended product feature vector corresponding to that mapping vector is determined as the target product feature vector. For example, in an insurance recommendation scenario, the server calculates the cosine similarity between the customer mapping vector and the critical illness insurance product mapping vector and the medical insurance product mapping vector, obtaining multiple similarity values. After sorting the similarity values, the critical illness insurance feature vector corresponding to the critical illness insurance product mapping vector with the highest similarity value is determined as the target product feature vector. This target product feature vector can be directly output as an insurance product configuration scheme containing the sum insured, premium, coverage scope, and term.

[0051] In the AI-based product recommendation method provided in this embodiment, by mapping the potential feature vectors of target customers and the feature vectors of recommended products to the same feature space for similarity calculation, the measurement bias caused by different feature spaces can be eliminated, making the matching results more accurate and comparable. By using quantitative similarity for objective matching judgment, the traditional manual experience judgment is replaced, improving the accuracy and consistency of supply and demand matching, reducing the risk of over-recommendation or products not matching needs, and providing accurate target product configuration schemes for subsequent personalized product recommendations.

[0052] In one embodiment, such as Figure 5 As shown, AI-based product recommendation methods also include: S501: Obtain market feedback data from the target customers regarding the target product; S502: Based on the potential feature vector corresponding to the target customer, the feature vector of the target product, and the market feedback data, update the model parameters of the reinforcement learning model to obtain an updated reinforcement learning model.

[0053] Among them, market feedback data refers to the actual response data generated by target customers after receiving the target product recommendation, including purchasing behavior, satisfaction evaluation, usage records, etc., which are used to measure the actual performance of the target product configuration in the real market environment.

[0054] As an example, in step S501, the server obtains market feedback data from target customers regarding the target product. The server collects actual response information from target customers through the feedback interface of the product recommendation system. In an insurance recommendation scenario, market feedback data includes whether the customer has viewed the details of the recommended product, submitted an insurance application, whether the policy has taken effect, and subsequent claims records. In a medical recommendation scenario, market feedback data includes whether the patient has adopted the recommended treatment plan, whether they have taken medication on time, whether they have undergone regular monitoring, follow-up visit feedback, and symptom improvement records.

[0055] As an example, in step S502, the server updates the model parameters of the reinforcement learning model based on the latent feature vectors corresponding to the target customer, the feature vectors of the target product, and market feedback data, thus obtaining an updated reinforcement learning model. Specifically, the server obtains the matching degree between the latent feature vectors corresponding to the target customer and the feature vectors of the target product, inputs the matching degree and market feedback data into a preset loss function to calculate the loss parameters, and then adjusts the model parameters of the reinforcement learning model through a gradient update algorithm based on the loss parameters, iteratively optimizing until the loss converges, so that the model strategy is aligned with the real market feedback.

[0056] In the AI-based product recommendation method provided in this embodiment, market feedback data from target customers on target products is obtained, and the model parameters of the reinforcement learning model are updated based on the latent feature vector, the target product feature vector, and the market feedback data. Since the market feedback data comes from real customer behavior, it can accurately reflect the actual effect of the product configuration scheme. The model parameters are deeply updated through backpropagation gradient descent, enabling the reinforcement learning model to continuously absorb real market information, realize the synchronous evolution of the model and the market, avoid the model strategy from being fixed in the simulated environment, and improve the long-term accuracy and market adaptability of product recommendations.

[0057] In one embodiment, such as Figure 6As shown, in step S502, based on the latent feature vector corresponding to the target customer, the feature vector of the target product, and the market feedback data, the model parameters of the reinforcement learning model are updated, including: S601: Obtain the matching degree between the potential feature vector corresponding to the target customer and the feature vector of the target product; S602: Based on a preset loss function, calculate the matching degree and the market feedback data to obtain the loss parameter; S603: Based on the loss parameters and combined with the gradient update algorithm, update the model parameters of the reinforcement learning model to obtain an updated reinforcement learning model; As an example, in step S601, the server obtains the matching degree between the latent feature vector corresponding to the target customer and the feature vector of the target product. Specifically, the server directly calls the similarity between the customer mapping vector and the target product mapping vector calculated in step S402 as the matching degree. This matching degree is the highest similarity value determined in S402, which is used for subsequent loss function calculation.

[0058] As an example, in step S602, the server uses the matching degree as an estimate of the fit between the customer demand predicted by the model and the product configuration; the market feedback data is converted into normalized label values. Specifically, the server obtains market feedback data of target customers on the target product, extracts customer behavior feedback indicators and product revenue indicators. Customer behavior feedback indicators include whether to purchase, whether to use, frequency of use, satisfaction rating, etc., and product revenue indicators include premium income, claims cost, treatment income, drug cost, etc. Each indicator is quantitatively scored according to business rules, such as purchase being scored as high score, no purchase as low score, high premium income as high score, and high claims cost as low score. The customer behavior feedback indicator scores are weighted and summed to obtain the market reaction score, and the product revenue indicator scores are weighted and summed to obtain the profitability indicator score. The market reaction score and the profitability indicator score are weighted and combined according to preset weights to obtain a comprehensive score. The comprehensive score is normalized and mapped to the [0,1] interval to obtain normalized label values. The normalized label value The comprehensive indicator reflects the market response to actual customer behavior and the profit indicators corresponding to actual product revenue, maintaining consistency with the market response and profit indicator framework of the comprehensive reward function in step S303. A preset loss function is used to calculate the deviation between the predicted estimate and the normalized label value, yielding the loss parameter. The preset loss function includes, but is not limited to, the mean squared error loss function and the cross-entropy loss function, which can be flexibly selected by those skilled in the art according to actual needs.

[0059] Among them, the gradient update algorithm refers to the optimization algorithm that adjusts the model parameters according to the gradient direction of the loss function, in order to reduce the model prediction error.

[0060] As an example, in step S603, the server updates the model parameters of the reinforcement learning model based on the loss parameter and a gradient update algorithm. The server calculates the gradient of the loss function with respect to the parameters of each layer of the policy network, propagates the gradient from the output layer to the input layer using backpropagation, and updates the network weights along the reverse gradient direction using gradient descent. Iterative training continues until the loss converges, resulting in the updated reinforcement learning model. The update formula is: in, This represents the updated parameters of the deep Q-network model. This represents the parameters of the current deep Q-network model. The learning rate controls the step size for each parameter update. Represents the loss function For parameters gradient, This represents the matching degree between the potential feature vector corresponding to the target customer and the feature vector of the target product. It is the degree of fit between the customer needs predicted by the model and the product configuration. The value range is [0,1]. The higher the value, the better the customer needs and product features match. This represents the normalized label value obtained from market feedback data, which comprehensively reflects the market reaction corresponding to the actual customer behavior feedback and the profit indicator corresponding to the actual product revenue. It is consistent with the market reaction and profit indicator framework of the comprehensive reward function in step S503, and the value range is [0,1]. The higher the value, the more positive the actual market reaction of the customer and the better the product's profit performance.

[0061] For example, in an insurance recommendation scenario, the server uses the target customer's matching degree as a predicted value for their willingness to purchase insurance. It converts the customer's actual insurance behavior and product premium revenue data into normalized label values ​​and calculates the mean squared error loss. Based on the loss calculation gradient, the server updates the weight parameters in the strategy network corresponding to the health concern group and critical illness insurance features, making subsequent recommendations for this group more accurate. In a medical recommendation scenario, the server uses the target patient's matching degree as a predicted value for medication adherence. It converts the patient's actual medication adherence and treatment plan benefits data into normalized label values ​​and calculates the mean squared error loss. Based on the loss calculation gradient, the server updates the weight parameters in the strategy network corresponding to the chronic disease management group and antihypertensive drug regimen features, making subsequent recommendations for this group more accurate.

[0062] In the AI-based product recommendation method provided in this embodiment, the matching degree between the latent feature vector and the target product feature vector is obtained, a loss function is constructed by combining real market feedback data, and the model parameters are optimized using the gradient descent algorithm. This enables the reinforcement learning model to continuously correct prediction bias based on real customer behavior. Through backpropagation and iterative training, the model strategy is deeply aligned with the real market environment, thereby improving the long-term accuracy of product recommendations and customer satisfaction.

[0063] As can be seen, in the above solution, for complex recommendation scenarios such as financial digital products and medical health management, the target customer segmentation needs group is first obtained through feature extraction and clustering, realizing accurate segmentation and differentiated positioning of customer needs. Then, a solution based on reinforcement learning to dynamically optimize product features is proposed. The reinforcement learning model analyzes and processes the target segmentation needs group and the original product feature vector to generate a suitable recommended product feature vector. Then, the target product feature vector is determined through matching and the corresponding product is recommended. This makes the product configuration scheme that better meets customer needs generated and recommended. It can effectively break through the limitations of traditional static rule recommendation, greatly improve the dynamic matching effect between products and customer needs, and improve the accuracy of personalized recommendations and user experience.

[0064] 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.

[0065] In one embodiment, an AI-based product recommendation device is provided, which corresponds one-to-one with the AI-based product recommendation method described in the above embodiments. For example... Figure 7 As shown, the AI-based product recommendation device includes a customer feature extraction module 71, a segmented demand determination module 72, a recommended product determination module 73, a target product determination module 74, and a product recommendation module 75.

[0066] The customer feature extraction module 71 is used to extract features from the customer data corresponding to the target customer and determine the potential feature vector corresponding to the target customer. The segmented demand determination module 72 is used to cluster the potential feature vectors corresponding to multiple target customers to obtain at least one cluster, and determine the target segmented demand group to which each target customer belongs, wherein the target segmented demand group corresponds to a cluster. The recommended product determination module 73 is used to analyze the target segment demand group and N original product feature vectors to which each target customer belongs using a reinforcement learning model, and determine M recommended product feature vectors corresponding to each target customer, where N and M are both integers greater than 1; The target product determination module 74 is used to perform matching processing on the potential feature vector and M recommended product feature vectors corresponding to each target customer, and determine the recommended product feature vector with the highest matching degree as the target product feature vector. Product recommendation module 75 is used to recommend the target product corresponding to the target product feature vector to the target customer.

[0067] In one embodiment, the customer feature extraction module 71 includes: The original feature extraction unit is used to extract features from the customer data corresponding to the target customer and determine the original feature vector corresponding to the target customer. The latent feature extraction unit is used to perform feature dimensionality reduction on the original feature vector corresponding to the target customer to determine the latent feature vector corresponding to the target customer.

[0068] In one embodiment, the segmentation requirement determination module 72 includes: The clustering grouping unit is used to cluster the potential feature vectors of multiple target customers using a clustering algorithm to obtain at least one original segmented demand group and to determine the central feature vector of each original segmented demand group. The attribution determination unit is used to match the potential feature vector of the target customer with the central feature vector of each original segmented demand group to determine the target segmented demand group to which each target customer belongs.

[0069] In one embodiment, the recommended product determination module 73 includes: A state space construction unit is used to acquire market environment data, extract features from the market environment data to obtain an environment vector, and use the target segment demand group to which the target customer belongs and the environment vector as the state vector. The strategy search optimization unit is used to construct a product feature vector set and select a product feature vector to be optimized from the product feature vector set as the action space. The product feature vector set initially includes all original product feature vector sets. The recommendation result generation unit is used to input the state vector and the action space into the reinforcement learning model for learning, obtain the output value of the comprehensive reward function in the reinforcement learning model, and determine the output value as the reward score of the product feature vector to be optimized; The strategy optimization iteration unit is used to optimize the product feature vector to be optimized based on the reward score, obtain the optimized product feature vector, write the optimized product feature vector into the product feature vector set, and return to execute the step of selecting a product feature vector to be optimized from the product feature vector set as the action space until the reward score reaches the optimization target. The optimal recommendation filtering unit is used to output M feature vectors of recommended products that enable the reward score to reach the optimization objective.

[0070] In one embodiment, the target product determination module 74 includes: The feature mapping alignment unit is used to map the potential feature vector of the target customer and the feature vectors of the M recommended products to the same feature space, and to determine the customer mapping vector corresponding to the target customer and the product mapping vectors corresponding to the feature vectors of the M recommended products respectively. The matching degree optimization unit determines the matching degree between the target customer and each recommended product based on the customer mapping vector and each product mapping vector, and determines the feature vector of the recommended product corresponding to the product mapping vector with the highest matching degree as the feature vector of the target product.

[0071] In one embodiment, the AI-based product recommendation device further includes: The feedback collection module is used to acquire market feedback data from the target customers regarding the target product. The model update module is used to update the model parameters of the reinforcement learning model based on the latent feature vector corresponding to the target customer, the feature vector of the target product, and the market feedback data, so as to obtain an updated reinforcement learning model.

[0072] In one embodiment, the model update module includes: The matching degree acquisition unit is used to acquire the matching degree between the potential feature vector corresponding to the target customer and the feature vector of the target product; The loss calculation unit is used to calculate the matching degree and the market feedback data based on a preset loss function to obtain loss parameters; The parameter update unit is used to update the model parameters of the reinforcement learning model according to the loss parameter and in combination with the gradient update algorithm, so as to obtain the updated reinforcement learning model.

[0073] This invention provides an artificial intelligence-based product recommendation device. First, it determines the target customer's specific needs group through feature extraction and clustering, achieving a refined segmentation of customer needs. Then, it proposes a reinforcement learning-based scheme for dynamically optimizing product recommendations. A reinforcement learning model generates multiple recommended product feature vectors that match the target specific needs group. Next, it matches the latent feature vector corresponding to the target customer with the multiple recommended product feature vectors, selecting products that are more suitable for the target customer. This effectively overcomes the limitations of traditional static rule-based recommendations, achieving continuous and automatic optimization of product features, greatly improving the effectiveness of personalized recommendations, and enhancing the accuracy and user experience of product recommendations.

[0074] Specific limitations regarding the AI-based product recommendation device can be found in the limitations of the AI-based product recommendation method described above, and will not be repeated here. Each module in the aforementioned AI-based product recommendation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0075] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side product recommendation method based on artificial intelligence.

[0076] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps of an artificial intelligence-based product recommendation method.

[0077] In one embodiment, 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 perform the following steps: Feature extraction is performed on the customer data corresponding to the target customer to determine the potential feature vector corresponding to the target customer; Clustering is performed on the potential feature vectors corresponding to multiple target customers to obtain at least one cluster, and the target segment demand group to which each target customer belongs is determined, wherein the target segment demand group corresponds to a cluster. A reinforcement learning model is used to analyze the target segment demand group and N original product feature vectors to which each target customer belongs, and to determine M recommended product feature vectors corresponding to each target customer, where N and M are both integers greater than 1; For each target customer, the potential feature vector and M recommended product feature vectors are matched, and the recommended product feature vector with the highest matching degree is determined as the target product feature vector. The target product corresponding to the target product feature vector is recommended to the target customer.

[0078] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, performs the following steps: Feature extraction is performed on the customer data corresponding to the target customer to determine the potential feature vector corresponding to the target customer; Clustering is performed on the potential feature vectors corresponding to multiple target customers to obtain at least one cluster, and the target segment demand group to which each target customer belongs is determined, wherein the target segment demand group corresponds to a cluster. A reinforcement learning model is used to analyze the target segment demand group and N original product feature vectors to which each target customer belongs, and to determine M recommended product feature vectors corresponding to each target customer, where N and M are both integers greater than 1; For each target customer, the potential feature vector and M recommended product feature vectors are matched, and the recommended product feature vector with the highest matching degree is determined as the target product feature vector. The target product corresponding to the target product feature vector is recommended to the target customer.

[0079] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0080] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and RAMbus direct RAM. (RDRAM), Direct Memory Bus Dynamic RAM (DRDRAM), and Memory Bus Dynamic RAM (RDRAM), etc.

[0081] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0082] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0083] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A product recommendation method based on artificial intelligence, characterized in that, include: Feature extraction is performed on the customer data corresponding to the target customer to determine the potential feature vector corresponding to the target customer; Clustering is performed on the potential feature vectors corresponding to multiple target customers to obtain at least one cluster, and the target segment demand group to which each target customer belongs is determined, wherein the target segment demand group corresponds to a cluster. A reinforcement learning model is used to analyze the target segment demand group and N original product feature vectors to which each target customer belongs, and to determine M recommended product feature vectors corresponding to each target customer, where N and M are both integers greater than 1; For each target customer, the potential feature vector and M recommended product feature vectors are matched, and the recommended product feature vector with the highest matching degree is determined as the target product feature vector. The target product corresponding to the target product feature vector is recommended to the target customer.

2. The product recommendation method based on artificial intelligence as described in claim 1, characterized in that, The step of extracting features from customer data corresponding to the target customer to determine the potential feature vector corresponding to the target customer includes: Extract features from the customer data corresponding to the target customer to determine the original feature vector corresponding to the target customer; The original feature vectors corresponding to the target customers are subjected to feature dimensionality reduction to determine the potential feature vectors corresponding to the target customers.

3. The product recommendation method based on artificial intelligence as described in claim 1, characterized in that, The process of clustering the potential feature vectors corresponding to multiple target customers to obtain at least one cluster, and determining the target segment demand group to which each target customer belongs, wherein the target segment demand group corresponds to a cluster, includes: Clustering algorithms are used to cluster the potential feature vectors of multiple target customers to obtain at least one original segmented demand group, and the central feature vector of each original segmented demand group is determined. The potential feature vectors of the target customers are matched with the central feature vectors of each of the original segmented demand groups to determine the target segmented demand group to which each target customer belongs.

4. The product recommendation method based on artificial intelligence as described in claim 1, characterized in that, The reinforcement learning model is configured with a comprehensive reward function. The reinforcement learning model analyzes the target customer's target segment demand group and N original product feature vectors to determine M recommended product feature vectors corresponding to each target customer, including: Acquire market environment data, extract features from the market environment data to obtain an environment vector; use the target segment demand group to which the target customer belongs and the environment vector as a state vector; Construct a product feature vector set, and select a product feature vector to be optimized from the product feature vector set as the action space. The product feature vector set initially includes all original product feature vector sets. The state vector and the action space are input into the reinforcement learning model for learning, and the output value of the comprehensive reward function in the reinforcement learning model is obtained. The output value is determined to be the reward score of the product feature vector to be optimized. Based on the reward score, the product feature vector to be optimized is optimized to obtain an optimized product feature vector. The optimized product feature vector is written into the product feature vector set. Then, the process of selecting a product feature vector to be optimized from the product feature vector set as the action space is returned until the reward score reaches the optimization target. Output M feature vectors of recommended products that enable the reward score to reach the optimization objective.

5. The product recommendation method based on artificial intelligence as described in claim 1, characterized in that, The step of matching the potential feature vector corresponding to each target customer with the M recommended product feature vectors to determine the recommended product feature vector with the highest matching degree as the target product feature vector includes: Map the potential feature vector of the target customer and the feature vectors of the M recommended products to the same feature space, and determine the customer mapping vector corresponding to the target customer and the product mapping vector corresponding to the feature vectors of the M recommended products respectively; Based on the customer mapping vector and each product mapping vector, the matching degree between the target customer and each recommended product is determined, and the feature vector of the recommended product corresponding to the product mapping vector with the highest matching degree is determined as the feature vector of the target product.

6. The product recommendation method based on artificial intelligence as described in any one of claims 1 to 5, characterized in that, Also includes: Obtain market feedback data from the target customers regarding the target product; Based on the potential feature vectors corresponding to the target customers, the feature vectors of the target products, and the market feedback data, the model parameters of the reinforcement learning model are updated to obtain an updated reinforcement learning model.

7. The product recommendation method based on artificial intelligence as described in claim 6, characterized in that, The step of updating the model parameters of the reinforcement learning model based on the latent feature vectors corresponding to the target customers, the feature vectors of the target products, and market feedback data to obtain an updated reinforcement learning model includes: Obtain the matching degree between the potential feature vector corresponding to the target customer and the feature vector of the target product; Based on a preset loss function, the matching degree and the market feedback data are calculated to obtain loss parameters; Based on the loss parameter, and combined with the gradient update algorithm, the model parameters of the reinforcement learning model are updated to obtain the updated reinforcement learning model.

8. A product recommendation device based on artificial intelligence, characterized in that, include: The customer feature extraction module is used to extract features from the customer data corresponding to the target customer and determine the potential feature vector corresponding to the target customer. The segmented demand determination module is used to cluster the potential feature vectors corresponding to multiple target customers to obtain at least one cluster, and to determine the target segmented demand group to which each target customer belongs, wherein the target segmented demand group corresponds to a cluster. The recommended product determination module is used to analyze the target segment demand group and N original product feature vectors to which each target customer belongs using a reinforcement learning model, and determine M recommended product feature vectors corresponding to each target customer, where N and M are both integers greater than 1; The target product determination module is used to match the potential feature vector and M recommended product feature vectors corresponding to each target customer, and determine the recommended product feature vector with the highest matching degree as the target product feature vector. The product recommendation module is used to recommend the target product corresponding to the target product feature vector to the target customer.

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 artificial intelligence-based 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 artificial intelligence-based product recommendation method as described in any one of claims 1 to 7.

Images