Product recommendation method and device based on sample optimization, equipment and storage medium

Abstract

The embodiment of the application provides a product recommendation method and device based on sample optimization, equipment and a storage medium, belonging to the technical field of artificial intelligence, and applied to a financial scene. The method comprises the following steps: performing product performance evaluation according to a candidate interaction behavior and a candidate user portrait, so as to select a target sample in the candidate sample according to a product performance score; performing product recommendation on a candidate user through an original product recommendation model and the target sample, updating the target sample according to performance change data after recommendation, training the original product recommendation model according to the updated sample, and performing product recommendation on a target user through a target product recommendation model and a target interaction behavior. According to the embodiment of the application, sample selection is performed according to product performance evaluation data, a product recommendation model is trained, and product recommendation is realized through the trained target product recommendation model, so that personalized products that are matched with the real needs of users and have high product conversion performance can be recommended to the users, and the product recommendation accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and is applicable to the financial field, particularly to a product recommendation method, apparatus, device, and storage medium based on sample optimization. Background Technology

[0002] Currently, traditional product recommendation methods primarily rely on pre-trained neural network models (such as LSTM, Long Short-Term Memory) and analysis of user behavior and static user profiles. For example, in financial scenarios, an LSTM model can be trained using randomly selected user training samples. This pre-trained LSTM can then analyze users' temporal behavioral data (such as browsing, clicking on critical illness insurance products, or seeking information about "diseases") and identify user profiles containing the tag "purchase health insurance," thus recommending critical illness insurance products. However, this method lacks specificity in selecting training samples for the neural network model, typically employing random sampling. This leads to the trained model neglecting the actual needs of different customer groups for different products, resulting in mismatches between recommended products and actual user needs, thus reducing accuracy. Furthermore, it fails to consider the varying impacts of different user groups on product performance (such as conversion rates), leading to low conversion efficiency. Therefore, improving the accuracy of product recommendations to increase conversion rates has become a pressing issue. Summary of the Invention

[0003] To address the technical problem of improving product conversion rates by enhancing the accuracy of product recommendations, the main objective of this application is to propose a product recommendation method, apparatus, device, and storage medium based on sample optimization, aiming to improve the accuracy of product recommendations and further enhance product conversion rates.

[0004] To achieve the above objectives, a first aspect of this application proposes a product recommendation method based on sample optimization, the method comprising:

[0005] Obtain candidate samples of candidate users; wherein, the candidate samples include candidate interaction behaviors of candidate users and candidate user profiles constructed based on the candidate interaction behaviors; wherein, the candidate interaction behaviors are the interaction operations of the candidate users on preset products; Product performance is evaluated based on the candidate interaction behaviors and the candidate user profiles to obtain candidate product performance scores. Based on the candidate product performance scores, the candidate samples are selected to obtain the target samples; By using a preset original product recommendation model and the target sample, product recommendations are made to the candidate users to obtain data on changes in the effectiveness of the recommended products. The target sample is updated based on the recommended product performance change data to obtain an updated sample; The original product recommendation model is trained based on the updated samples to obtain the target product recommendation model. The target user's target interaction behavior is obtained, and product recommendations are made to the target user based on the target product recommendation model and the target interaction behavior.

[0006] In some embodiments, the step of evaluating product performance based on the candidate interaction behaviors and the candidate user profiles to obtain candidate product performance scores includes: Activity level is obtained by performing activity detection based on the candidate interaction behaviors; Based on the candidate user profiles, a sample risk assessment is performed to obtain the user risk level; Based on the candidate user profile, sample period detection is performed to obtain the user period stage; The product performance score is obtained by scoring the candidate product based on the behavioral activity level, the user risk level, and the user cycle stage.

[0007] In some embodiments, updating the target sample based on the recommended product performance change data to obtain an updated sample includes: Based on the recommended product performance change data, construct a product performance change function; An initial sample sampling strategy is obtained, and the initial sample sampling strategy is optimized according to the product performance change function to obtain an optimized sample sampling strategy. The target sample is updated by sampling according to the optimized sample sampling strategy to obtain the updated sample.

[0008] In some embodiments, the updated sample includes real product performance change data and real recommended products; The step of training the original product recommendation model based on the updated samples to obtain the target product recommendation model includes: Based on the original product recommendation model, product recommendation prediction is performed on the updated sample to obtain the trained predicted recommended products; Obtain updated candidate users from the updated samples, and recommend the trained prediction recommendation products to the target candidate users to obtain data on the change in the effectiveness of the predicted products; The target loss value is obtained by calculating the loss based on the trained predicted recommended products, the predicted product performance change data, the actual recommended products, the actual product performance change data, and the optimized sample sampling strategy. The original product recommendation model is trained based on the target loss value to obtain the target product recommendation model.

[0009] In some embodiments, the candidate user profile constructed based on the candidate interaction behaviors includes: Obtain candidate user information of the candidate users, and extract features from the candidate user information to obtain candidate user information features; Feature extraction is performed on the candidate interaction behaviors to obtain candidate interaction behavior features; The candidate user information features and the candidate interaction behavior features are weighted and fused to obtain user profile features; Based on the user profile features, candidate user attribute labels are generated; The candidate user profile is generated based on the candidate user attribute tags.

[0010] In some embodiments, the candidate interaction behaviors include product click behavior, product dwell behavior, and product conversion behavior; The step of extracting features from the candidate interaction behaviors to obtain candidate interaction behavior features includes: Click frequency analysis is performed on the product click behavior to obtain click behavior characteristics; Dwell time sequence analysis is performed on the product dwell behavior to obtain dwell behavior characteristics; The conversion effectiveness of the product conversion behavior is analyzed to obtain the characteristics of the conversion effectiveness behavior; The click behavior feature, the dwell behavior feature, and the conversion effect behavior feature are weighted to obtain the candidate interaction behavior feature.

[0011] In some embodiments, recommending products to the target user based on the target product recommendation model and the target interaction behavior includes: Obtain the target user information of the target user, and construct a target user profile based on the target user information and the target interaction behavior; Based on the target product recommendation model and the target user profile, product preference prediction is performed to obtain product preference data; Based on the product preference data, product demand is inferred to obtain the product demand probability; Based on the product demand probability, a product recommendation prediction is made to obtain the target predicted recommended product, and the target predicted recommended product is pushed to the target user.

[0012] To achieve the above objectives, a second aspect of this application provides a product recommendation apparatus based on sample optimization, the apparatus comprising: A candidate sample acquisition module is used to acquire candidate samples of candidate users; wherein, the candidate samples include candidate interaction behaviors of candidate users and candidate user profiles constructed based on the candidate interaction behaviors; wherein, the candidate interaction behaviors are the interaction operations of the candidate users on preset products; The product performance evaluation module is used to evaluate product performance based on the candidate interaction behavior and the candidate user profile, and obtain the candidate product performance score. The sample selection module is used to select target samples from the candidate samples based on the candidate product performance scores. The training product recommendation module is used to recommend products to the candidate users using a preset original product recommendation model and the target sample, and to obtain data on the change in the effectiveness of the recommended products. The sample update module is used to update the target sample based on the recommended product performance change data to obtain an updated sample; The model training module is used to train the original product recommendation model based on the updated samples to obtain the target product recommendation model. The target product recommendation module is used to acquire the target user's target interaction behavior and recommend products to the target user based on the target product recommendation model and the target interaction behavior.

[0013] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0014] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of the first aspect described above.

[0015] This application proposes a sample-optimized product recommendation method, apparatus, device, and storage medium. First, it acquires candidate samples of candidate users, including candidate interaction behaviors and user profiles constructed based on these behaviors. Then, it evaluates the product effectiveness of the candidate interaction behaviors and user profiles to obtain candidate product effectiveness scores, accurately identifying the true needs and conversion rates of different user groups. Second, it selects candidate samples based on their effectiveness scores and updates target samples according to changes in recommended product effectiveness data. This allows for targeted selection of user groups that contribute significantly to product conversion, avoiding the need mismatch problem caused by random sample selection in traditional methods. Furthermore, it trains the original product recommendation model using the target samples to obtain a target product recommendation model. This ensures that the model's learning process is directly aligned with the product conversion effectiveness target, improving the model's accuracy in identifying product needs and recommending products to different users. Finally, it recommends products to target users using the target product recommendation model and target interaction behaviors, outputting personalized product solutions that match users' true needs and have high conversion effectiveness, significantly improving the accuracy of product recommendations and product conversion rates. Attached Figure Description

[0016] Figure 1 This is a flowchart of the product recommendation method based on sample optimization provided in the embodiments of this application; Figure 2 yes Figure 1 The flowchart of step S101 in the text; Figure 3 yes Figure 2 The flowchart of step S202 in the document; Figure 4 yes Figure 1 The flowchart of step S102 in the document; Figure 5 yes Figure 1 The flowchart of step S105 in the process; Figure 6 yes Figure 1 The flowchart of step S106 in the process; Figure 7 yes Figure 1 The flowchart of step S107 in the process; Figure 8 This is a schematic diagram of the product recommendation device based on sample optimization provided in the embodiments of this application; Figure 9 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0020] First, let's analyze some of the terms used in this application: Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0021] This application provides a product recommendation method, apparatus, device, and storage medium based on sample optimization, aiming to improve the accuracy of product recommendations and further increase product conversion rates.

[0022] The product recommendation method, apparatus, device, and storage medium based on sample optimization provided in this application are specifically described through the following embodiments. First, the product recommendation method based on sample optimization in this application is described.

[0023] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0024] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0025] The sample-optimized product recommendation method provided in this application relates to the field of artificial intelligence technology. This sample-optimized product recommendation method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the sample-optimized product recommendation method, but is not limited to the above forms.

[0026] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions performed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can be located on local and remote computer storage media, including storage devices.

[0027] Figure 1This is an optional flowchart of the product recommendation method based on sample optimization provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S107.

[0028] Step S101: Obtain candidate samples of candidate users; wherein, the candidate samples include candidate interaction behaviors of candidate users and candidate user profiles constructed based on the candidate interaction behaviors; wherein, the candidate interaction behaviors are the interaction operations of candidate users on preset products.

[0029] Step S102: Evaluate product performance based on candidate interaction behaviors and candidate user profiles to obtain candidate product performance scores.

[0030] Step S103: Select candidate samples based on candidate product performance scores to obtain target samples.

[0031] Step S104: Recommend products to candidate users using a preset original product recommendation model and target samples to obtain data on changes in the effectiveness of recommended products.

[0032] Step S105: Update the target sample based on the recommended product performance change data to obtain the updated sample.

[0033] Step S106: Train the original product recommendation model based on the updated samples to obtain the target product recommendation model.

[0034] Step S107: Obtain the target user's target interaction behavior, and recommend products to the target user based on the target product recommendation model and the target interaction behavior.

[0035] Steps S101 to S107 of this embodiment first involve acquiring candidate samples of candidate users, including candidate interaction behaviors and user profiles built based on these interactions. Product performance is then evaluated on the candidate interaction behaviors and user profiles to obtain candidate product performance scores, accurately identifying the real needs and conversion rates of different user groups. Second, candidate samples are selected based on their performance scores, and target samples are updated based on changes in recommended product performance data. This allows for targeted selection of user groups that significantly contribute to product conversion, avoiding the misalignment of needs caused by random sample selection in traditional methods. Furthermore, the original product recommendation model is trained using the target samples to obtain a target product recommendation model. This ensures that the model's learning process is directly aligned with the product conversion performance target, improving the model's accuracy in identifying product needs from different users and enhancing product recommendation accuracy. Finally, product recommendations are made to target users using the target product recommendation model and target interaction behaviors, outputting personalized product solutions that match users' real needs and have high conversion performance, significantly improving the accuracy of product recommendations and product conversion rates.

[0036] In step S101 of some embodiments, specifically, the candidate sample refers to a set of user data related to the business, used to train the original product recommendation model for product recommendation. The candidate sample includes candidate user interaction behaviors and candidate user profiles constructed based on the candidate interaction behaviors. The candidate interaction behaviors are the interaction operations of candidate users on preset products, which may include product click behavior, product dwell behavior, and product conversion behavior, etc. The candidate user profile is a set of tags that characterize the attributes, preferences, and potential needs of candidate users based on the candidate interaction behaviors.

[0037] Specifically, candidate samples of potential users can be extracted from the interaction log database of the business platform, the interaction log data of the business APP, and the customer relationship management system, according to preset time windows (such as from 2 days ago to the present) and user filtering conditions (such as non-blacklisted users, registered users, and users who have accessed product pages).

[0038] For example, in an insurance scenario, candidate samples can be registered users of an insurance platform. A user's product click behavior can be identified and extracted from user interaction logs on the insurance platform, including click events related to insurance products such as clicking on the product details page, clicking on the product plan introduction page, or clicking the customer service consultation button. Product dwell behavior can be identified and extracted from user interaction logs, including dwell events on the insurance product page, such as product page browsing time, product page scrolling depth, and interaction duration. Product conversion behavior can be identified and extracted from user interaction logs, including conversion events related to insurance products such as placing an order, renewing a policy, completing a consultation, or purchasing insurance.

[0039] Furthermore, in the insurance scenario, if a candidate user clicks on the health insurance product page multiple times on the insurance platform, reviews the terms of critical illness insurance in detail, spends a considerable amount of time on the health insurance page, and performs interactive operations such as purchasing critical illness insurance, it indicates that the user has a high interest in health insurance. In this case, a candidate user profile can be constructed that includes tags such as "purchasing critical illness insurance," "interested in health insurance," "paying attention to critical illness insurance terms," ​​and "high awareness of health risks."

[0040] In this embodiment, by acquiring candidate user interaction behaviors, including candidate interaction actions on preset products, and candidate samples of user profiles built based on interaction behaviors, comprehensive user interaction behavior and profile data can be collected, providing high-quality data support for subsequent model training and product recommendation.

[0041] Please see Figure 2 In some embodiments, step S101 may include, but is not limited to, steps S201 to S205: Step S201: Obtain candidate user information of candidate users, and extract features from candidate user information to obtain candidate user information features.

[0042] Step S202: Extract features from candidate interaction behaviors to obtain candidate interaction behavior features.

[0043] Step S203: The candidate user information features and candidate interaction behavior features are weighted and fused to obtain user profile features.

[0044] Step S204: Generate candidate user attribute labels based on user profile features.

[0045] Step S205: Generate candidate user profiles based on candidate user attribute tags.

[0046] In step S201 of some embodiments, specifically, the candidate user information refers to the static attribute information of the candidate user, which may include the candidate user's age, gender, and number of insurance policies, etc.

[0047] Specifically, candidate user information features are user attribute vector representations extracted from candidate user information.

[0048] Specifically, word embedding can be performed on candidate user information using a user information encoder (such as a Transformer network) to capture the features of the candidate user information.

[0049] Please see Figure 3 In some embodiments, candidate interaction behaviors include product click behavior, product dwell behavior, and product conversion behavior, and step S202 includes, but is not limited to, steps S301 to S304: Step S301: Analyze the click frequency of product click behavior to obtain click behavior characteristics.

[0050] Step S303: Perform dwell time sequence analysis on product dwell behavior to obtain dwell behavior characteristics.

[0051] Step S303: Analyze the conversion effectiveness of product conversion behavior to obtain the characteristics of conversion effectiveness behavior.

[0052] Step S304: Weight the click behavior features, dwell behavior features, and conversion effect behavior features to obtain candidate interaction behavior features.

[0053] In step S301 of some embodiments, specifically, the click behavior feature is a click frequency statistical vector representation extracted from the product click behavior.

[0054] Specifically, product click behavior event sequences can be identified from candidate interaction behaviors. Behavioral encoders (such as RNNs, Recurrent Neural Networks) can be used to count the click frequency of event sequences based on dimensions such as product category, interaction channel (such as APP, web page, mini program), and time period (such as the last 48 hours, 7 days, etc.) and generate structured vectors.

[0055] For example, in an insurance scenario, if a candidate user clicks on the "Medical Insurance" product page 25 times (18 times on the APP channel and 7 times on the web channel) and clicks on the "Car Insurance" page 12 times (only on the APP channel) in the past 7 days, then the click behavior feature can be represented as [25,18,7,12], where each vector element corresponds to the frequency statistics value of different dimensions.

[0056] In step S302 of some embodiments, specifically, the dwell behavior feature is a dwell time sequence vector representation extracted from the product dwell behavior.

[0057] Specifically, for time-series data on product dwell behavior (such as page dwell time, scrolling depth, and frequency of interactive operations), the RNN network can be used to capture the time-series vector of dwell event sequences based on dimensions such as dwell time range (such as 0 to 10 seconds, 10 to 30 seconds, and more than 30 seconds), dwell page depth (such as the first screen, terms page, and insurance page), and time period (such as the last 48 hours).

[0058] For example, in an insurance scenario, if a candidate user spends 180 seconds on the details page of "Critical Illness Insurance Product A" (deeply browsing to the "Disclaimer" section), 20 seconds on the overview page of "Critical Illness Insurance Product B", and finally returns to the "Insurance Notice" page of "Critical Illness Insurance Product A" and spends 90 seconds there, then the user's dwell behavior characteristics can be [180, 20, 90], where each element in the vector represents the corresponding dwell time.

[0059] In step S303 of some embodiments, specifically, the conversion effectiveness behavior feature is a conversion behavior vector representation extracted from product conversion behavior.

[0060] Specifically, RNN networks can be used to capture behavioral characteristics of conversion effectiveness based on product conversion behaviors (such as placing an order, renewing, completing a consultation, and successfully purchasing insurance), according to dimensions such as conversion type (e.g., first-time insurance purchase, renewal, consultation) and contribution of conversion business (e.g., policy amount, business cycle value).

[0061] For example, in an insurance scenario, if a candidate user completes the "car insurance renewal" conversion this month (policy amount of 5,000 yuan, business value weight of 2.0) and the "first health insurance purchase" conversion last month (policy amount of 2,000 yuan, business value weight of 1.0), then the conversion effectiveness behavioral characteristics can be represented as [10000, 2000], where the vector elements correspond to the quantitative values ​​of the conversion effectiveness of different conversion events.

[0062] In step S304 of some embodiments, specifically, the candidate interaction behavior features are behavior vector representations that integrate click behavior features, dwell behavior features, and conversion effectiveness behavior features.

[0063] Specifically, for each behavioral feature, a time decay function can be applied to calculate the decay weight of each behavioral feature based on the timestamp of the behavior, and then the click behavior feature, dwell behavior feature and conversion effect behavior feature can be weighted and fused according to the weight to obtain candidate interaction behavior features.

[0064] Furthermore, the behavioral feature decay weights can be calculated using the following formula:

[0065] in, Indicates the attenuation weight of behavioral features. This represents the attenuation coefficient, which is a constant greater than 0, and The larger the value, the greater the effect of time decay. Indicates the current time. Indicates the time when the action occurred.

[0066] Steps S301 to S304 reflect users' recent preferences and behavioral patterns, providing dynamic interactive behavior data support for subsequent user profile construction and product recommendation model training, which helps improve the accuracy of subsequent product recommendations.

[0067] In step S203 of some embodiments, specifically, the user profile feature refers to the profile vector representation that integrates candidate user information features and candidate interaction behavior features.

[0068] Specifically, the weighted fusion process can be performed using the following formula:

[0069] in, This represents the user profile features of the i-th candidate user. This represents the fusion weight, used to balance candidate user information features and candidate interaction behavior features. Indicates user information encoder, This represents the information features of the i-th candidate user. Indicates a behavior encoder. This represents the candidate interaction behavior characteristics of the i-th candidate user.

[0070] Furthermore, the fusion weights can be optimized through comparative learning. This makes the vector distance between user profiles of the same type closer (i.e., positive sample pairs) and the vector distance between user profiles of different types farther (i.e., negative sample pairs).

[0071] Specifically, comparative learning can be represented by the following formula:

[0072] in, This represents the contrastive learning loss function; This represents the user profile features of the i-th candidate user; The positive sample represents the user profile features of the j-th candidate user, which are similar to the user profile features of the i-th candidate user. Indicates the temperature coefficient; This represents the user profile features of the k-th candidate user, which can be either a positive or negative sample.

[0073] In step S204 of some embodiments, specifically, the candidate user attribute label refers to a set of labels for candidate users that have clear business meanings. These labels are used to reflect the candidate user's user preference tendencies and user risk level.

[0074] Specifically, the user profile features can be activated using the Softmax activation function to output the probability of each candidate user attribute label, and the final candidate user attribute label can be determined by setting a threshold (such as 0.5).

[0075] For example, in an insurance scenario, after the user profile features are activated, the probability of outputting each candidate user attribute label can be "high-net-worth user" (probability 0.85), "medical insurance preference user" (probability 0.92), "low-risk policy cancellation user" (probability 0.78), and "sensitive to car insurance renewal" (probability 0.78). If the preset threshold is 0.5, the final generated candidate user attribute labels will be high-net-worth user, medical insurance preference user, and low-risk policy cancellation user.

[0076] In step S205 of some embodiments, specifically, the candidate user profile refers to the visualized entity data after associating the candidate user attribute tags and the corresponding candidate user identification information.

[0077] Specifically, candidate user attribute tags and user profile features can be visualized to obtain a user profile with candidate user identification information (such as the unique user identifier UserID) as the primary key, and containing at least the following fields: user profile features, candidate user identification information, candidate user attribute tags, and profile generation timestamp.

[0078] Steps S201 to S205 integrate the static attributes and dynamic behaviors of candidate users to form personalized candidate user profiles, enabling a better understanding of user behavior preferences and product needs. This helps improve the accuracy of subsequent product recommendations, thereby increasing product conversion rates.

[0079] Please see Figure 4 In some embodiments, step S102 includes, but is not limited to, steps S401 to S404: Step S401: Perform activity detection based on candidate interaction behaviors to obtain behavior activity.

[0080] Step S402: Conduct a sample risk assessment based on the candidate user profile to obtain the user risk level.

[0081] Step S403: Perform sample period detection based on candidate user profiles to obtain user period stages.

[0082] Step S404: Based on behavioral activity, user risk level, and user lifecycle stage, product performance is scored to obtain candidate product performance scores.

[0083] In step S401 of some embodiments, specifically, behavioral activity refers to the frequency and intensity of candidate users' interaction with the product, which is used to reflect the candidate users' attention to and participation in the product.

[0084] Specifically, the candidate interaction behavior features are obtained, and the assigned behavior feature weights of the candidate interaction behavior features are determined according to business rules through a preset activity evaluation function. The candidate interaction behavior features are then weighted according to the behavior feature weights to obtain a behavior activity score, which is used to evaluate the behavior activity of candidate users.

[0085] For example, in an insurance scenario, for a candidate user who has recently clicked on health insurance products multiple times and spent a long time reading the terms but has not yet purchased insurance, their candidate interaction behavior feature vector includes normalized behavior dimension data (e.g., click behavior intensity is 0.8, dwell behavior intensity is 0.9, and conversion behavior intensity is 0.2). Using a pre-defined activity evaluation function, click behavior, dwell behavior, and conversion behavior are assigned weights of 0.5, 0.4, and 0.1 respectively, emphasizing the dominant role of proactive exploration behavior in activity evaluation. Then, the feature vector is weighted according to the assigned weights, resulting in a behavior activity score of 0.78.

[0086] In step S402 of some embodiments, specifically, the user risk level describes the potential risk situation of the candidate user.

[0087] Specifically, user profile features of candidate user profiles are obtained and input into a pre-trained risk assessment model (such as the XGBoost model) to perform sample risk assessment, output sample risk probability, and determine the degree of user risk based on the sample risk probability.

[0088] For example, in an insurance scenario, if a candidate user's profile characteristics indicate that the candidate user has frequently consulted about the policy cancellation process recently, and the candidate user has a history of complaints about their policies, then the XGBoost model can predict that the candidate user's risk of policy cancellation is 0.75.

[0089] In step S403 of some embodiments, specifically, the user lifecycle stage refers to the stage in the user's business lifecycle, which is used to reflect the candidate user's demand for the product and purchase intention.

[0090] Specifically, a periodic prediction model can be used to predict the stage of a candidate user based on the user profile characteristics of the candidate user profile, so as to output the expected value data of the candidate user and divide the life cycle stage of the user according to the preset expected value threshold.

[0091] For example, in an insurance scenario, if the expected value of a candidate user is less than the expected value threshold (e.g., 5,000 yuan), it indicates that the user is a potential customer in the lifecycle stage; if the expected value is within the expected value threshold range (e.g., 5,000 to 20,000 yuan), it indicates that the user is a growth customer in the lifecycle stage; if the expected value is greater than the expected value threshold (e.g., 20,000 yuan), it indicates that the user is a high-net-worth user in the lifecycle stage.

[0092] In step S404 of some embodiments, specifically, the product performance score is used to assess the contribution of candidate samples to expected business objectives by comprehensively evaluating behavioral activity, user risk level, and user lifecycle stage factors.

[0093] For example, in the insurance scenario, the expected business goal could be the conversion rate of insurance products.

[0094] Specifically, the performance score of candidate products can be calculated using the following formula:

[0095] in, This represents the candidate product performance score for the i-th candidate user. This represents the activity weighting parameter. The behavioral activity level representing candidate interaction behavior features. This represents the candidate interaction behavior features of the i-th candidate user. Represents the periodic weight parameter. This represents the user lifecycle stage of the user profile characteristics of the i-th candidate user. This represents the user profile features of the i-th candidate user. This represents the risk weight parameter. This represents the user risk level of the user profile features of the i-th candidate user.

[0096] Through steps S401 to S404, the product performance score corresponding to each candidate sample can be used to consider the differentiated impact of different users on the performance of business products, which helps to achieve differentiated sample selection in the future, thereby improving product conversion rate.

[0097] In step S103 of some embodiments, specifically, the target sample refers to the sample selected from the candidate samples that has a high contribution to the expected business objectives.

[0098] Specifically, the selection of candidate samples based on the candidate product performance scores to obtain target samples may include calculating the sample weights of candidate samples based on the product performance scores to obtain candidate sample weights; and selecting candidate samples based on the candidate sample weights to obtain target samples.

[0099] Specifically, the candidate sample weights can be calculated using the following formula:

[0100] in, This represents the candidate sample weight for the i-th candidate user. This represents the candidate product performance score for the i-th candidate user. This represents the candidate product performance score for the j-th candidate user.

[0101] For example, in an insurance scenario, after product performance evaluation, if a user sample with high recent activity, is in the high-net-worth user cycle segment, and has a low risk level, its candidate product performance score is 0.9, while a user sample with low activity and a high risk level has a candidate product performance score of 0.2, then a threshold method based on ranking is used to select all samples with product performance scores higher than 0.6 as target samples, that is, filtering out low-scoring samples of 0.2. The overall business product performance of the target samples selected in this way is much higher than that of the initial candidate sample set.

[0102] In step S104 of some embodiments, specifically, the original product recommendation model can be an untrained neural network model, such as a Transformer model.

[0103] Specifically, the data on changes in product effectiveness refers to the data on the recommendation effect after recommending products to target candidate users based on the target sample.

[0104] For example, in the insurance scenario, data on changes in the effectiveness of recommended products can include data on changes in insurance product conversion rates, changes in insurance product renewal rates, and changes in user lifecycle stages.

[0105] Specifically, the target candidate interaction behavior and target candidate user profile in the target sample can be input into the Transformer model to identify the recommended product, obtain the target recommended product, recommend the target recommended product to the target candidate user, and observe the product performance change data after the target candidate user receives the target recommended product recommendation information.

[0106] Please see Figure 5 In some embodiments, step S105 includes, but is not limited to, steps S501 to S503: Step S501: Construct a product performance change function based on the recommended product performance change data.

[0107] Step S502: Obtain the initial sample sampling strategy, and optimize the initial sample sampling strategy according to the product performance change function to obtain the optimized sample sampling strategy.

[0108] Step S503: Update the target sample according to the optimized sample sampling strategy to obtain the updated sample.

[0109] In step S501 of some embodiments, specifically, the product performance change function refers to a mapping function that quantifies the recommended product performance change data into a reinforcement learning reward signal.

[0110] Specifically, the product performance variation function can be determined using the following formula:

[0111] in, This represents the sample reward signal for the target sample n. Indicates the periodic change weight parameter. This represents data that changes throughout a user's lifecycle. The parameter representing the weighting of conversion rate changes This indicates data on changes in product conversion rates. This indicates the weighting parameter for changes in the renewal rate. This indicates data on changes in product renewal rates.

[0112] In step S502 of some embodiments, specifically, the initial sample sampling strategy refers to the basic sample sampling decision that has not undergone reinforcement learning reward feedback.

[0113] Specifically, optimizing the sample sampling strategy refers to the sample sampling decision optimized through reinforcement learning reward feedback.

[0114] Specifically, the expected return of the initial sample sampling strategy can be determined using the following formula:

[0115] in, Indicates the initial sample sampling strategy Expected returns; This represents the parameters of the reinforcement learning network; Indicates the initial sample sampling strategy; This indicates the sampling strategy used for initial samples. The expected value of the target sample n; This represents the sample reward signal for the target sample n.

[0116] Specifically, the policy can be optimized using the following policy gradient optimization formula, taking into account the expected return of the samples:

[0117] in, Indicates the sampling strategy based on the initial sample. The expected return corresponding reinforcement learning network parameters The gradient; Indicates the initial sample sampling strategy Expected returns; Indicates the initial sample sampling strategy; This represents the sample reward signal for the target sample n; This represents the probability that the target sample n is selected.

[0118] In step S503 of some embodiments, specifically, updating the sample refers to a sample with a higher reward expectation that is reselected from the target sample through an optimized sampling strategy.

[0119] Specifically, an optimized sample sampling strategy can be implemented. For each target sample in the target sample pool, the probability of the sample being selected is calculated through reinforcement learning. The sample is then randomly sampled according to this probability distribution, and the selected target sample constitutes the updated sample.

[0120] For example, in an insurance scenario, if the current target sample pool contains 10,000 registered users of an insurance platform, the application optimizes the sample sampling strategy to calculate the probability of a sample being selected from the target sample pool, and then extracts 3,000 samples from it as update samples based on the probability.

[0121] Through steps S501 to S503, the sample selection strategy can be optimized based on reinforcement learning to select user groups that contribute significantly to product conversion. This avoids the demand mismatch problem caused by the randomness of sample selection in traditional methods, thereby improving the accuracy of subsequent product recommendation models.

[0122] Please see Figure 6 In some embodiments, the updated sample includes real product performance change data and real recommended products, and step S106 also includes, but is not limited to, steps S601 to S604: Step S601: Based on the original product recommendation model, perform product recommendation prediction on the updated samples to obtain the trained predicted recommended products.

[0123] Step S602: Obtain updated candidate users of the updated sample, recommend the trained prediction recommendation product to the target candidate users, and obtain the predicted product performance change data.

[0124] Step S603: Calculate the loss based on the training prediction recommended products, the predicted product performance change data, the actual recommended products, the actual product performance change data, and the optimized sample sampling strategy to obtain the target loss value.

[0125] Step S604: Train the original product recommendation model based on the target loss value to obtain the target product recommendation model.

[0126] In step S601 of some embodiments, specifically, training the prediction and recommendation of products refers to the original product recommendation model using the product data recommended by the user corresponding to the update sample.

[0127] Specifically, the target candidate user profiles and behavioral features from the updated samples can be input into the original product recommendation model to generate one or more training prediction recommendation products for the target candidate users.

[0128] For example, in an insurance scenario, for a user in an updated sample who represents "middle-aged, has a family, and is currently focused on health management", the original product recommendation model may output predicted recommended products as "family medical insurance" and "group accident insurance".

[0129] In step S602 of some embodiments, specifically, the predicted product performance change data refers to the recommendation effect data after recommending products to updated candidate users based on the updated samples.

[0130] For example, in the insurance scenario, data predicting changes in product performance can include data predicting changes in insurance product conversion rates, insurance product renewal rates, and changes in user lifecycle stages.

[0131] Specifically, the method for determining the predicted product performance change data is the same as the method for recommending products to candidate users through a pre-set original product recommendation model and target samples to obtain the recommended product performance change data, which will not be elaborated here.

[0132] In step S603 of some embodiments, specifically, the target loss value is the difference between the training predicted recommended product, the predicted product performance change data, the real recommended product, and the real product performance change data, which incorporates the constraints of the optimized sample sampling strategy.

[0133] Specifically, the recommendation task loss value among training predicted recommended products, predicted product performance change data, real recommended products, and real product performance change data can be calculated using the cross-entropy loss function. This loss value is then summed with the entropy constraint term of the optimized sample sampling strategy to obtain the target loss value. This regularization constraint term is used to prevent data bias caused by excessive bias of sampled samples towards high-yield customers.

[0134] Specifically, the target loss value can be determined using the following formula:

[0135] in, Indicates the target loss value. This represents the recommended task loss value. This represents a hyperparameter used to balance the importance of the recommendation task loss and the entropy constraint term. Indicates the optimization of sample sampling strategy Entropy constraint terms.

[0136] In step S604 of some embodiments, specifically, the target product recommendation model is a neural network model trained from the original product recommendation model.

[0137] Specifically, it can be determined whether the target loss value meets the preset loss condition (such as being less than the preset loss threshold of 0.6). If it does not meet the condition, the gradient parameters of the model in the original product recommendation model can be updated through the backpropagation algorithm to reduce the loss and obtain the updated target product recommendation model. Then, the steps of updating the speech synthesis model pair and updating the sample pair to update the candidate users for product recommendation are executed until the target loss value meets the loss condition and the target product recommendation model is obtained.

[0138] Steps S601 to S604 ensure that the model's learning process is directly aligned with the product conversion efficiency target, thereby improving the model's accuracy in identifying product needs of different users and the accuracy of product recommendations.

[0139] In step S107 of some embodiments, specifically, the target user's target interaction behavior is the target user's interaction operation on a preset product. The target interaction behavior may include target product click behavior, target product dwell behavior, and target product conversion behavior, etc.

[0140] Specifically, target users can come from candidate users or newly registered users on the business platform.

[0141] Specifically, target interaction behaviors can be extracted from the interaction log database of the business platform, the interaction log data of the business APP, and the customer relationship management system, according to a preset time window (such as from 2 days ago to the present).

[0142] Please see Figure 7 In some embodiments, step S107 may include, but is not limited to, steps S701 to S704: Step S701: Obtain target user information of the target user, and construct a target user profile based on the target user information and target interaction behavior.

[0143] Step S702: Based on the target product recommendation model and the target user profile, predict product preferences to obtain product preference data.

[0144] Step S703: Infer product demand based on product preference data to obtain product demand probability.

[0145] Step S704: Based on the product demand probability, perform product recommendation prediction to obtain the target predicted recommended product, and push the target predicted recommended product to the target user.

[0146] In step S701 of some embodiments, specifically, the target user profile refers to the visualized entity data after associating the target user attribute tags and the corresponding target user identification information.

[0147] Specifically, the method for constructing a target user profile based on target user information and target interaction behavior is the same as the method for constructing a candidate user profile based on candidate interaction behavior, and will not be elaborated here.

[0148] In step S702 of some embodiments, specifically, the product preference data refers to the product preference probability distribution data of the target user.

[0149] For example, in the insurance scenario, a target user profile containing "youth, first-tier city, and strong intention to explore critical illness insurance" can be input into the target product recommendation model to calculate the preference probability distribution as 0.65 for consumer critical illness insurance, 0.20 for million-dollar medical insurance, 0.10 for term life insurance, and 0.05 for accident insurance. This preference probability distribution is used to quantify the model's assessment of the target user's interest in various types of products.

[0150] In step S703 of some embodiments, specifically, the product demand probability refers to the intensity of product demand from the target user.

[0151] For example, in an insurance scenario, the business rule base can reveal that although the target user prefers "consumer-type critical illness insurance", the target user profile shows that "they already have social security and ordinary medical insurance". The rule base can then determine that the target user has a stronger need for "supplementing high medical expenses". Therefore, the probability of demand for "million-dollar medical insurance" is increased from 0.20 to 0.35, and other probabilities are adjusted accordingly to form the final product demand probability list.

[0152] In step S704 of some embodiments, specifically, the target predicted recommended product refers to the specific product list or a single product actually pushed to the target user.

[0153] Specifically, products can be sorted from high to low according to their demand probability, and the top-N (e.g., top-3) products or all products with a probability exceeding a certain threshold can be selected to form a final recommendation list. The products in the list can then be pushed to target users through the recommendation engine's display interface (e.g., APP information flow, SMS, customer service workbench).

[0154] For example, in an insurance scenario, after demand inference, the final product recommendation ranking is: million-dollar medical insurance 0.35, term critical illness insurance 0.30, and comprehensive accident insurance 0.25. A recommendation list of million-dollar medical insurance, term critical illness insurance, and comprehensive accident insurance is generated, and the personalized promotion plan of "million-dollar medical insurance" is given priority to the target users in the "Recommended for You" module on the APP homepage.

[0155] Through steps S701 to S704, it is possible to capture users' real needs based on dynamic user profiles, and recommend personalized product solutions that match the real needs of target users and have high product conversion efficiency, which significantly improves the accuracy of product recommendations and product conversion rate.

[0156] This application first acquires candidate samples of user data, including candidate interaction behaviors and user profiles built based on these behaviors. Then, it evaluates the product performance of these candidate interactions and user profiles to obtain candidate product performance scores, accurately identifying the true needs and conversion rates of different user groups. Second, it selects candidate samples based on their performance scores and updates target samples based on changes in recommended product performance data. This allows for targeted selection of user groups that significantly contribute to product conversion, avoiding the misalignment of needs caused by random sample selection in traditional methods. Furthermore, it trains the original product recommendation model using the target samples to obtain a target product recommendation model. This ensures that the model's learning process is directly aligned with the product conversion performance target, improving the model's accuracy in identifying product needs and recommending products to different users. Finally, it recommends products to target users using the target product recommendation model and target interaction behaviors, outputting personalized product solutions that match users' true needs and have high conversion performance, significantly improving the accuracy of product recommendations and product conversion rates.

[0157] Please see Figure 8 This application also provides a sample-optimized product recommendation apparatus, which can implement the above-described sample-optimized product recommendation method. The apparatus includes: The candidate sample acquisition module is used to acquire candidate samples of candidate users; wherein, the candidate samples include candidate interaction behaviors of candidate users and candidate user profiles constructed based on the candidate interaction behaviors; wherein, the candidate interaction behaviors are the interaction operations of candidate users on preset products; The product performance evaluation module is used to evaluate product performance based on candidate interaction behaviors and candidate user profiles, and obtain candidate product performance scores. The sample selection module is used to select target samples from the candidate samples based on the candidate product performance scores. The training product recommendation module is used to recommend products to the candidate users using a preset original product recommendation model and the target sample, and to obtain data on the change in the effectiveness of the recommended products. The sample update module is used to update the target sample based on the recommended product performance change data to obtain an updated sample; The model training module is used to train the original product recommendation model based on the updated samples to obtain the target product recommendation model. The target product recommendation module is used to acquire the target user's target interaction behavior and recommend products to the target user based on the target product recommendation model and the target interaction behavior.

[0158] The specific implementation of this sample-optimized product recommendation device is basically the same as the specific implementation of the sample-optimized product recommendation method described above, and will not be repeated here.

[0159] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described sample-based optimization product recommendation method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0160] Please see Figure 9 , Figure 9 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to perform related programs to implement the technical solutions provided in the embodiments of this application. The memory 902 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 902 can store the processing system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and called by the processor 901 to perform the sample-optimized product recommendation method of the embodiments of this application. The input / output interface 903 is used to implement information input and output; The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904); The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0161] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described sample-optimized product recommendation method.

[0162] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0163] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0164] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0165] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be placed in one location or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0166] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0167] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0168] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0169] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or omitted. The couplings or direct couplings or communication connections shown or discussed may be through some interfaces, or indirect couplings or communication connections between devices or units, and may be electrical, mechanical, or other forms.

[0170] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be placed in one location or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0171] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0172] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0173] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A product recommendation method based on sample optimization, characterized in that, The method includes: Obtain candidate samples of candidate users; wherein, the candidate samples include candidate interaction behaviors of candidate users and candidate user profiles constructed based on the candidate interaction behaviors; wherein, the candidate interaction behaviors are the interaction operations of the candidate users on preset products; Product performance is evaluated based on the candidate interaction behaviors and the candidate user profiles to obtain candidate product performance scores. Based on the candidate product performance scores, the candidate samples are selected to obtain the target samples; By using a preset original product recommendation model and the target sample, product recommendations are made to the candidate users to obtain data on changes in the effectiveness of the recommended products. The target sample is updated based on the recommended product performance change data to obtain an updated sample; The original product recommendation model is trained based on the updated samples to obtain the target product recommendation model. The target user's target interaction behavior is obtained, and product recommendations are made to the target user based on the target product recommendation model and the target interaction behavior.

2. The method according to claim 1, characterized in that, The step of evaluating product performance based on the candidate interaction behaviors and the candidate user profiles to obtain candidate product performance scores includes: Activity level is obtained by performing activity detection based on the candidate interaction behaviors; Based on the candidate user profiles, a sample risk assessment is performed to obtain the user risk level; Based on the candidate user profile, sample period detection is performed to obtain the user period stage; The product performance score is obtained by scoring the candidate product based on the behavioral activity level, the user risk level, and the user cycle stage.

3. The method according to claim 1, characterized in that, The step of updating the target sample based on the recommended product performance change data to obtain an updated sample includes: Based on the recommended product performance change data, construct a product performance change function; An initial sample sampling strategy is obtained, and the initial sample sampling strategy is optimized according to the product performance change function to obtain an optimized sample sampling strategy. The target sample is updated by sampling according to the optimized sample sampling strategy to obtain the updated sample.

4. The method according to claim 3, characterized in that, The updated sample includes real product performance change data and real recommended products; The step of training the original product recommendation model based on the updated samples to obtain the target product recommendation model includes: Based on the original product recommendation model, product recommendation prediction is performed on the updated sample to obtain the trained predicted recommended products; Obtain updated candidate users from the updated samples, and recommend the trained prediction recommendation products to the target candidate users to obtain data on the change in the effectiveness of the predicted products; The target loss value is obtained by calculating the loss based on the trained predicted recommended products, the predicted product performance change data, the actual recommended products, the actual product performance change data, and the optimized sample sampling strategy. The original product recommendation model is trained based on the target loss value to obtain the target product recommendation model.

5. The method according to claim 1, characterized in that, The candidate user profile constructed based on the candidate interaction behaviors includes: Obtain candidate user information of the candidate users, and extract features from the candidate user information to obtain candidate user information features; Feature extraction is performed on the candidate interaction behaviors to obtain candidate interaction behavior features; The candidate user information features and the candidate interaction behavior features are weighted and fused to obtain user profile features; Based on the user profile features, candidate user attribute labels are generated; The candidate user profile is generated based on the candidate user attribute tags.

6. The method according to claim 5, characterized in that, The candidate interactive behaviors include product click behavior, product dwell behavior, and product conversion behavior; The step of extracting features from the candidate interaction behaviors to obtain candidate interaction behavior features includes: Click frequency analysis is performed on the product click behavior to obtain click behavior characteristics; Dwell time sequence analysis is performed on the product dwell behavior to obtain dwell behavior characteristics; The conversion effectiveness of the product conversion behavior is analyzed to obtain the characteristics of the conversion effectiveness behavior; The click behavior feature, the dwell behavior feature, and the conversion effect behavior feature are weighted to obtain the candidate interaction behavior feature.

7. The method according to any one of claims 1 to 6, characterized in that, The step of recommending products to the target user based on the target product recommendation model and the target interaction behavior includes: Obtain the target user information of the target user, and construct a target user profile based on the target user information and the target interaction behavior; Based on the target product recommendation model and the target user profile, product preference prediction is performed to obtain product preference data; Based on the product preference data, product demand is inferred to obtain the product demand probability; Based on the product demand probability, a product recommendation prediction is made to obtain the target predicted recommended product, and the target predicted recommended product is pushed to the target user.

8. A product recommendation device based on sample optimization, characterized in that, The device includes: A candidate sample acquisition module is used to acquire candidate samples of candidate users; wherein, the candidate samples include candidate interaction behaviors of candidate users and candidate user profiles constructed based on the candidate interaction behaviors; wherein, the candidate interaction behaviors are the interaction operations of the candidate users on preset products; The product performance evaluation module is used to evaluate product performance based on the candidate interaction behavior and the candidate user profile, and obtain the candidate product performance score. The sample selection module is used to select target samples from the candidate samples based on the candidate product performance scores. The training product recommendation module is used to recommend products to the candidate users using a preset original product recommendation model and the target sample, and to obtain data on the change in the effectiveness of the recommended products. The sample update module is used to update the target sample based on the recommended product performance change data to obtain an updated sample; The model training module is used to train the original product recommendation model based on the updated samples to obtain the target product recommendation model. The target product recommendation module is used to acquire the target user's target interaction behavior and recommend products to the target user based on the target product recommendation model and the target interaction behavior.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the sample-optimized product recommendation method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, The computer program, when executed by a processor, implements the sample-optimized product recommendation method according to any one of claims 1 to 7.

Images