Smart portfolio pairing method, system, storage medium and product for equity portfolio package

Abstract

This application provides a method, system, storage medium, and product for intelligent combination and pairing of benefit packages, relating to the field of benefit package push. The method includes: calculating the vector matching degree and estimated interaction response probability of each candidate benefit for the target user based on the user state vector corresponding to the target user's historical behavior data and the benefit feature vector corresponding to the attribute configuration data of each candidate benefit in the benefit library; filtering a set of candidate benefits from the benefit library based on the vector matching degree; calculating the dynamic expected occupancy weight of each target benefit based on the estimated interaction response probability and the static resource occupancy weight; calculating the score value corresponding to each candidate benefit package based on the vector matching degree and the correlation gain coefficient between each target benefit in the candidate benefit packages that meet preset resource constraints; and pushing the candidate benefit package with the highest score to the target user. This method can improve the effective information density and pairing accuracy of benefit packages.

Description

Technical Field

[0001] This application relates to the field of rights and benefits package push, and in particular to a method, system, storage medium and product for intelligent combination and pairing of rights and benefits packages. Background Technology

[0002] In internet information distribution and digital service platforms, to enhance user activity and maintain a healthy platform ecosystem, intelligent recommendation systems for benefit packages typically push data packages (i.e., benefit packages) containing various heterogeneous benefits (such as digital content services and lifestyle service vouchers) to users. This process is essentially a data processing step of selecting and constructing a set of target objects from a massive candidate object library under limited system resource constraints. The system needs to ensure that the data packages pushed to users meet specific constraints (such as quantity limits and total value thresholds) while maximizing the depth of user interaction with those data packages.

[0003] When constructing such combined data packets, related technologies typically employ a resource allocation logic based on the "knapsack problem" principle. Specifically, the system pre-configures a static resource occupancy weight for each data object (right) in the candidate object library (this weight is usually based on the object's inherent attributes, such as face value or fixed procurement costs), and sets a total resource distribution threshold for a single user. During recommendation generation, the system adds data objects to the combined list sequentially based on object popularity or basic tag matching results, accumulating their corresponding static resource occupancy weights until the accumulated value reaches the upper limit of the total resource threshold, thus outputting a combined data packet that meets the nominal constraints.

[0004] However, the resource allocation weights in related technologies are static, representing the theoretical maximum resource consumption when the data object is fully redeemed. In contrast, users' actual interaction intentions towards different objects are highly random and varied. When the system filters data for a specific user, if that user has extremely low interaction intentions towards certain objects with high static weights, the accumulation logic of the related technologies will still calculate and reserve space based on the full static weight of that object. This may cause the system's internal calculation logic to determine that "the resource threshold is full," thus prematurely stopping the filtering process or crowding out the admission space for other objects. This processing method results in the final generated combined data package (i.e., the rights and benefits package) nominally reaching the resource threshold, but containing a low effective information density (i.e., the expected probability of the user's actual response), causing a resource mismatch between the distribution strategy and the user's actual needs, and reducing the accuracy of the system's data distribution. Summary of the Invention

[0005] This application provides a method, system, storage medium, and product for intelligent combination and pairing of equity portfolios, which addresses the problem in related technologies where the use of static resource occupancy weights for constraint control during the generation of equity portfolios may lead to an imbalance between the nominal resource occupancy and actual effectiveness of data objects, thereby reducing the effective information density and pairing accuracy of equity portfolios.

[0006] Firstly, this application provides an intelligent combination and matching method for equity portfolios, applied to an intelligent recommendation system for equity portfolios, the method comprising: Feature extraction is performed on the historical behavior data of the target user and the attribute configuration data of each candidate right in the right pool to generate user state vector and right feature vector of each candidate right. The attribute configuration data includes the static resource occupation weight of the right, the classification attribute of the domain, and the association attribute of the historical co-occurrence relationship between the rights. Based on the user state vector and the benefit feature vector, calculate the vector matching degree of each candidate benefit for the target user and the estimated probability of interaction response; A candidate equity set is selected from the equity pool based on the vector matching degree, and the candidate equity set includes multiple target equitys; The dynamic expected occupancy weight of each target right is calculated based on the estimated interaction response probability and the static resource occupancy weight of each target right; Based on preset resource constraints, the target rights in the candidate rights set are arranged and combined to generate several candidate rights combinations. The resource constraints include the upper limit of the number of rights in the rights combination package, the mutual exclusion rule of the classification attributes, and the cumulative threshold of the dynamic expected occupancy weight. The score value corresponding to each of the candidate benefit combinations is calculated based on the vector matching degree and the correlation gain coefficient between each target benefit in the candidate benefit combination, and the candidate benefit combination with the highest score value is pushed to the target user. The correlation gain coefficient is determined based on the correlation attribute.

[0007] By adopting the above technical solution, the system uses a probabilistic prediction model to correct the static resource value of rights and interests, generating a dynamic expected occupancy weight value. This transforms the judgment criterion from nominal resource consumption to actual expected resource occupancy, enabling the inclusion of more high-potential-value rights and interests within a limited resource threshold. Simultaneously, by combining classification mutual exclusion and correlation gain coefficients to screen and score rights and interests, the system ensures that the final pushed bundles not only meet resource constraints but also possess logical complementarity among their internal rights and interests, thereby improving the effective information density and matching accuracy of the rights and interests bundles.

[0008] In some embodiments, the step of calculating the vector matching degree of each candidate benefit for the target user and estimating the probability of interaction response based on the user state vector and the benefit feature vector specifically includes: The user state vector and the feature vector of each candidate right are calculated in the same feature space to obtain the vector matching degree of each candidate right for the target user; The user state vector and the benefit feature vector are input into a pre-trained interaction behavior prediction model, which outputs the probability value of each candidate benefit being effectively interacted with by the target user. The probability value is determined as the estimated interactive response probability.

[0009] By employing the above technical solution, the system calculates vector similarity within the same feature space to capture the potential interest matching relationship between users and rights, and further combines this with an interaction behavior prediction model to output specific probability values. This dual evaluation mechanism retains the generalized matching capability based on feature distance while introducing a nonlinear model to accurately judge the user's deep interaction intent. This improves the accuracy and reliability of subsequent calculations of dynamically expected weight allocation.

[0010] In some embodiments, before the step of inputting the user state vector and the benefit feature vector into a pre-trained interaction behavior prediction model and outputting the probability value of each candidate benefit being executed by the target user in an effective interaction behavior, the method further includes: Based on users' historical behavior data and the attribute configuration data of each historical benefit, a training dataset containing positive and negative samples is constructed. The positive samples are records in which users generate effective interactive responses to the benefits, and the negative samples are records in which users only generate exposure to the benefits but do not generate effective interactive responses. The deep neural network model is trained using the training dataset under supervised learning until the loss function value converges or the preset number of training rounds is reached, at which point training stops, thus obtaining the interactive behavior prediction model.

[0011] By employing the aforementioned technical solution, the system conducts supervised training of a deep neural network by meticulously constructing a negative sample dataset containing records of unresponsive exposures. This process enables the model to effectively distinguish between users' "not seen" and "not interested," thereby learning deeper-level characteristic patterns of user rejection or ignoring behavior. The model, after sufficient iterative training, is able to more acutely perceive the impact of different attribute configuration data on users' willingness to interact.

[0012] In some embodiments, the step of calculating the dynamic expected occupancy weight of each target interest based on the estimated interaction response probability and the static resource occupancy weight corresponding to each target interest specifically includes: Obtain the basic resource consumption value corresponding to the inherent resource consumption attribute of each target benefit from the attribute configuration data; Multiply the base occupancy value corresponding to each target benefit by the estimated interaction response probability to obtain the dynamic expected occupancy weight value corresponding to each target benefit. The dynamic expected occupancy weight value represents the expected resource occupancy of the platform after correction based on the interaction response probability when distributing target benefits to the target user.

[0013] By adopting the above technical solution, the system calculates a weighted average of the inherent base value of a right and the probability of user interaction with it. This derivation process quantifies the risk cost of distributing rights on the platform, enabling dynamic adjustment of resource allocation. For rights with low interaction probabilities, their allocated system resource budget is reduced accordingly, thereby freeing up more resource quota space for other candidate rights. This also allows the system to flexibly adjust its combination strategy under strict overall resource constraints, maximizing the expected interaction value within a unit resource quota and improving the marginal efficiency of resource allocation.

[0014] In some embodiments, the step of calculating the score value corresponding to each of the candidate equity portfolios based on the vector matching degree and the correlation gain coefficient between each target equity in the candidate equity portfolio specifically includes: Based on the associated attributes in the attribute configuration data, obtain the co-occurrence frequency statistics of any two target rights in the candidate rights set in historical data; Based on the co-occurrence frequency statistics and the feature distance between the classification attributes of the two target interests, the synergistic correlation degree between any two target interests is calculated, and the synergistic correlation degree is determined as the correlation gain coefficient. Set the sum of all vector matching degrees in the candidate equity portfolio as the basic matching score; The sum of the correlation gain coefficients between each pair of target interests in the candidate interest portfolio is set as the portfolio correlation score; The basic matching score and the combined association score are weighted and summed according to preset weight coefficients to obtain the score value corresponding to each candidate benefit combination.

[0015] By employing the aforementioned technical solutions, the system mines historical co-occurrence data and the distance between classification attributes to quantify the synergistic effect between benefits. During the scoring phase, not only is the matching degree between individual benefits and users considered, but a weighted summation of combined correlation scores is also introduced. This mechanism ensures that the selected benefits are not isolated individuals, but rather an organic whole with complementary or beneficial relationships. By improving the logical coherence within the bundle, user acceptance and satisfaction with the overall bundle are enhanced.

[0016] In some embodiments, after the steps of calculating the vector matching degree of each candidate benefit for the target user and estimating the probability of the interaction response based on the user state vector and the benefit feature vector, the method further includes: Get the remaining distribution quota of each candidate equity in the equity pool and the quota consumption rate within a preset time window; The resource competition coefficient of each candidate interest is calculated based on the remaining allocation quota and the quota consumption rate. The resource competition coefficient is negatively correlated with the remaining allocation quota and positively correlated with the quota consumption rate. A penalty factor is constructed based on the resource competition coefficient, and when the resource competition coefficient is higher than a preset warning value, the vector matching degree of the corresponding candidate interest is reduced based on the penalty factor.

[0017] By adopting the above technical solution, the system introduces a dynamic control mechanism based on inventory consumption rate. By calculating the resource competition coefficient, the scarcity and consumption intensity of benefits are perceived in real time. When a specific benefit faces the risk of overheating or shortage, the system proactively reduces its matching degree using a penalty factor, thereby achieving automatic traffic smoothing and load balancing at the algorithm level. This helps prevent popular benefits from being exhausted prematurely, while guiding users to focus on other high-quality benefits with sufficient inventory, ensuring the stability and long-term effectiveness of the platform's overall benefit distribution service.

[0018] In some embodiments, after the step of pushing the candidate benefit combination with the highest rating to the target user, the method further includes: Real-time monitoring of target users' interaction behavior data with the proposed combination of benefits; If the interactive behavior data is a negative feedback instruction, then the feature vector of each target benefit in the candidate benefit combination is extracted as a negative feedback feature; The user state vector is orthogonally corrected based on the negative feedback features to remove feature components in the user state vector that are similar to the negative feedback features, thereby obtaining the corrected user state vector. The vector matching degree of each candidate benefit with respect to the revised user state vector is recalculated to generate a new combination of candidate benefits for updated recommendations.

[0019] By employing the above technical solution, after receiving negative feedback from a user, the system extracts negative feedback features and orthogonally corrects the user's state vector. This process essentially removes feature components in the vector space that align with the direction of the content the user dislikes, thereby quickly reconstructing the user's real-time preference expression. This correction method not only responds rapidly to the user's intention to reject the feedback but also effectively avoids interests with similar attributes when generating new recommendations, improving the diversity and targeting of updated recommendations.

[0020] Secondly, this application provides an intelligent recommendation system for equity portfolio packages, the system comprising: one or more processors and a memory; The memory is coupled to the one or more processors. The memory is used to store computer program code, which includes computer instructions. The one or more processors call the computer instructions so that the system can implement the intelligent combination and pairing method of the rights and interests package provided in the above embodiments, which will not be described in detail here.

[0021] Thirdly, this application provides a computer-readable storage medium including instructions that, when executed on a smart recommendation system for benefit packages, enable the system to implement a smart combination and matching method for benefit packages provided in the above embodiments, which will not be elaborated further here.

[0022] Fourthly, this application provides a computer program product that, when running on an intelligent recommendation system for equity portfolios, enables the system to implement an intelligent combination and matching method for equity portfolios provided in the above embodiments, which will not be elaborated here.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. The system uses a deep learning model to predict the probability of user interaction with each benefit, and then multiplies this probability by the inherent resource consumption value of the benefit as a correction coefficient. This transforms the rigid nominal resource occupation into a flexible expected occupation that reflects distribution risk. This method can identify and reduce the computational weight of benefits with low user engagement in the system, thereby freeing up more resource quota space. This allows the system to load more high-potential-value benefit objects under a fixed total resource constraint, improving the effective information density within a unit resource quota and the conversion rate generated by the increased value of the bundled packages.

[0024] 2. By analyzing the co-occurrence frequency statistics and feature distances of classification attributes among benefits in historical data, the system quantifies the synergistic correlation when different benefits coexist in the same bundle. When generating recommendations, it not only considers the popularity of individual benefits but also introduces a weighted summation of bundle correlation scores to filter out sets of benefits with logical complementarity or scenario coherence. This approach avoids the feeling of a patchwork of functional conflicts or logical disjointedness within the bundle, ensuring that what is pushed to users is not merely a collection of high-scoring individual items, but an organic whole with inherent synergistic relationships, thus improving the consistency of the user experience.

[0025] 3. When a user's rejection or negative action is detected, the system does not use traditional demotion or blacklisting logic. Instead, it directly extracts the feature vector of the negative feedback benefit and orthogonally processes the user's state vector accordingly. This process can accurately remove feature components that are consistent with (i.e., similar to) the negative feedback content while preserving other dimensions of user interest, thus correcting the user's real-time preferences. By regenerating recommendations using this corrected state vector, the system can avoid the recurrence of homogeneous content to the greatest extent, improving the system's error correction capability and recommendation accuracy in continuous interaction scenarios. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a smart combination and pairing method for a rights and interests package in an embodiment of this application; Figure 2 This is another flowchart illustrating a smart combination and pairing method for a rights and interests package in an embodiment of this application; Figure 3 This is a schematic diagram of the physical device structure of an intelligent recommendation system for rights and interests packages in this application embodiment. Detailed Implementation

[0027] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0028] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0029] In the scenario of pushing benefit packages on internet information distribution and digital service platforms, the system needs to push benefit packages with high matching degree and high effective information density to users under limited resource constraints. Related technologies are prone to resource mismatch due to the use of static resource occupancy weights. This application transforms static resource constraints into dynamic expectation constraints through a full-link operation of feature extraction, matching degree and probability calculation, benefit screening, dynamic weight calculation, combination generation and scoring push. At the same time, it constructs high-quality combinations by combining benefit association characteristics, thereby solving the resource mismatch problem and improving the matching accuracy of benefit packages and the probability of user interaction.

[0030] For ease of understanding, the method provided in this implementation is described in process below. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating a smart combination and pairing method for a rights package in an embodiment of this application.

[0031] S101. Extract features from the target user's historical behavior data and the attribute configuration data of each candidate right in the rights and benefits pool, respectively, to generate user state vector and rights feature vector of each candidate right.

[0032] Among them, the target user refers to the specific user to whom the system intends to push the benefit package; historical behavior data refers to the relevant records of the target user's interaction with various benefits on the platform, which is used to reflect the user's benefit preferences; the benefit library refers to the database on the platform that stores all distributable benefits, and candidate benefits refer to various benefits in the benefit library that can be screened for combination; the user state vector is a vector that quantitatively represents the characteristics of the target user, which is used to reflect the user's preferences, behavioral characteristics and other states; the benefit feature vector is a vector that quantitatively represents the characteristics of the candidate benefits, which is used to reflect the various attribute characteristics of the benefits.

[0033] This step is executed at the initial stage of the formal start of the benefits package matching process after the system receives a request to push a benefits package to a specific target user. The application scenario is all platform service scenarios that need to generate personalized benefits packages for users, including benefit pushes on e-commerce platforms and benefits distribution on digital service platforms.

[0034] Specifically, the system identifies the target user to whom the benefit package needs to be pushed, retrieves all historical behavior data related to benefits within the platform for that user, and simultaneously retrieves the attribute configuration data of all candidate benefits in the benefit library. This attribute configuration data includes the static resource occupancy weight of the benefit, the classification attribute of its domain, and the correlation attribute of historical co-occurrence relationships between benefits. Subsequently, the system performs feature extraction on the two types of raw data. For the historical behavior data, it extracts core features that reflect the user's benefit preferences, interaction habits, and interest tendencies. For the attribute configuration data, it extracts core features that reflect the core attributes, resource characteristics, and correlation characteristics of the benefits. The extracted user features are then quantified to generate a user state vector representing the user's overall state. The extracted benefit features are also quantified to generate a benefit feature vector corresponding to each candidate benefit, and all user state vectors and benefit feature vectors are located in the same feature space.

[0035] In some embodiments, the system can perform data cleaning on historical behavior data and attribute configuration data, removing invalid, missing, and abnormal raw data; then, the TF-IDF feature extraction algorithm is used to extract core features from the two types of cleaned data, forming a user feature set and a rights feature set; finally, the two feature sets are vectorized and encoded, using one-hot encoding to convert discrete features into numerical features, and then feature normalization is performed to generate user state vectors and rights feature vectors with uniform dimensions. It is understood that other methods can also be used to achieve feature extraction and vector generation, such as using a deep learning model CNN for feature extraction followed by vectorization; this is not limited here.

[0036] S102. Calculate the vector matching degree of each candidate right for the target user and the estimated probability of interactive response based on the user state vector and the right feature vector.

[0037] Among them, vector matching degree refers to the similarity between the user state vector and the candidate benefit feature vector, which is used to characterize the basic matching degree between the candidate benefit and the target user; the estimated interaction response probability refers to the probability that the system predicts that the target user will generate effective interaction behavior with the candidate benefit, which is expressed in the form of probability value.

[0038] Specifically, the system performs two core calculations based on user state vectors within the same feature space and the feature vectors of each candidate benefit. First, it calculates the vector matching degree. Using a professional vector similarity calculation method, it measures the similarity between the feature vector of each candidate benefit and the target user's state vector, obtaining the vector matching degree for each candidate benefit relative to the target user. A higher matching degree indicates a better basic match between the candidate benefit and the user. Second, it calculates the estimated interaction response probability. Using a professional prediction model or algorithm, the system combines user and benefit features contained in the user state vector and benefit feature vector to predict the probability that the target user will generate an effective interaction behavior for each candidate benefit, obtaining the estimated interaction response probability for each candidate benefit. A higher probability value indicates a greater likelihood that the user will generate an effective interaction for that benefit.

[0039] Optionally, the system can use a cosine similarity algorithm to calculate the cosine similarity between the user's state vector and the feature vectors of each candidate right, thereby obtaining the vector matching degree of each candidate right. Then, based on the statistical data of the user's historical interactions, a probabilistic statistical model is constructed, and the vector matching degree is used as the core feature input to the model. Combined with the classification attribute features of the right, the estimated interaction response probability of each candidate right is calculated. Finally, the calculated probability values ​​are normalized to make them fall within a reasonable range of 0-1.

[0040] S103. Select a set of candidate rights from the rights pool based on the vector matching degree.

[0041] Specifically, the system presets a threshold for vector matching, which can be dynamically adjusted based on factors such as the platform's rights distribution strategy, resource availability, and user level. The system then compares the vector matching degree of each candidate right in the rights pool with this threshold, selecting all candidate rights with a vector matching degree greater than or equal to the preset threshold. These rights are then integrated into a new rights set, i.e., the candidate rights set, and each right in this set is the target right.

[0042] If some candidate rights have the same vector matching degree and are at the critical point of the screening threshold, the system can perform secondary screening by combining the basic attributes of the rights to ensure that the target rights in the candidate rights set have basic matching with the target user and can further participate in the subsequent dynamic weight calculation and rights combination operation. At the same time, the size of the candidate rights set not only meets the diversity needs of subsequent combinations, but also does not lead to low computational efficiency due to too many rights.

[0043] In some embodiments, the system can set personalized vector matching degree filtering thresholds for target users based on historical distribution data and user profiles. Then, it compares the vector matching degree of each candidate benefit with the personalized threshold one by one, filtering out benefits that meet the threshold requirements. Finally, it deduplicates the filtered benefits, removing duplicate entries, generating a candidate benefit set, and marking the benefits within it with the target benefit identifier. Optionally, the system can also set a unified basic matching degree filtering threshold to filter out all candidate benefits that meet the basic threshold. Then, it sorts the filtered benefits according to their vector matching degree from high to low, selecting the top-ranked benefits based on the platform's preset upper limit for the candidate benefit set. Finally, it performs compliance verification on the selected benefits, removing benefits that do not comply with the platform's distribution rules, forming the final candidate benefit set, which is the target benefit. It is understood that other methods can also be used to achieve benefit filtering, such as using quantile filtering to determine the filtering threshold, and then combining it with the remaining benefit quota for secondary filtering; this is not limited here.

[0044] S104. Calculate the dynamic expected occupancy weight of each target right based on the estimated interaction response probability and static resource occupancy weight corresponding to each target right.

[0045] Among them, the dynamic expected resource occupancy weight value refers to the expected resource occupancy value after being corrected based on the interaction response probability when the platform distributes target benefits to target users. It is a dynamic adjustment of the static resource occupancy weight value.

[0046] Specifically, the system retrieves the corresponding static resource occupancy weight from the attribute configuration data of each target benefit. This weight is the basic value corresponding to the inherent resource consumption attribute of the benefit. Then, the system treats each target benefit as an independent accounting unit, correlates its corresponding estimated interaction response probability with the static resource occupancy weight, and dynamically adjusts the static resource occupancy weight using the estimated interaction response probability as a correction coefficient through reasonable calculation logic, thus obtaining the dynamic expected occupancy weight for each target benefit. This weight accurately reflects the actual expected resource occupancy based on the probability of user interaction during the actual distribution process. For target benefits with low estimated interaction response probabilities, their dynamic expected occupancy weight will be reduced accordingly, thereby releasing more resource quotas and providing combination space for other highly matched, high-interaction-probability benefits, achieving a rational allocation of resources.

[0047] Optionally, the system can extract the static resource occupancy weight from the attribute configuration data of the target right, perform range verification on the weight to ensure that it is within the reasonable resource value range preset by the platform; then retrieve the estimated interaction response probability corresponding to the target right, and perform a direct multiplication operation between the probability value and the static resource occupancy weight to obtain the preliminary dynamic expected occupancy weight; finally, the preliminary calculation result is rounded to the nearest integer to generate the dynamic expected occupancy weight that meets the platform's resource accounting accuracy requirements.

[0048] S105. Based on the preset resource constraints, the target rights in the candidate rights set are arranged and combined to generate several combinations of candidate rights.

[0049] Among them, resource constraints refer to the various restrictive rules set by the platform for generating benefit packages, which are the basic criteria for benefit packages; candidate benefit packages refer to benefit packages that meet the resource constraints and are qualified for further scoring and screening, which are the candidate forms of benefit packages.

[0050] Specifically, the system first obtains preset resource constraints, which include three core aspects: First, the upper limit of the number of rights in a rights package, that is, the number of target rights in a single package shall not exceed the upper limit; second, the mutual exclusion rule of classification attributes, that is, target rights with mutually exclusive classification attributes in some fields shall not appear in the same rights package at the same time; and third, the cumulative threshold of dynamic expected occupancy weight, that is, the cumulative sum of the dynamic expected occupancy weight of all target rights in a single package shall not exceed the preset threshold.

[0051] The system then uses this resource constraint as the basic criterion to perform permutation and combination operations on all target rights in the candidate rights set. During the combination process, each combination is checked to see if it meets the three constraint rules. All invalid combinations that do not meet the rules are eliminated. The valid combinations that meet all the rules are integrated to generate several candidate rights combinations. Each candidate rights combination meets the platform's resource distribution requirements and is qualified to participate in the scoring and become the final push combination package.

[0052] Optionally, the system can group the target rights in the candidate rights set according to their classification attributes, mark the rights groups that cannot be combined according to the mutual exclusion rules of the classification attributes, and then combine different quantities and different rights from the non-mutually exclusive rights groups according to the requirement of the upper limit of the number of rights. Finally, the dynamic expected occupancy weight of each preliminary combination is accumulated and calculated, and combinations with accumulated values ​​exceeding the preset threshold are eliminated, and combinations that meet all constraints are retained to generate candidate rights combinations.

[0053] It should be noted that, in this embodiment, the permutation and combination of target interests in the candidate interest set based on preset resource constraints does not refer to a complete exhaustive search in a mathematical sense, but rather to a combination construction process based on a preset optimization algorithm. For example, the system can use a dynamic programming (KnapsackProblem variant) algorithm or a greedy algorithm to prioritize adding the target interests with the highest cost-effectiveness (score / dynamic expected occupancy weight) to the combination container until the boundary of the resource constraints is reached (such as the cumulative threshold of the dynamic expected occupancy weight). In this process, the set that meets the constraints is the candidate interest combination. In this way, the system can complete the combination generation within a controllable computational complexity.

[0054] S106. Calculate the score value corresponding to each candidate benefit combination based on the vector matching degree and the correlation gain coefficient between each target benefit in the candidate benefit combination, and push the candidate benefit combination with the highest score value to the target user.

[0055] Among them, the correlation gain coefficient refers to the coefficient determined based on the correlation attributes between the equity and characterizes the degree of synergistic correlation between two target equitys in the candidate equity portfolio. The higher the coefficient, the better the synergistic gain effect between the two. The score value refers to the numerical value for quantitatively evaluating the overall quality of each candidate equity portfolio.

[0056] Specifically, for each candidate benefit combination, the system retrieves the vector matching degree of each target benefit in the combination relative to the target user, and calculates the correlation gain coefficient between any two target benefits in the combination based on the association attributes of the benefits. Then, the system uses a preset scoring calculation rule to fuse the vector matching degree and the correlation gain coefficient to obtain a score value corresponding to each candidate benefit combination. This score value comprehensively considers the matching degree between individual benefits in the combination and the user, as well as the synergistic association between the benefits. A higher score value indicates a higher overall quality of the candidate benefit combination and a better match with the target user's benefit needs.

[0057] Finally, the system sorts all candidate benefit combinations from highest to lowest score, selects the candidate benefit combination with the highest score as the final benefit package, and pushes it to the target user, thus completing the entire intelligent matching and push process for the benefit package.

[0058] In the above embodiments, the system uses a probabilistic prediction model to correct the static resource value of rights and interests, generating a dynamic expected occupancy weight value. This transforms the judgment criterion from nominal resource consumption to actual expected resource occupancy, enabling more high-potential-value rights and interests to be accommodated within a limited resource threshold. Simultaneously, by combining classification mutual exclusion and correlation gain coefficients to screen and score rights and interests, the system ensures that the final pushed bundle not only meets resource constraints but also has logical complementarity among its internal rights and interests, thereby improving the effective information density and matching accuracy of the rights and interests bundle.

[0059] The following provides a more detailed description of the process of the method provided in this implementation. Please refer to [link / reference]. Figure 2 This is another flowchart illustrating a smart combination and pairing method for a rights package in an embodiment of this application.

[0060] S201. Based on the user's historical behavior data and the attribute configuration data of each historical right, construct a training dataset containing positive and negative samples.

[0061] Positive samples refer to records where users generate effective interactive responses to benefits, and are used to characterize the benefit features that users are interested in; negative samples refer to records where users only generate exposure to benefits but do not generate effective interactive responses, and are used to characterize the benefit features that users are not interested in.

[0062] This step is executed during the model training preparation stage before the system starts the equity portfolio matching process, earlier than the calculation steps of vector matching degree and estimated interaction response probability in the basic process. The application scenario is the early data construction stage of the interaction behavior prediction model, which is a prerequisite for the model to achieve accurate prediction.

[0063] Specifically, the system retrieves massive amounts of historical user behavior data from within the platform, along with complete attribute configuration data for each historical benefit these users have encountered. It then correlates and matches user data with corresponding benefit data, forming one-to-one user-benefit interaction data pairs. Subsequently, the system categorizes and filters these data pairs, selecting records where users have completed effective interactions such as clicking, claiming, and using the benefits, as positive samples in the training dataset; and selecting records where users only saw the benefit exposure but did not engage in any effective interaction, as negative samples in the training dataset. Finally, the system annotates the positive and negative samples, highlighting key information about user characteristics, benefit characteristics, and interaction results. The annotated positive and negative samples are then integrated to form a structurally complete and balanced training dataset, ensuring that the dataset covers interaction scenarios with different user preferences and benefit attributes.

[0064] S202. Supervised learning training of the deep neural network model is performed using the training dataset until the loss function value converges or the preset number of training rounds is reached, at which point training is stopped to obtain the interactive behavior prediction model.

[0065] Among them, deep neural network model refers to machine learning model based on multi-layer neural network structure with feature self-learning ability, used to explore the non-linear relationship between user features and rights features; supervised learning training refers to the training method that uses a labeled training dataset to let the model learn the mapping relationship between input features and output labels; loss function value refers to the numerical value used to measure the error between the model's prediction result and the actual label, the smaller the error, the smaller the loss function value; convergence means that the model's loss function value gradually stabilizes as the number of training epochs increases, and no longer shows a significant decrease; training epochs refer to the number of times the model learns completely from the training dataset.

[0066] Specifically, the system selects a suitable deep neural network model architecture and sets hyperparameters such as the number of layers and neurons based on the scenario requirements of interest interaction prediction. Then, the constructed training dataset is input into the deep neural network model to initiate the supervised learning training process. During training, the model predicts interaction probabilities based on the input user features and interest features, compares the prediction results with the actual interaction labels of the samples, and calculates the loss function value. The system continuously iterates the training process, calculating the loss function value after each training epoch and simultaneously validating the model's prediction performance using a validation set. When the loss function value tends to converge (i.e., the loss function value does not decrease significantly after several consecutive epochs and the validation set performance is stable), or when the model's training epochs reach the platform's preset maximum number of epochs, the system immediately stops training, saves the model parameters at this point, and obtains a trained and ready-to-use interaction behavior prediction model.

[0067] Optionally, the system can first select a deep neural network architecture of CNN + fully connected layers, set appropriate hyperparameters such as convolutional kernel size and the number of neurons in the fully connected layers, and initialize the model parameters; then input the training dataset into the model in batches for supervised learning training, use the gradient descent algorithm to continuously adjust the model parameters to reduce the loss function value, and use the validation set to evaluate the model accuracy after every 10 rounds of training; finally, when the loss function value does not decrease for 20 consecutive rounds or the training rounds reach 100 rounds, stop training, save the optimal model parameters, generate an interactive behavior prediction model, and perform model testing.

[0068] Optionally, the system can first build a deep neural network model using the Transformer infrastructure, and set hyperparameters such as the number of attention heads and the dimension of hidden layers according to the sample feature dimensions; then input the labeled training dataset into the model, calculate the prediction error using the cross-entropy loss function, update the model parameters through the Adam optimizer, and monitor the loss function values ​​of the training set and validation set in real time during the training process; finally, when the loss function value of the validation set converges or the training epochs reach a preset threshold, training is stopped, the model is pruned and optimized to improve the prediction speed, and the final interactive behavior prediction model is obtained.

[0069] It is understandable that other methods can be used to train the model, such as using an LSTM neural network architecture combined with an early stopping mechanism for training; this is not a limitation here.

[0070] S203. Calculate the vector similarity between the user state vector and the feature vector of each candidate right in the same feature space to obtain the vector matching degree of each candidate right for the target user.

[0071] The same feature space refers to the space in which the user state vector and the rights feature vector are located after undergoing the same feature processing and dimension mapping, and the feature dimensions are unified and the feature meanings are corresponding, so as to ensure that similarity calculation can be performed between the vectors.

[0072] Specifically, the system determines whether the target user's state vector and the feature vectors of all candidate benefits reside in the same feature space. If there are inconsistencies in vector dimensions or feature mappings, feature space alignment is immediately performed to ensure that the computational dimensions and feature meanings of all vectors correspond completely. Subsequently, the system selects a suitable vector similarity calculation algorithm and, using the target user's state vector as a benchmark, performs a one-to-one similarity calculation with the feature vector of each candidate benefit. During the calculation, all feature dimension information of the vectors is preserved to fully explore the potential matching relationships between user features and benefit features. Finally, the system quantifies the similarity results obtained from each calculation, converting them into standardized values. This value represents the vector matching degree of the corresponding candidate benefit for the target user; a higher matching degree indicates a stronger feature match between the candidate benefit and the target user.

[0073] Optionally, the system can perform L2 normalization on the user state vector and the rights feature vector to ensure that the magnitude of all vectors is 1, thus completing the final alignment of the feature space. Then, the cosine similarity algorithm is used to calculate the cosine value between the target user state vector and each candidate rights feature vector. This cosine value is the initial vector matching degree. Finally, the initial matching degree is linearly mapped to a numerical range of 0-100 to obtain the standardized vector matching degree and record it.

[0074] S204. Calculate the resource competition coefficient of each candidate right based on the current remaining distribution quota of each candidate right in the rights pool and the quota consumption rate within the preset time window.

[0075] Among them, the remaining distribution quota refers to the amount of each candidate benefit in the benefit pool that has not yet been distributed and can continue to be pushed to users; the preset time window refers to a fixed time period set by the platform according to the benefit distribution rules to count the consumption of quota, such as 1 hour, 24 hours, etc.; the quota consumption rate refers to the number of quotas distributed to candidate benefits per unit time within the preset time window, which is used to characterize the distribution popularity of benefits; the resource competition coefficient is a quantitative indicator used to characterize the scarcity of candidate benefits and the degree of distribution competition, which is negatively correlated with the remaining distribution quota and positively correlated with the quota consumption rate.

[0076] Specifically, the system retrieves the current remaining allocation quota for each candidate right in the rights pool in real time, and simultaneously retrieves the total quota consumption of each candidate right within a preset time window. Based on the total consumption and the duration of the time window, the system calculates the quota consumption rate for each candidate right, i.e., the amount of quota consumed per unit time. Subsequently, the system incorporates the remaining allocation quota and quota consumption rate into the calculation according to preset calculation rules, fully considering their correlation. The lower the remaining allocation quota, the scarcer the right, and the higher the resource competition coefficient; the faster the quota consumption rate, the higher the right's distribution popularity, and the higher the resource competition coefficient. The system performs a weighted fusion calculation of the two indicators to obtain the resource competition coefficient corresponding to each candidate right. This coefficient accurately reflects the current resource competition status of each right, providing a quantitative basis for subsequent matching degree adjustment.

[0077] In some embodiments, the system can calculate the quota consumption rate of each candidate interest, divide the total consumption quota within a preset time window by the window duration to obtain the consumption value per unit time; then normalize the remaining allocation quota and the quota consumption rate to convert them into values ​​of 0-1, and set preset weights for both, with the quota consumption rate having a higher weight than the remaining allocation quota; finally, add (1 - the normalized remaining allocation quota value) to the normalized quota consumption rate value, multiply by the corresponding weight, and sum to obtain the resource competition coefficient of each candidate interest.

[0078] S205. Construct a penalty factor based on the resource competition coefficient, and reduce the vector matching degree of the corresponding candidate interest based on the penalty factor when the resource competition coefficient is higher than the preset warning value.

[0079] Among them, the penalty factor refers to the quantitative coefficient constructed based on the resource competition coefficient, which is used to lower the vector matching degree of candidate rights. The magnitude of the coefficient is positively correlated with the resource competition coefficient. The preset warning value refers to the resource competition coefficient threshold set by the platform based on the resource carrying capacity of rights distribution, which is used to determine whether rights are in a high-competition state. Lowering the vector matching degree of the corresponding candidate rights means calculating the initial vector matching degree of rights in a high-competition state with the penalty factor to obtain the lowered matching degree.

[0080] Specifically, the system constructs a tiered penalty factor system based on the numerical range of the resource competition coefficient. The higher the resource competition coefficient, the larger the corresponding penalty factor, and the greater the reduction in vector matching degree. The penalty factor ranges from 0 to 1 to ensure that the matching degree remains valid after reduction. The system then compares the resource competition coefficient of each candidate benefit with the platform's preset warning value to determine whether each benefit is in a high resource competition state. For benefits with a resource competition coefficient lower than or equal to the preset warning value, their initial vector matching degree remains unchanged. For benefits with a resource competition coefficient higher than the preset warning value, the system retrieves their corresponding penalty factor, multiplies the initial vector matching degree by the penalty factor, and obtains the reduced vector matching degree. This achieves dynamic control of the matching degree of highly competitive benefits, preventing such benefits from being over-recommended due to excessively high matching degrees, which could lead to rapid quota depletion.

[0081] Optionally, the system can divide the resource competition coefficient into multiple tiered intervals and set a corresponding penalty factor for each interval. The higher the coefficient, the smaller the penalty factor for the interval. For example, a coefficient of 0.8-1.0 corresponds to a penalty factor of 0.5, and a coefficient of 0.6-0.8 corresponds to a penalty factor of 0.7. Then, the resource competition coefficient of each candidate interest is compared with the preset warning value to filter out high-competition interests. Finally, for high-competition interests, their initial vector matching degree is multiplied by the penalty factor of the interval to obtain the lowered matching degree, and the adjustment range and reason for the matching degree are recorded.

[0082] S206. Input the user state vector and the benefit feature vector into the interaction behavior prediction model, output the probability value of each candidate benefit being executed by the target user in an effective interaction behavior, and determine it as the estimated interaction response probability.

[0083] Specifically, the system performs format validation on the target user's state vector and the feature vectors of each candidate benefit. Then, the system pairs the validated user state vector with the feature vector of each candidate benefit, using these pairs as input features for the model. Each set of features is sequentially input into the trained interaction behavior prediction model. Based on the user-benefit interaction patterns learned during training, the model processes each set of input features, uncovering the non-linear correlation between user features and benefit features, and outputs a corresponding probability value. This probability value represents the likelihood that the target user will perform an effective interaction with the corresponding candidate benefit. Finally, the system organizes all the probability values ​​output by the model, directly determining each probability value as the estimated interaction response probability of the corresponding candidate benefit for the target user, and stores it in association with the adjusted vector matching degree.

[0084] Optionally, the system can concatenate the user state vector and the rights feature vector into a single high-dimensional vector to construct the model input. The concatenated high-dimensional vector is then input into the interaction behavior prediction model in batches, with the batch size set to improve computational efficiency. Finally, the system receives the probability values ​​output by the model, rounds them to two decimal places, and determines these as the estimated interaction response probabilities. Alternatively, the system can perform feature normalization on the user state vector and the rights feature vector separately to eliminate the influence of dimensions on model prediction. The two vectors are then input into different input layers of the interaction behavior prediction model, and the model's fusion layer performs feature fusion and computation on the two vectors. Finally, the system obtains the probability value from the model's output layer. If the probability value is lower than a preset minimum threshold, it is marked as a low-interaction-probability rights interest, and this probability value is determined as the estimated interaction response probability, completing the probability prediction for all candidate rights interests. It is understood that other methods can also be used to achieve model input and probability acquisition, such as performing feature dimensionality reduction on the vectors before inputting them into the model; this is not limited here.

[0085] S207. Obtain the basic occupancy value corresponding to the inherent resource consumption attribute of each target right from the attribute configuration data, and multiply the basic occupancy value corresponding to each target right by the estimated interaction response probability to obtain the dynamic expected occupancy weight value corresponding to each target right.

[0086] Among them, the inherent resource consumption attribute refers to the inherent characteristics of the rights themselves that determine their resource consumption, such as face value, procurement cost, platform redemption cost, etc.; the basic occupancy value refers to the static resource occupancy value determined according to the inherent resource consumption attribute of the rights, which is the original resource occupancy value of the rights before probability correction.

[0087] Specifically, the system identifies all target rights from the candidate rights set, retrieves the attribute configuration data for each target right one by one, extracts information related to inherent resource consumption attributes from the attribute configuration data, and determines the basic occupancy value corresponding to each target right based on this information, ensuring that the basic occupancy value accurately reflects the inherent resource consumption of the right. Subsequently, the system treats each target right as an independent accounting unit, performs a one-to-one multiplication operation between its corresponding basic occupancy value and the previously calculated estimated interaction response probability, and uses the estimated interaction response probability as a correction coefficient to dynamically adjust the static basic occupancy value.

[0088] For target benefits with a low estimated probability of interaction response, the adjusted dynamic expected resource occupancy weight will be reduced accordingly; for benefits with a high probability, the weight will be closer to the base occupancy value. Finally, the system determines the result of each multiplication operation as the dynamic expected resource occupancy weight for the corresponding target benefit. This weight accurately reflects the actual expected resource occupancy based on the user interaction probability when the platform distributes the benefit.

[0089] S208. Based on the co-occurrence frequency statistics of any two target interests in the candidate interest set in historical data and the feature distance between the classification attributes of the two target interests, calculate the synergistic correlation degree between any two target interests, and determine the synergistic correlation degree as the correlation gain coefficient.

[0090] Among them, the co-occurrence frequency statistics refer to the statistical results of the number of times any two target rights in the candidate rights set appear simultaneously in the platform's historical rights combination packages, which is used to characterize the degree of historical pairing between the two; the feature distance refers to the distance value between the classification attributes of the two target rights in the feature space, which is used to characterize the degree of similarity or difference between the classification attributes of the two; the synergistic association degree refers to the synergistic pairing effect index between the two target rights, which is quantified after comprehensively considering the historical co-occurrence and classification attribute features.

[0091] Specifically, the system retrieves the co-occurrence frequency statistics for each pair of target interests from the platform's historical data set. It then verifies the data's validity, removes outliers, and standardizes the data to ensure comparability of co-occurrence frequencies between different interest pairs. Next, the system maps the classification attributes of the two target interests to the same feature space and calculates the feature distance between their classification attributes. A smaller feature distance indicates greater similarity between the two attributes, while a larger feature distance indicates greater attribute differences. The system then combines the co-occurrence frequency statistics and feature distance, performing a comprehensive calculation according to a pre-defined algorithm. Interest pairs with higher co-occurrence frequencies and more reasonable feature distances have higher collaborative correlation values. Finally, the calculated collaborative correlation value is directly determined as the correlation gain coefficient for that pair of target interests, completing the calculation and storage of correlation gain coefficients between all pairwise target interests.

[0092] Optionally, the system can first normalize the co-occurrence frequency statistics to convert them into a numerical range of 0-1, then perform reverse mapping on the feature distance of the classification attribute to convert the distance value into a similarity value of 0-1; then set preset weights for the normalized co-occurrence frequency value and the feature distance similarity value, and perform a weighted summation on the two; finally, determine the result of the weighted summation as the collaborative correlation degree, that is, the correlation gain coefficient of the target interest.

[0093] It should be noted that for text-based classification attributes (such as 'lifestyle services' and 'digital entertainment'), the system pre-constructs a domain-specific knowledge graph or embedding vector space. After mapping the classification attributes to high-dimensional vectors, the distance between them is calculated using Euclidean distance or cosine distance formulas. For example, 'movie ticket' is closer to 'popcorn' in the vector space, while it is farther from 'gas coupon'. This distance value is the feature distance.

[0094] S209. Set the sum of the correlation gain coefficients between each pair of target equitys in the candidate equity portfolio as the portfolio correlation score.

[0095] Among them, the combination correlation score refers to the quantitative score used to characterize the synergistic effect between all target interests within a single candidate interest portfolio. The higher the score, the stronger the synergy and the more reasonable the combination of interests within the portfolio.

[0096] Specifically, the system identifies a single candidate equity portfolio, extracts all target equitys within that portfolio, and identifies the pairwise pairings of all target equitys within the portfolio. The system then retrieves the association gain coefficient for each pair of equitys from the equity association database, and sequentially sums all association gain coefficients to obtain the total association gain coefficient for the candidate equity portfolio. Finally, the system directly sets this total as the portfolio association score for that candidate equity portfolio. After calculating the portfolio association score for a single portfolio, the same logic is applied to all candidate equity portfolios to calculate and set their portfolio association scores, uniquely associating each portfolio with its corresponding candidate equity portfolio, preparing for the subsequent weighted summation calculation of the final score.

[0097] Optionally, the system can number the target interests within a single candidate interest combination, generate a table of all pairwise pairings based on the numbering to avoid missing pairings; then retrieve the corresponding correlation gain coefficients one by one according to the relationship table, calculate the sum of coefficients by accumulating them one by one; finally, set the accumulated result directly as the combination correlation score of the combination and mark it in the system.

[0098] S210. Set the sum of all vector matching degrees in the candidate equity portfolio as the basic matching score, and perform a weighted sum of the basic matching score and the portfolio association score according to the preset weight coefficient to obtain the score value corresponding to each candidate equity portfolio.

[0099] Among them, the basic matching score refers to the quantitative score used to characterize the basic matching degree between all target benefits within a single candidate benefit combination and the target user. The higher the score, the stronger the individual matching between the benefits within the combination and the user. The weighting coefficient refers to the weight ratio set by the platform according to the benefit distribution strategy for the basic matching score and the combination association score, which is used to balance the proportion of the two sub-scores in the final score.

[0100] Specifically, for a single candidate benefit combination, the system extracts the vector matching degree of all target benefits in the combination for the target user, accumulates all vector matching degrees, and sets the sum as the base matching score for that candidate benefit combination. The system then retrieves preset weight coefficients from the platform, which correspond to the base matching score and the combination association score, respectively, with the sum of their weight coefficients being 1. The platform can adjust the proportion of these two coefficients according to actual business needs. Next, the system multiplies the base matching score by its corresponding weight coefficient, and the combination association score by its corresponding weight coefficient, then sums the two products to obtain the score value corresponding to that candidate benefit combination.

[0101] Finally, following the same logic, the system performs basic matching score calculation and weighted summation calculation on all candidate equity combinations in sequence, obtains the score value of all candidate equity combinations, and sorts them from high to low scores.

[0102] In some embodiments, the system can perform validity checks on the target equity vector matching degree within a single candidate equity portfolio, and remove abnormal matching values; then accumulate the valid matching values ​​to obtain a basic matching score; subsequently, retrieve the fixed weight coefficients preset by the platform, set the basic matching score weight to 0.6, and the portfolio association score weight to 0.4, multiply them respectively, and sum them to obtain the score value; finally, round the score value to two decimal places and record it.

[0103] Optionally, the system can also set personalized weight coefficients for target users based on their profile characteristics. For users with high requirements for matching benefits, the basic matching score weight is increased, and for users with high requirements for the synergy of benefit combinations, the combination association score weight is increased. Then, the basic matching score of the candidate benefit combination is calculated, and the basic matching score and combination association score are multiplied by the personalized weight coefficients respectively and summed to obtain the score value. Finally, the score values ​​of all candidate benefit combinations are standardized and converted into a range of 0-100 for easy and intuitive comparison.

[0104] S211. After pushing the candidate benefit combination with the highest score to the target user, monitor the target user's interaction behavior data of the candidate benefit combination in real time.

[0105] Interactive behavior data refers to the records of relevant behaviors of target users after receiving the push notification package of benefits and the individual benefits within it, including various behaviors such as clicking, claiming, using, closing, and rejecting.

[0106] Specifically, after pushing out the optimal benefit package, the system immediately initiates a real-time monitoring mechanism, setting the monitoring behavioral dimensions and data collection frequency. The behavioral dimensions cover user actions on the entire package and actions on individual benefits within the package. Through the platform's user behavior collection module, the system instantly captures every interaction of the target user, performing real-time analysis and classification of the behavioral data to distinguish between positive, non-interactive, and negative interactions. It also labels each type of behavioral data with time, operation type, and operation object to ensure data integrity and traceability. Simultaneously, the system stores the collected interaction data in real-time to the user behavior database, providing data support for subsequent negative feedback judgments and user state vector corrections. The monitoring process continues until the preset monitoring period ends or a clear negative feedback instruction is captured.

[0107] S212. If the interaction behavior data is a negative feedback instruction, then extract the feature vector of each target interest in the candidate interest portfolio as a negative feedback feature.

[0108] Among them, negative feedback instructions refer to explicit behavioral instructions that target users convey to the system through specific operations to express dissatisfaction or rejection of the pushed benefit package, such as actively refusing to receive it, closing the package and having no interaction for a long time, or giving a negative evaluation of the package; negative feedback features refer to quantitative vectors used to characterize the benefit features that target users dislike, which are a precise extraction of users' negative preferences.

[0109] Specifically, the system comprehensively judges the interactive behavior data monitored in real time, and confirms whether the user's behavior is a clear negative feedback according to the preset negative feedback instruction judgment criteria. If it is determined to be a negative feedback instruction, the negative feedback feature extraction process is initiated.

[0110] The system then identifies the benefit package pushed to the user, extracts all target benefits included in the package, and retrieves the complete benefit feature vector for each target benefit from the benefit feature database, ensuring that the vector's dimension and feature space are consistent with the original user state vector. Finally, the system integrates the feature vectors of all retrieved target benefits as the negative feedback feature corresponding to this user's negative feedback, and labels the negative feedback feature to identify the corresponding user and the pushed benefit package.

[0111] S213. Based on the negative feedback characteristics, the user state vector is orthogonally corrected to remove feature components in the user state vector that are similar to the negative feedback characteristics, so as to obtain the corrected user state vector.

[0112] Orthogonalization correction refers to the process of eliminating or removing components in the user state vector that are consistent with the direction of negative feedback features and have similar features, through orthogonalization methods of linear algebra in the feature space; feature components refer to the eigenvalues ​​of the user state vector in each dimension of the feature space, which are the basic units that constitute the user state vector.

[0113] Specifically, the system maps negative feedback features and the target user's original state vector to the same feature space, ensuring a complete correspondence in their feature dimensions and meanings. Then, the system employs a professional orthogonalization algorithm to calculate the similarity between the user's state vector and the negative feedback features, identifying feature components in the user's state vector similar to the negative feedback features. These components correspond to the user's averse preferences for certain rights and interests. Next, the system uses orthogonalization to eliminate or remove these similar feature components, retaining other feature components in the user's state vector unrelated to the negative feedback features during the elimination process, ensuring that the user's other preference dimensions remain unaffected. Finally, the system normalizes and standardizes the orthogonally-processed vector to obtain a corrected user state vector that more realistically and accurately reflects the target user's real-time rights and interests preferences.

[0114] Optionally, the system can be corrected using Gram-Schmidt process or vector projection. Assume the user state vector is U, and the negative feedback eigenvector is N. The system first calculates the projection component of U in the N direction. The system then performs a subtraction operation. At this point, we obtain... This is the corrected user state vector, which is geometrically perpendicular (or orthogonal) to the negative feedback feature N. This mathematically eliminates the feature dimension that users dislike, while retaining the user's preference information in other dimensions.

[0115] S214. Recalculate the vector matching degree of each candidate benefit with respect to the revised user state vector to generate a new combination of candidate benefits for updated recommendations.

[0116] Specifically, the system uses the corrected user state vector as the new basis for user preferences, retrieving all candidate benefits and their corresponding feature vectors from the benefit pool to ensure that the feature vectors and the corrected user state vectors are in the same feature space. Then, following the same vector similarity calculation method as the original process, the system recalculates the vector matching degree of each candidate benefit relative to the corrected user state vector, fully incorporating the corrected user preferences during the calculation process to avoid benefits similar to negative feedback features. Next, based on the recalculated vector matching degree, the system sequentially completes the steps of candidate benefit set screening, dynamic expected weight calculation, candidate benefit combination generation, and combination score calculation, following the original process's screening rules, resource constraints, and combination scoring logic, ultimately selecting the new candidate benefit combination with the highest score. Finally, the system pushes this new benefit combination package to the target user, completing the updated recommendation, while continuing to monitor user interaction data with the new combination package in real time, forming a closed-loop optimization of the recommendation strategy.

[0117] Optionally, the system can perform feature verification on the corrected user state vector to ensure its validity and accuracy. Then, using the same cosine similarity algorithm as the original process, it can recalculate the vector matching degree of each candidate benefit, while applying an additional penalty coefficient to candidate benefits similar to negative feedback features to further reduce their matching degree. Next, it generates new combinations of candidate benefits according to the complete steps of the original process, calculates the score, and selects the optimal combination. Finally, it pushes the new combination package to the user through the original push channel, along with an updated recommendation notification. It is understood that other methods can also be used to implement updated recommendations; this is not limited here.

[0118] The rights and benefits package intelligent recommendation system of this invention is applied to electronic devices. Figure 3 A schematic diagram of the architecture of an electronic device suitable for implementing embodiments of the present invention is shown.

[0119] It should be noted that, Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0120] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions (computer programs), or by instructions (computer programs) controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. The electronic device of this embodiment includes a storage medium and a processor, wherein the storage medium stores multiple instructions that can be loaded by the processor to execute any step of the method provided in the embodiments of the present invention.

[0121] Specifically, the storage medium and the processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more signal lines. The storage medium stores computer-executable instructions that implement data access control methods, including at least one software functional module that can be stored in the storage medium in the form of software or firmware. The processor executes various functional applications and data processing by running the software program and module stored in the storage medium. The storage medium can be, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The storage medium stores the program, and the processor executes the program after receiving the execution instructions.

[0122] Furthermore, the software programs and modules within the aforementioned storage medium may also include an operating system, which may include various software components and / or drivers for managing system tasks (e.g., memory management, storage device control, power management, etc.) and can communicate with various hardware or software components to provide an operating environment for other software components. The processor may be an integrated circuit chip with signal processing capabilities. The aforementioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc., which can implement or execute the methods, steps, and logic flowcharts disclosed in this embodiment. The general-purpose processor may be a microprocessor or any conventional processor.

[0123] Since the instructions stored in the storage medium can execute the steps in any of the methods provided in the embodiments of the present invention, the beneficial effects of any of the methods provided in the embodiments of the present invention can be achieved, as detailed in the preceding embodiments, and will not be repeated here.

[0124] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligently combining and matching equity portfolios, applied to an intelligent recommendation system for equity portfolios, characterized in that, The method includes: Feature extraction is performed on the historical behavior data of the target user and the attribute configuration data of each candidate right in the right pool to generate user state vector and right feature vector of each candidate right. The attribute configuration data includes the static resource occupation weight of the right, the classification attribute of the domain, and the association attribute of the historical co-occurrence relationship between the rights. Based on the user state vector and the benefit feature vector, calculate the vector matching degree of each candidate benefit for the target user and the estimated probability of interaction response; A candidate equity set is selected from the equity pool based on the vector matching degree, and the candidate equity set includes multiple target equitys; The dynamic expected occupancy weight of each target right is calculated based on the estimated interaction response probability and the static resource occupancy weight of each target right; Based on preset resource constraints, the target rights in the candidate rights set are arranged and combined to generate several candidate rights combinations. The resource constraints include the upper limit of the number of rights in the rights combination package, the mutual exclusion rule of the classification attributes, and the cumulative threshold of the dynamic expected occupancy weight. The score value corresponding to each of the candidate benefit combinations is calculated based on the vector matching degree and the correlation gain coefficient between each target benefit in the candidate benefit combination, and the candidate benefit combination with the highest score value is pushed to the target user. The correlation gain coefficient is determined based on the correlation attribute.

2. The method according to claim 1, characterized in that, The steps of calculating the vector matching degree of each candidate benefit for the target user and estimating the probability of interaction response based on the user state vector and the benefit feature vector specifically include: The user state vector and the feature vector of each candidate right are calculated in the same feature space to obtain the vector matching degree of each candidate right for the target user; The user state vector and the benefit feature vector are input into a pre-trained interaction behavior prediction model, which outputs the probability value of each candidate benefit being effectively interacted with by the target user. The probability value is determined as the estimated interactive response probability.

3. The method according to claim 2, characterized in that, Before the step of inputting the user state vector and the benefit feature vector into the pre-trained interaction behavior prediction model and outputting the probability value of each candidate benefit being executed by the target user in an effective interaction behavior, the method further includes: Based on users' historical behavior data and the attribute configuration data of each historical benefit, a training dataset containing positive and negative samples is constructed. The positive samples are records in which users generate effective interactive responses to the benefits, and the negative samples are records in which users only generate exposure to the benefits but do not generate effective interactive responses. The deep neural network model is trained using the training dataset under supervised learning until the loss function value converges or the preset number of training rounds is reached, at which point training stops, thus obtaining the interactive behavior prediction model.

4. The method according to claim 2, characterized in that, The step of calculating the dynamic expected occupancy weight of each target interest based on the estimated interaction response probability and the static resource occupancy weight corresponding to each target interest specifically includes: Obtain the basic resource consumption value corresponding to the inherent resource consumption attribute of each target benefit from the attribute configuration data; Multiply the base occupancy value corresponding to each target benefit by the estimated interaction response probability to obtain the dynamic expected occupancy weight value corresponding to each target benefit. The dynamic expected occupancy weight value represents the expected resource occupancy of the platform after correction based on the interaction response probability when distributing target benefits to the target user.

5. The method according to claim 1, characterized in that, The step of calculating the score value corresponding to each of the candidate equity combinations based on the vector matching degree and the correlation gain coefficient between each target equity in the candidate equity combination specifically includes: Based on the associated attributes in the attribute configuration data, obtain the co-occurrence frequency statistics of any two target rights in the candidate rights set in historical data; Based on the co-occurrence frequency statistics and the feature distance between the classification attributes of the two target interests, the synergistic correlation degree between any two target interests is calculated, and the synergistic correlation degree is determined as the correlation gain coefficient. Set the sum of all vector matching degrees in the candidate equity portfolio as the basic matching score; The sum of the correlation gain coefficients between each pair of target interests in the candidate interest portfolio is set as the portfolio correlation score; The basic matching score and the combined association score are weighted and summed according to preset weight coefficients to obtain the score value corresponding to each candidate benefit combination.

6. The method according to claim 2, characterized in that, After the steps of calculating the vector matching degree of each candidate benefit for the target user and estimating the probability of interactive response based on the user state vector and the benefit feature vector, the method further includes: Get the remaining distribution quota of each candidate equity in the equity pool and the quota consumption rate within a preset time window; The resource competition coefficient of each candidate interest is calculated based on the remaining allocation quota and the quota consumption rate. The resource competition coefficient is negatively correlated with the remaining allocation quota and positively correlated with the quota consumption rate. A penalty factor is constructed based on the resource competition coefficient, and when the resource competition coefficient is higher than a preset warning value, the vector matching degree of the corresponding candidate interest is reduced based on the penalty factor.

7. The method according to claim 1, characterized in that, Following the step of pushing the candidate benefit combination with the highest rating to the target user, the method further includes: Real-time monitoring of target users' interaction behavior data with the proposed combination of benefits; If the interactive behavior data is a negative feedback instruction, then the feature vector of each target benefit in the candidate benefit combination is extracted as a negative feedback feature; The user state vector is orthogonally corrected based on the negative feedback features to remove feature components in the user state vector that are similar to the negative feedback features, thereby obtaining the corrected user state vector. The vector matching degree of each candidate benefit with respect to the revised user state vector is recalculated to generate a new combination of candidate benefits for updated recommendations.

8. A smart recommendation system for benefit packages, characterized in that, The system includes: one or more processors and memory; The memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is run on the rights and benefits package intelligent recommendation system, the system performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the rights and interests package intelligent recommendation system, the system performs the method as described in any one of claims 1-7.

Images