User real-time interaction information-based right commodity configuration method

Abstract

The invention provides a right commodity configuration method based on user real-time interaction information, and belongs to the technical field of interaction information data processing.The right commodity configuration method comprises the steps that firstly, multi-dimensional user behavior feature vectors are constructed, and a complete behavior sequence is formed by capturing user interaction data in real time and combining historical interaction records; modeling a user interaction state through an interaction state neural network in combination with multi-dimensional data such as a user geographic position, a timestamp and equipment information; performing differential weight assignment on the behavior characteristics, and generating an intention intensity matrix through a multi-source data fusion algorithm; inputting the weighted behavior characteristic data into an interactive perception model of a deep bidirectional attention network architecture to generate a right recommendation list; finally, a recommendation strategy is dynamically adjusted according to the real-time response data of the user, optimal distribution is executed, a final right and interest putting scheme is determined, and accurate recognition of actual demands and intentions of the user and efficient configuration of right and interest commodities are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of interactive information data processing, and more particularly, relates to a method for configuring rights and interests commodities based on real-time user interaction information. Background Art

[0002] In the fields of e-commerce and digital marketing, the precise configuration of rights and interests commodities (such as discount vouchers, cash vouchers, full reduction vouchers, and integral vouchers, etc.) is an important means to improve user conversion rates and promote consumption. Traditional methods for configuring rights and interests commodities mainly rely on the historical behavior data of a single user. By collecting information such as the user's browsing history, click behavior, purchase records, etc., a static user portrait and preference model are constructed for the matching and pushing of rights and interests commodities. Such methods have been widely used in e-commerce platforms, membership point systems, and marketing activities, such as recommendation algorithms based on collaborative filtering, rule-based matching systems, and simple user portrait analysis. However, traditional methods for analyzing single-user behavior data have obvious limitations. First, these methods overly rely on historical behavior data and cannot capture the current intention changes and demand transfers of users in real time. Second, static user portraits are difficult to reflect the differentiated needs of users in different scenarios and time points. Third, simple behavior frequency statistics and category matching lack an understanding of the depth and quality of user behavior, resulting in a deviation between rights and interests configuration and the actual needs of users. Especially in the current situation where user behavior is becoming increasingly complex and changeable, single-dimensional and static user behavior analysis can no longer meet the needs of precision marketing. In actual application scenarios, users' demands for rights and interests commodities are often affected by various factors, including situational factors such as time, location, and social environment, as well as deep psychological needs and decision-making intentions. Existing technologies mainly focus on surface behavior data and lack the ability to perceive the deep intentions reflected in the real-time interaction process of users, making it difficult to accurately identify the true needs of users. This has led to the misallocation and waste of rights and interests resources, and the rights and interests commodities obtained by users do not match their actual needs, not only reducing the user experience and conversion rate, but also affecting the marketing effect and resource utilization efficiency of merchants. There is an urgent need for a new method for configuring rights and interests commodities that can effectively identify the actual needs and intentions of users. Summary of the Invention

[0003] In view of this, the present invention provides a method for configuring rights and interests commodities based on real-time user interaction information, which can solve the technical problem in the prior art that the configuration of rights and interests commodities based on single-user behavior data cannot effectively identify the actual needs and intentions of users.

[0004] The present invention is implemented as follows: The present invention provides a method for configuring rights and interests commodities based on real-time user interaction information, including: constructing a multi-dimensional user behavior feature vector; inputting user behavior feature data into an interaction state neural network to generate a user association state score, constructing a user relationship graph, and calculating an associated vector of the same-industry user group; inputting the user relationship graph into a minimum spanning tree algorithm to perform a structural analysis on the same-industry user group, identifying key influencing nodes and calculating the internal rights and interests propagation path of the group, and outputting a user influence index; introducing a user intention recognition module to assign differential weight values to the user behavior feature data to form weighted behavior feature data; establishing an intention intensity scoring system to generate an intention intensity matrix; calculating the similarity between the associated vector of the same-industry user group and the standard associated vector to determine the input weight coefficient and generate a rights and interests recommendation list; and implementing a multi-objective optimization allocation strategy to determine the final rights and interests delivery plan.

[0005] Among them, the step of constructing a multi-dimensional user behavior feature vector is specifically to form a behavior sequence by capturing user interaction data in real time and combining historical interaction records, and obtain user behavior feature data.

[0006] Among them, the step of inputting user behavior feature data into an interaction state neural network is specifically to generate a user association state score by combining user geographical location data, timestamp data, and device information data, construct a user relationship graph according to the user association state score, and calculate an associated vector of the same-industry user group.

[0007] Among them, the interaction state neural network refers to a feedforward neural network structure for analyzing the interaction association between users. The input layer receives user behavior feature data, geographical location data, timestamp data, and device information data. The hidden layer uses a fully connected structure for feature extraction, and the output layer generates a user association state score representing the user association strength.

[0008] Among them, the interaction state neural network is trained using a cosine similarity loss function, and the number of network parameters is controlled within 1 million to ensure real-time inference performance.

[0009] Among them, the associated vector of the same-industry user group refers to a multi-dimensional feature vector representing the association pattern between users, including a geographical distance component, a time synchronization component, an interaction frequency component, an interest overlap component, and a social relationship strength component. The vector dimension is 32, and the value range of each component is from 0 to 1.

[0010] Among them, the standard associated vector refers to an ideal association state vector statistically obtained from the association patterns of historical high-conversion user groups, which is used as a reference benchmark for evaluating the actual user group association state. The vector dimension is the same as that of the associated vector of the same-industry user group.

[0011] Among them, the similarity weight function refers to a non-linear function that maps the cosine similarity between the peer user group association vector and the standard association vector to the input weight coefficient. The function expression is a sigmoid function, and the steepness parameter is optimized and determined by historical conversion data. The output range is from 0.5 to 1.5.

[0012] Among them, the minimum spanning tree algorithm refers to an algorithm for finding the connected subgraph with the minimum total weight in the user relationship graph, which is used to identify the most cost-effective rights and interests dissemination path in the user group. The key influencing node is the node with the highest degree centrality in the minimum spanning tree.

[0013] Among them, the user influence index is calculated by weighted summation of the weights of the connecting edges of the key influencing nodes in the minimum spanning tree.

[0014] The steps of entering the user intention recognition module are specifically to assign different weight values to the user behavior feature data according to the interaction depth value, the stay duration value, and the interaction frequency value, and form weighted behavior feature data.

[0015] The steps of establishing the intention intensity scoring system are specifically to input the weighted behavior feature data into the multi-source data fusion algorithm, calculate the intention scores of the user for different commodity categories, and generate an intention intensity matrix.

[0016] Among them, the multi-source data fusion algorithm refers to a calculation method for integrating the user's explicit behavior data and implicit interest signals. Through tensor decomposition technology, heterogeneous data is mapped to a unified feature space to achieve dimensional consistency and semantic relevance.

[0017] Among them, the rights and interests recommendation list refers to a sorted set of rights and interests commodities generated according to the intention intensity matrix and the commodity attribute matching degree, including key attribute information such as rights and interests type, usage conditions, expiration date, and expected conversion rate, and dynamically adjusts the recommendation priority of the commodities in the rights and interests recommendation list according to the user's real-time response data.

[0018] The steps of implementing the multi-objective optimization allocation strategy are specifically to use the rights and interests recommendation list and the user influence index as inputs, comprehensively consider the individual conversion probability value and the group influence value, and determine the final rights and interests placement plan. The multi-objective optimization allocation strategy refers to a resource allocation method that maximizes both the individual conversion rate and the group influence under the constraint of limited rights and interests resources. The problem is modeled as a multi-objective optimization problem with constraints, and the Pareto optimal solution set is used as the candidate solution, and the comprehensive score is calculated by the weighted summation method.

[0019] Through technical means such as constructing multi-dimensional user behavior feature vectors, real-time interaction state modeling, introducing user intention recognition modules, and multi-source data fusion, the present invention realizes the accurate recognition of users' actual needs and intentions. This method not only focuses on users' historical behavior patterns but also attaches more importance to the deep intention signals shown by users during real-time interactions, providing a more accurate decision-making basis for the configuration of rights and interests products. Compared with traditional technologies, the present invention solves the limitations of single-user behavior data analysis. By capturing user interaction information in real time and combining historical records to form a behavior sequence, it realizes the accurate grasp of users' dynamic intentions. The system can assign differential weight values to user behavior characteristics according to indicators such as interaction depth, dwell time, and interaction frequency, and more precisely understand users' real needs. At the same time, through a multi-source data fusion algorithm, the system integrates users' explicit behaviors and implicit interest signals to generate an intention intensity matrix that comprehensively reflects users' intentions. The interaction perception model adopted by the present invention has a deep bidirectional attention network architecture, which can effectively capture the complex interaction patterns between user behaviors and product attributes, and realize the accurate matching of user intentions and rights and interests products. The system dynamically adjusts the recommendation strategy according to users' real-time response data, ensuring the timeliness and adaptability of rights and interests configuration. Through these innovative technical means, the present invention successfully solves the core technical problem of the inability to effectively identify users' actual needs and intentions in the prior art, and significantly improves the accuracy of rights and interests product configuration and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] To make the purposes, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0022] As Figure 1 shown, it is a flowchart of a method for configuring rights and interests products based on users' real-time interaction information provided by the present invention. This method includes the following steps: S01. Construct multi-dimensional user behavior feature vectors, form a behavior sequence by capturing user interaction data in real time and combining historical interaction records, and obtain user behavior feature data; S02. Input the user behavior feature data into an interaction state neural network, generate a user association state score in combination with user geographical location data, timestamp data, and device information data, construct a user relationship graph according to the user association state score, and calculate the peer user group association vector; S03. Input the user relationship graph into the minimum spanning tree algorithm, perform a structured analysis on the peer user group, identify key influencing nodes, calculate the internal equity propagation path of the group, and output the user influence index; S04. Introduce a user intention recognition module, assign differential weight values to the user behavior feature data according to the interaction depth value, stay duration value, and interaction frequency value to form weighted behavior feature data; S05. Establish an intention intensity scoring system, input the weighted behavior feature data into a multi-source data fusion algorithm, calculate the intention scores of the user for different commodity categories, and generate an intention intensity matrix; S06. Calculate the similarity between the peer user group association vector and the standard association vector, determine the input weight coefficient through the similarity weight function, multiply the intention intensity matrix and the user influence index by the input weight coefficient, and then input them into a pre-trained interaction perception model to generate an equity recommendation list, and dynamically adjust the recommendation priority of the commodities in the equity recommendation list according to the user real-time response data; S07. Execute a multi-objective optimization and allocation strategy, take the equity recommendation list and the user influence index as inputs, comprehensively consider the individual conversion probability value and the group influence value, and determine the final equity delivery plan.

[0023] Among them, the interaction state neural network specifically refers to a feedforward neural network structure for analyzing the interaction correlation between users. The input layer receives user behavior feature data, geographical location data, timestamp data, and device information data. The hidden layer uses a fully connected structure for feature extraction. The output layer generates a user association state score representing the user association strength. The interaction state neural network is trained using a cosine similarity loss function, and the number of network parameters is controlled within 1 million to ensure real-time inference performance.

[0024] Among them, the peer user group association vector specifically refers to a multi-dimensional feature vector representing the association pattern between users, including geographical distance component, time synchronization component, interaction frequency component, interest overlap component, and social relationship strength component. The vector dimension is 32, and the value range of each component is from 0 to 1, which is calculated through the user association state score.

[0025] Among them, the standard association vector specifically refers to an ideal association state vector statistically obtained from the association patterns of historical high-conversion user groups, which is used as a reference benchmark for evaluating the actual user group association state. The vector dimension is the same as that of the peer user group association vector.

[0026] Among them, the similarity weight function specifically refers to a non-linear function that maps the cosine similarity between the peer user group association vector and the standard association vector to an input weight coefficient. The function expression is a sigmoid function, and the steepness parameter is optimized and determined by historical conversion data. The output range is from 0.5 to 1.5, and the higher the similarity, the larger the output weight coefficient.

[0027] Among them, the minimum spanning tree algorithm specifically refers to an algorithm for finding the connected subgraph with the minimum total weight in the user relationship graph, which is used to identify the most cost-effective rights and interests dissemination path in the user group. The edge weight represents the intensity of the influence relationship between users, and the node weight represents the influence of users on group decisions. The key influence node is the node with the highest degree centrality in the minimum spanning tree, and the user influence index is calculated by weighted summation of the weights of the connecting edges of the key influence node in the minimum spanning tree.

[0028] Among them, the multi-source data fusion algorithm specifically refers to a calculation method for integrating user explicit behavior data and implicit interest signals. Through tensor decomposition technology, heterogeneous data is mapped to a unified feature space to achieve dimensional consistency and semantic relevance. The weighted behavior feature data is input, and the intention intensity matrix is output.

[0029] Among them, the rights and interests recommendation list specifically refers to a sorted set of rights and interests products generated according to the intention intensity matrix and the commodity attribute matching degree, including key attribute information such as rights and interests type, usage conditions, expiration date, and expected conversion rate. The rights and interests type includes discount coupons, vouchers, full reduction coupons, and integral coupons. The usage conditions include the applicable commodity range, minimum consumption amount, and usage time period limit. The expiration date includes the start time, end time, and number of valid days after activation. The expected conversion rate is obtained through statistical analysis of historical data.

[0030] Among them, the multi-objective optimization allocation strategy specifically refers to a resource allocation method that maximizes both individual conversion rate and group influence under the constraint of limited rights and interests resources. The problem is modeled as a multi-objective optimization problem with constraints, and the Pareto optimal solution set is used as the candidate solution. The comprehensive score is calculated by the weighted summation method, and the solution with the highest comprehensive score is selected as the final rights and interests placement plan.

[0031] Among them, the specific structure of the interaction perception model is a deep bidirectional attention network architecture, which consists of a user behavior encoding layer, a commodity attribute encoding layer, a cross-modal interaction layer, and a multi-head self-attention layer. The core mechanism is a bidirectional attention mechanism and an adaptive feature fusion module. The number of attention heads is dynamically adjusted according to the dimension of the peer user group association vector, and the depth of feature fusion is determined by the user association status score. The slope parameter of the activation function is determined by the mean value of the intention intensity matrix.

[0032] The steps for establishing the training dataset of the interaction perception model specifically include collecting multi-scenario user interaction behavior data, including browsing trajectory data, clickstream data, dwell time data, interaction operation data, and conversion event data, extracting user individual features and group association features, constructing a positive sample set containing successful rights and interests conversion cases and corresponding pre-interaction sequences, constructing a negative sample set containing failed rights and interests push cases and corresponding pre-interaction sequences, splitting the training set, validation set, and test set in chronological order, cleaning, normalizing, and augmenting the training data, and generating a multi-modal training sample library containing sequence features, graph structure features, and label information.

[0033] The steps for training the interaction perception model specifically include randomly initializing the model network parameters, using the Xavier initialization method to set the initial values of the weights, designing a joint loss function to optimize the two objectives of individual conversion prediction and group influence prediction simultaneously, adopting a hierarchical learning rate strategy to apply different update strengths to different network layers, implementing the batch gradient descent algorithm for parameter optimization, applying an early stopping strategy to avoid overfitting, periodically evaluating the performance metrics of the validation set, including the prediction accuracy of the individual conversion rate and the group response consistency, dynamically adjusting the attention mechanism parameters and the feature fusion intensity, freezing the parameters of the best-performing model as a pre-trained model for online deployment, and designing a model compression strategy to achieve lightweightness to meet the real-time inference requirements.

[0034] The following describes the specific implementation manners of the above steps in detail.

[0035] The specific implementation manner of step S01 is to obtain user interaction behavior data through multiple acquisition channels and construct a multi-dimensional behavior feature vector. First, collect explicit interaction data such as user clicks, swipes, and dwells, and record the original data including interaction type, interaction object, interaction duration, interaction frequency, etc.; then, combine the user's historical interaction records to construct a behavior sequence according to the time series, and set the sequence length to the interaction behavior within the most recent 48 hours; next, perform feature extraction on the behavior sequence, including behavior frequency statistics, time interval analysis, and behavior conversion path identification; subsequently, apply a sequence embedding algorithm to convert the behavior sequence into a fixed-dimensional vector, and set the vector dimension to 128; finally, perform normalization processing to unify the range of each dimension feature value to 0-1, forming standardized user behavior feature data. The sequence embedding algorithm adopted in this step is based on the long short-term memory network, which can effectively capture the temporal dependence of user behavior and extract the feature representation containing the user's latent intention, laying a data foundation for subsequent analysis.

[0036] The specific implementation of step S02 is to construct a user relationship graph and calculate the group association vector. First, splice the user behavior feature data with the geographical location data, timestamp data, and device information data to form an input vector. Then, input the input vector into an interaction state neural network, which includes an input layer, three fully connected hidden layers, and an output layer. The number of neurons in the hidden layers is 512, 256, and 128 respectively, and the ReLU activation function is used. Next, obtain the user association state score from the network output. The score range is 0 to 100, indicating the association strength between users. Subsequently, construct a user relationship graph based on the user association state score. The nodes in the graph represent users, and the edges represent the association relationships between users. The edge weights are determined by the association state score, and only the edges with a score greater than 30 are retained. Finally, calculate the group association vector of peer users and extract a 32-dimensional vector including geographical distance components, time synchronization components, interaction frequency components, interest overlap components, and social relationship strength components. This step utilizes the powerful expression ability of the neural network to effectively mine the implicit association patterns between users and provides key data support for subsequent group influence analysis.

[0037] The specific implementation of step S03 is to analyze the user group structure using the minimum spanning tree algorithm. First, use the user relationship graph constructed in step S02 as the input, and define the edge weight as 100 minus the user association state score to ensure that the higher the association strength, the smaller the weight. Then, apply the Kruskal algorithm or Prim algorithm to construct a minimum spanning tree and find the subgraph that connects all users with the minimum total weight. Next, calculate the degree centrality of each node in the spanning tree. The degree centrality calculation formula is the number of edges connected to the node divided by the maximum possible number of connections. Subsequently, identify the node with the highest degree centrality as the key influence node, and set the threshold to 0.6, that is, the nodes with a degree centrality greater than 0.6 are regarded as key influence nodes. Finally, calculate the user influence index by weighted summing the weights of the edges connected to the key influence nodes. The weight coefficient is inversely proportional to the level of the edge in the spanning tree. The higher the level, the lower the weight coefficient, and the influence index range is normalized to 0 to 100. This step effectively identifies the key influence nodes in the group through graph theory algorithms, quantitatively evaluates the user influence, and lays a foundation for the subsequent optimal allocation of rights and interests resources.

[0038] The specific implementation of step S04 is to introduce a user intention recognition module to perform weighted processing on user behavior feature data. First, three key indicators, namely interaction depth value, stay duration value, and interaction frequency value, are extracted from the user behavior feature data; then a weight assignment mode is designed, where the weight range of interaction depth is 1.2 - 1.8, the weight range of stay duration is 0.8 - 1.5, and the weight range of interaction frequency is 1.0 - 1.6; then specific weight values are set based on the behavior type, with higher weights assigned to direct conversion behaviors such as adding to the shopping cart and favoriting, and lower weights assigned to browsing behaviors; subsequently, the corresponding weight values are applied to the behavior feature data of each dimension, and the weighting formula is the original value of the feature multiplied by the corresponding weight value; finally, the weighted results are normalized to ensure the rationality of the proportions of different feature dimensions, forming weighted behavior feature data. This step highlights the high-value interaction behaviors of users through differential weight assignment, enhances the ability to identify the true intentions of users, and improves the accuracy of subsequent recommendations.

[0039] The specific implementation of step S05 is to establish an intention intensity scoring system. First, the weighted behavior feature data obtained in step S04 is used as input, in combination with multi-source data such as the user's historical purchase records, search keywords, and favorited products; then a multi-source data fusion algorithm is applied. This algorithm is based on tensor decomposition technology and maps heterogeneous data to a unified 64-dimensional feature space; then reliability coefficients are set for data from different sources, with the coefficient for historical purchase behavior data being 0.8 - 1.0, the coefficient for browsing behavior data being 0.3 - 0.6, and the coefficient for search behavior data being 0.5 - 0.8; subsequently, the intention scores of the user for each product category are calculated, with the score range being 0 - 10, and the score formula being the sum of the contribution scores of each data source multiplied by the reliability coefficient; finally, an intention intensity matrix is generated, with the matrix dimension being the number of users multiplied by the number of product categories, and the matrix elements being the intention scores of the corresponding users for the corresponding product categories. This step comprehensively evaluates the intention intensity of users through multi-source data fusion technology, generates a structured intention representation, and provides data support for accurate rights and interests matching.

[0040] The specific implementation of step S06 is to generate and dynamically adjust the rights and interests recommendation list. First, calculate the cosine similarity between the associated vector of the peer user group and the standard associated vector. The cosine similarity calculation formula is the inner product of the two vectors divided by the product of the norms of the two vectors; then determine the input weight coefficient through the similarity weight function. The function uses a sigmoid-type function, and the expression is , where k is the steepness parameter with a value range of 5 to 10, x is the cosine similarity, and the output range is 0.5 to 1.5; then multiply the intention intensity matrix and the user influence index by the input weight coefficients respectively; subsequently, input the weighted data into a pre-trained interaction perception model, which adopts a deep bidirectional attention network architecture; finally, generate an interest recommendation list, and adjust the recommendation priority according to real-time user response data such as click-through rate, redemption rate, sharing rate, etc., with an adjustment frequency of once every 30 minutes. The priority increase amplitude is proportional to the positive degree of the response, and the maximum increase amplitude is 30%. This step enhances the response sensitivity of the model to high-quality groups through a similarity-driven weight adjustment mechanism, and at the same time combines real-time feedback for dynamic optimization to improve the adaptability of the recommendation system.

[0041] The specific implementation of step S07 is to execute a multi-objective optimization allocation strategy. First, use the interest recommendation list and the user influence index as input data; then set optimization goals, including maximizing individual conversion rate and maximizing group influence diffusion in two dimensions; then define constraint conditions, including total interest quantity limit, maximum interest quantity limit per user, interest type diversity requirements, etc.; subsequently, model the problem as a multi-objective optimization problem with constraints and convert it into a single-objective problem using the weighted summation method, with a weight ratio of 0.6 for individual conversion rate and 0.4 for group influence; finally, solve the optimal allocation plan through a genetic algorithm, with the population size set to 100 and the number of iterations to 50 - 100. The selection operation uses the roulette wheel method, with a crossover probability of 0.8 and a mutation probability of 0.1. Finally, select the individual with the highest fitness as the interest placement plan. This step balances the two dimensions of individual conversion and group influence through multi-objective optimization technology, realizes efficient resource allocation, and improves the overall marketing effect.

[0042] It should be noted that through in-depth exploration of user group interaction correlation information, the present invention has achieved a qualitative breakthrough in the configuration of rights and interests commodities. The interaction state neural network established by the system not only analyzes the behavior of a single user, but also pays more attention to the dynamic interaction patterns among users. By constructing a multi-dimensional associated vector of peer user groups through multi-dimensional features such as geographical location, time synchronization, and interest overlap, the internal influence conduction mechanism within the group is accurately characterized. The combination of the user relationship graph and the minimum spanning tree algorithm enables the system to identify the key influence nodes in the group decision-making process, depict the optimal propagation path of the rights and interests value within the group, and realize the mode transformation from point-to-point push to point-driven surface propagation. The similarity weight function innovatively compares the actual group association pattern with the historical high-conversion standard associated vector, dynamically adjusts the intensity of the recommendation strategy, and enables the system to have the ability to adapt to changes in the group structure. The bidirectional attention mechanism and the adaptive feature fusion module of the interaction perception model can simultaneously process individual demand signals and group influence evaluation, and the number of attention heads is flexibly adjusted according to the dimension of the group associated vector to ensure accurate capture of the influence path in a complex social network structure. The multi-objective optimization allocation strategy takes the individual conversion probability and group influence as dual objectives, searches for the resource allocation balance point through the Pareto optimal solution set, and realizes the social network amplification effect of the rights and interests value. This group interaction perception-based recommendation method significantly improves the accuracy and coverage efficiency of rights and interests placement, preferentially allocates limited rights and interests resources to key node users with the potential for fission propagation, forms a self-reinforcing group marketing ecosystem, significantly enhances the recommendation accuracy, reduces ineffective rights and interests placement, improves the overall marketing efficiency, and at the same time maximizes the transmission of the rights and interests value through the influence diffusion of key nodes, enabling the limited resources to generate a geometric amplification effect and forming a sustainable positive marketing cycle.

[0043] Specifically, the principle of the present invention is as follows: The core technical principle of the present invention to solve the problem of identifying the actual needs and intentions of users lies in constructing a rights and interests commodity configuration framework that integrates multi-dimensional feature extraction, real-time interaction state modeling, and deep intention recognition. Through systematic technical means, this framework has achieved a breakthrough in user intention recognition from static to dynamic and from the surface to the deep layer.

[0044] At the level of user feature representation, the present invention has broken through the traditional isolated behavior analysis mode. By capturing user interaction data in real time and combining historical records to form a behavior sequence, a multi-dimensional user behavior feature vector containing time series information is constructed. This dynamic feature representation method can more comprehensively depict the change trajectory of the user behavior pattern and provides a rich information basis for intention recognition. Different from traditional methods, the present invention not only pays attention to the occurrence frequency of behaviors, but also focuses on the quality characteristics of behaviors. By assigning different weights to indicators such as interaction depth, stay duration, and interaction frequency, the system can distinguish the difference in intention intensity behind different behaviors and more accurately understand the true needs of users.

[0045] At the level of interactive state modeling, the present invention innovatively proposes an interactive state neural network, which fuses user behavior features with context data such as geographical location, timestamp, and device information, achieving an all-round perception of the user's interactive state. Through the training mechanism of the cosine similarity loss function, the network can effectively capture the intention differences of users in different contexts, overcoming the defect of insufficient consideration of context factors in traditional methods. This method of interactive state modeling that integrates context information enables the system to understand the changing needs of users under certain spatio-temporal conditions, greatly improving the accuracy of intention recognition.

[0046] At the level of intention recognition and matching, the present invention establishes an intention intensity scoring system, inputs weighted behavior feature data into a multi-source data fusion algorithm, and generates an intention intensity matrix representing the degree of interest of users in different commodity categories. The multi-source data fusion algorithm uses tensor decomposition technology to map the explicit behavior data and implicit interest signals of users into a unified feature space, achieving dimensional consistency and semantic relevance, and being able to comprehensively capture the multi-level intention information of users. The interactive perception model adopts a deep bidirectional attention network architecture, and through the coordinated action of the user behavior encoding layer, commodity attribute encoding layer, cross-modal interaction layer, and multi-head self-attention layer, accurately analyzes the complex interaction patterns between user behavior and commodity attributes, and realizes the optimal matching of user intentions and rights and interests commodities.

[0047] In terms of the dynamic adjustment mechanism, the present invention dynamically adjusts the recommendation priorities of commodities in the rights and interests recommendation list according to the real-time response data of users, forming a closed-loop feedback system. This dynamic adjustment mechanism enables the system to timely correct the deviation of intention recognition, continuously optimize the rights and interests allocation strategy, and ensure the adaptability and robustness of the system. Through the design of the joint loss function and the application of the hierarchical learning rate strategy, the interactive perception model can continuously improve the accuracy of intention recognition during the continuous learning process.

[0048] In summary, the present invention constructs a rights and interests commodity allocation system that can effectively identify the actual needs and intentions of users through the organic combination of technical means such as multi-dimensional feature representation, real-time interactive state modeling, deep intention recognition, and dynamic adjustment mechanism, fundamentally solves the core problems in the prior art, and realizes the precision and personalization of rights and interests commodity allocation.

[0049] The following provides a specific Embodiment 1 of the present invention, and the specific implementation manners of each step in this Embodiment 1 are described in detail as follows.

[0050] The specific implementation of step S01 is to obtain user interaction behavior data through multiple acquisition channels and construct a multi-dimensional behavior feature vector. First, explicit interaction data such as user clicks, swipes, and stays are collected, and the original data including interaction type, interaction object, interaction duration, interaction frequency, etc. is recorded; then, combined with the user's historical interaction records, a behavior sequence is constructed according to the time series, and the sequence length is set to the interaction behavior within the most recent 48 hours; then, feature extraction is performed on the behavior sequence, including behavior frequency statistics, time interval analysis, and behavior conversion path recognition; subsequently, a sequence embedding algorithm is applied to convert the behavior sequence into a fixed-dimensional vector. The sequence embedding algorithm is based on a long short-term memory network, and for the input behavior sequence , where represents the th behavior event, and the feature of each behavior event is represented as , where represents the interaction type encoding, represents the interaction object encoding, represents the interaction duration, represents the interaction frequency. The sequence embedding calculation formula is as follows: ; ; In the formula, represents the hidden state of the long short-term memory network at time step , with a dimension of 128; represents the calculation function of the long short-term memory network; represents the finally obtained behavior sequence embedding vector, with a dimension of 128; represents the fully connected layer mapping function, which maps the average pooled hidden state to the feature space. Finally, normalization processing is performed, and the normalization formula is: ; In the formula, represents the normalized feature vector, with a value range of 0 to 1; and respectively represent the minimum and maximum values in the vector . This step converts the time-series behavior data into a fixed-dimensional feature representation through sequence embedding technology, effectively capturing the time-series dependence of user behavior and laying a data foundation for subsequent analysis.

[0051] The specific implementation of step S02 is to construct a user relationship graph and calculate the group association vector. First, the user behavior feature data is concatenated with the geographical location data, timestamp data, and device information data to form an input vector , which is specifically represented as: ; In the formula, is the input vector after feature splicing, with a dimension of 162; is the normalized user behavior feature vector obtained in step S01, with a dimension of 128; is the geographical location data vector, with a dimension of 10; is the timestamp data vector, with a dimension of 8; is the device information data vector, with a dimension of 16. Then the input vector is passed into the interaction state neural network, which is a feedforward neural network structure, including an input layer, three fully connected hidden layers, and an output layer. The forward propagation calculation process of the network is as follows: ; ; ; ; ; ; ; ; In the formula, , , , are the weight matrices of each layer respectively; , , , are the bias vectors of each layer respectively; are the linear transformation outputs of each layer respectively; are the outputs of each hidden layer after passing through the activation function respectively; is the rectified linear unit activation function, defined as ; is the sigmoid activation function, defined as ; is the user association state score, ranging from 0 to 100. Then, according to the user association state score, a user relationship graph is constructed, where represents the set of user nodes, represents the set of user association relationship edges, represents the set of edge weights. The construction condition of the edge set is: ; In the formula, represents user and user The correlation status score between them, with the threshold set at 30, that is, only the edges with a correlation status score greater than 30 are retained. Finally, the peer user group correlation vector is calculated , with a dimension of 32 and containing five types of components. The calculation method is as follows: ; ; ; ; ; ; In the formula, is the geographical distance component, with a dimension of 8; represents the user and the user The geographical distance between them, in meters; represents the maximum considered distance, set to 10000 meters; is the time synchronization component, with a dimension of 6; and respectively represent the activity timestamps of the user and the user ; is the time decay coefficient, set to 3600 seconds; is the interaction frequency component, with a dimension of 6; represents the interaction frequency of the user ; is the frequency threshold, set to 10 times per hour; is the interest overlap component, with a dimension of 6; and respectively represent the interest sets of the user and the user ; is the social relationship strength component, with a dimension of 6; represents the correlation status score between the user and the user . In this step, the neural network is used to calculate the user correlation status score, construct the user relationship graph, and extract the multi-dimensional group correlation feature vector, providing key data support for the subsequent user group influence analysis.

[0052] The specific implementation of step S03 is to analyze the user group structure using the minimum spanning tree algorithm. First, take the user relationship graph constructed in step S02 as the input, and define the edge weight conversion function: ; In the formula, is the converted edge weight; is the user and the user is the association status score between them. This conversion ensures that the higher the association strength, the smaller the corresponding edge weight. Then the Kruskal algorithm is applied to construct a minimum spanning tree , where and , and the algorithm process is as follows: (1) Sort all edges in ascending order of weight ; (2) Initially, each node forms an independent connected component; (3) Traverse each edge in ascending order of weight , if node and node are not in the same connected component, then add this edge to the minimum spanning tree , and merge these two connected components; (4) Repeat step (3) until the minimum spanning tree contains edges.

[0053] Next, calculate the degree centrality of each node in the spanning tree , and the calculation formula is: ; In the formula, represents the degree centrality of node ; represents the degree of node in the minimum spanning tree , that is, the number of edges connected to node ; represents the total number of nodes. Subsequently, identify the node with the highest degree centrality as the key influencing node, and the judgment condition is: ; In the formula, represents the set of key influencing nodes; the threshold is set to 0.6, that is, nodes with a degree centrality greater than 0.6 are regarded as key influencing nodes. Finally, calculate the user influence index , and the calculation formula is: ; ; In the formula, represents the influence index of user ; the summation range is all edges directly connected to user ; is the weight coefficient, related to the level of node in the spanning tree They are related, and the higher the level, the lower the weight coefficient. is the edge weight after conversion. Finally, the influence index is normalized to the range of 0-100: ; In the formula, is the influence index of the user after normalization . and respectively represent the minimum and maximum values of the influence index among all users. This step effectively identifies the key influence nodes in the group through graph theory algorithms, quantitatively evaluates the user influence, and lays a foundation for the subsequent optimal allocation of rights and interests resources.

[0054] The specific implementation of step S04 is to introduce a user intention recognition module to perform weighted processing on the user behavior feature data. First, three key indicators, namely the interaction depth value, the stay duration value, and the interaction frequency value, are extracted from the user behavior feature data to form a feature vector : ; In the formula, represents the interaction depth feature sub-vector; represents the stay duration feature sub-vector; represents the interaction frequency feature sub-vector. Then, a weight allocation vector is designed: ; In the formula, represents the interaction depth weight sub-vector, and its value range is 1.2-1.8; represents the stay duration weight sub-vector, and its value range is 0.8-1.5; represents the interaction frequency weight sub-vector, and its value range is 1.0-1.6. The specific weight values are determined according to the behavior type, and the calculation formula is: ; ; ; In the formula, represents the importance score of the th behavior type, and the range is 0-10. For direct conversion behaviors such as adding to the shopping cart and collecting, the score is relatively high, and for browsing behaviors, the score is relatively low; represents the standard score of the stay duration of the th behavior, and the range is 0-10; represents the standard score of the frequency of the th behavior, and the range is 0-10. Subsequently, the corresponding weight values are applied to the behavior feature data of each dimension to calculate the weighted feature vector : ; In the formula, represents the Hadamard product operator, that is, element-wise multiplication; represents the weighted feature vector. Finally, the weighted result is normalized, and the calculation formula is: ; In the formula, represents the normalized weighted feature vector, that is, the weighted behavior feature data; and respectively represent the minimum and maximum values in the vector . This step highlights the high-value interaction behaviors of users through differential weight allocation, enhances the ability to identify the true intentions of users, and improves the accuracy of subsequent recommendations.

[0055] The specific implementation of step S05 is to establish an intention intensity scoring system. First, the weighted behavior feature data obtained in step S04 is used as the input, and combined with multi-source data such as the user's historical purchase records, search keywords, and favorite products, to form a multi-source feature matrix : ; In the formula, represents the multi-source feature matrix; represents the historical purchase behavior feature vector; represents the search behavior feature vector; represents the favorite behavior feature vector. Then, a multi-source data fusion algorithm is applied. This algorithm is based on tensor decomposition technology and maps heterogeneous data to a unified feature space. The tensor decomposition process is expressed as: ; In the formula, represents the original multi-source data tensor; represents the core tensor; , , respectively represent the factor matrices of the three dimensions; represents the tensor product operation along the th dimension. Through tensor decomposition, the heterogeneous data is mapped to a unified 64-dimensional feature space, and a unified representation vector is obtained. Then, a reliability coefficient vector is set for data from different sources: ; In the formula, represents the reliability coefficient of the historical purchase behavior data, and its value range is 0.8 - 1.0; Represents the reliability coefficient of browsing behavior data, with a value range of 0.3 to 0.6; Represents the reliability coefficient of search behavior data, with a value range of 0.5 to 0.8. Subsequently, calculate the intention score matrix of each commodity category for the user , and the calculation formula is: ; In the formula, Represents the user 's intention score for the commodity category , and the score range is 0 to 10; Represents the eigenvalue of the user for the commodity category on the data source ; Represents the reliability coefficient of the data source ; Represents the total number of data sources. Finally, generate the intention intensity matrix , that is, the intention score matrix after normalization: ; In the formula, Represents the intention intensity matrix; and respectively represent the minimum value and the maximum value in the matrix . This step uses multi-source data fusion technology to comprehensively evaluate the user intention intensity, generate a structured intention representation, and provide data support for accurate rights and interests matching.

[0056] The specific implementation method of step S06 is to generate and dynamically adjust the rights and interests recommendation list. First, calculate the cosine similarity between the peer user group association vector and the standard association vector , and the calculation formula is: ; In the formula, represents the cosine similarity between the vector and the vector ; represents the inner product of the two vectors; and respectively represent the norm of the vector and the vector . Then, determine the input weight coefficient through the similarity weight function. The function uses a sigmoid-type function, and the expression is: ; In the formula, Denotes the input weight coefficient, with a value range of 0.5 to 1.5; Denotes the steepness parameter, with a value of 5 to 10, which is optimized and determined according to historical conversion data; Denotes the cosine similarity. Then the intent strength matrix and the user influence index are respectively multiplied by the input weight coefficient : ; ; In the formula, Denotes the weighted intent strength matrix; Denotes the weighted user influence index. Subsequently, the weighted data is input into a pre-trained interaction perception model to obtain an initial rights and interests recommendation list , and each rights and interests item contains the following attributes: rights and interests type, usage conditions, expiration date, expected conversion rate, etc. Finally, the recommendation priority is dynamically adjusted according to the user's real-time response data, and the adjustment formula is: ; In the formula, Denotes the adjusted priority of the rights and interests item ; Denotes the initial priority; Denotes the adjustment coefficient, with a maximum value of 0.3; Denotes the real-time response index of the rights and interests item , such as click-through rate, redemption rate, sharing rate, etc.; Denotes the average response index of all rights and interests items; Denotes the standard deviation of the response index. The adjustment frequency is once every 30 minutes, and the increase in priority is proportional to the positive degree of the response. This step enhances the response sensitivity of the model to high-quality groups through a similarity-driven weight adjustment mechanism, and at the same time combines real-time feedback for dynamic optimization to improve the adaptability of the recommendation system.

[0057] The specific implementation method of step S07 is to execute a multi-objective optimization and allocation strategy. First, the rights and interests recommendation list and the user influence index are used as input data; then the optimization objectives are set, including the individual conversion rate maximization objective function and the group influence diffusion maximization objective function : ; ; In the formula, Denotes the rights and interests allocation plan, Denotes whether to allocate the rights and interests Allocated to the user , taking values of 0 or 1; Indicates the user 's expected conversion rate for the rights and interests ; Indicates the influence index of the user ; Indicates the propagation coefficient of the rights and interests , which is related to the type and attributes of the rights and interests; Indicates the number of users; Indicates the number of types of rights and interests. Then define the constraints: ; ; ; ; In the formula, Indicates the total limit of the number of rights and interests; Indicates the maximum limit of the number of rights and interests per user, with a default value of 3; Indicates the maximum limit of the number of users to whom a single right and interest can be allocated; Indicates the th minimum allocation quantity of the rights and interests of the category to ensure the diversity of the types of rights and interests. Subsequently, the problem is modeled as a multi-objective optimization problem with constraints and transformed into a single-objective problem using the weighted summation method: ; Satisfy all the above constraints In the formula, Indicates the comprehensive objective function; and respectively represent the weights of the individual conversion rate objective and the group influence objective, with default values of 0.6 and 0.4 respectively. Finally, the optimal allocation scheme is solved through a genetic algorithm, and the algorithm parameters are set as follows: the population size is 100, the chromosome encoding adopts binary encoding, and each position represents the corresponding value; the selection operation adopts the roulette wheel method, and the selection probability is proportional to the fitness; the crossover operation adopts single-point crossover, and the crossover probability is 0.8; the mutation operation adopts bit flip mutation, and the mutation probability is 0.1; the termination condition is to reach the maximum number of iterations of 50 - 100 or no improvement in the optimal solution for 10 consecutive generations. The fitness function is defined as: ; In the formula, Indicates the fitness value of the individual ; It represents a penalty function for constraint violation, which is proportional to the degree of violation. Finally, the individual with the highest fitness is selected as the rights and interests investment plan. This step uses multi-objective optimization technology to balance the two dimensions of individual transformation and group influence, achieve efficient resource allocation, and improve the overall marketing effect.

[0058] In this embodiment, the detailed structure of the interaction state neural network is as follows: This network is a feedforward neural network for analyzing the interaction association patterns among users. The input layer consists of 4 parts: 128 user behavior feature data input units, 10 geographical location data input units, 8 timestamp data input units, and 16 device information data input units. The total number of neurons in the input layer is 162. The hidden layer adopts a fully connected structure and consists of 3 layers. The number of neurons in the first hidden layer is 512, the second hidden layer is 256, and the third hidden layer is 128. They are fully connected between layers, and the activation function uses the ReLU function. The number of neurons in the output layer is 1, which is used to generate a user association state score in the range of 0 to 100. The activation function of the output layer uses the sigmoid function and is linearly scaled to match the target range. The network is trained using the cosine similarity loss function, and the expression is , where is the predicted value, is the true value. To ensure real-time inference performance, the number of network parameters is strictly controlled within 1 million. The specific parameter calculation is as follows: The first layer weight matrix has 162×512 = 82944 parameters, the first layer bias has 512 parameters, the second layer weight matrix has 512×256 = 131072 parameters, the second layer bias has 256 parameters, the third layer weight matrix has 256×128 = 32768 parameters, the third layer bias has 128 parameters, the output layer weight matrix has 128×1 = 128 parameters, and the output layer bias has 1 parameter, totaling 247809 parameters. The network training uses the batch gradient descent algorithm. The initial value of the learning rate is set to 0.001, and the cosine annealing strategy is used to dynamically adjust the learning rate. The batch size is set to 64, and the number of training epochs is 100.

[0059] In this embodiment, the detailed steps for establishing the training dataset of the interaction state neural network are as follows: First, determine the target user group, and select users with high activity and rich behavioral data as the initial sample set. The number of users is not less than 50,000, covering people of different age groups, regions, and consumption capabilities to ensure sample diversity. Then, collect user interaction behavior data, including raw data such as click behavior, browsing behavior, purchase behavior, and social behavior for at least 3 months, with no less than 100 behavior records for each user. Next, collect user geographical location data, including permanent residence locations, activity trajectories, travel radii, etc., ensuring that the location accuracy is within 100 meters. Subsequently, collect timestamp data to record temporal information such as the time points, durations, and periodic characteristics of user activities, with the time accuracy controlled at the second level. At the same time, collect device information data, including technical parameters such as device types, operating systems, network environments, and application versions. Then, clean and preprocess the collected raw data, removing outliers, missing values, and noise data. The abnormal determination criterion is data points that exceed 3 standard deviations of the normal distribution, and features with a missing rate exceeding 20% are subject to filling or removal operations. Next, construct label data, generating user association state scores through expert annotation or automatic generation based on rules. The annotation rules include weighted calculations of multi-dimensional indicators such as interaction frequency, content similarity, and spatio-temporal overlap, generating association strength scores within the range of 0 to 100. Then, perform feature engineering, reducing the dimension, combining, and transforming the original features to generate a more expressive feature set. The dimension reduction method uses principal component analysis, retaining the principal components that can explain 85% of the variance. The feature combination uses cross-feature generation technology, and the feature transformation uses normalization and standardization processing. Subsequently, divide the dataset into a training set, a validation set, and a test set in chronological order, with a ratio of 8:1:1, avoiding data leakage problems caused by random partitioning. Finally, perform data augmentation, expanding the training samples through interpolation, sampling, adding slight noise, etc., to enhance the generalization ability of the model. The size of the augmented training set is expanded to 1.5 times the original data. The entire process of constructing the training dataset strictly follows the data quality control standards, regularly checking data consistency and monitoring distribution drift to ensure the reliability and effectiveness of the data basis for model training.

[0060] In this embodiment, the detailed structure of the interaction perception model is as follows: The model adopts a deep bidirectional attention network architecture, and the core design concept is to capture the complex interaction relationship between user behaviors and product attributes. The model consists of four main components: a user behavior encoding layer, a product attribute encoding layer, a cross-modal interaction layer, and a multi-head self-attention layer. The user behavior encoding layer uses a bidirectional long short-term memory network structure with 256 hidden units. The input is the user behavior sequence, and the output is the user behavior representation vector. The product attribute encoding layer uses a multi-layer perceptron structure with 3 layers, and the number of neurons in each layer is 128, 64, and 32 respectively. The input is the product attribute features, and the output is the product attribute representation vector. The cross-modal interaction layer realizes the bidirectional attention interaction between the user behavior representation and the product attribute representation. The number of attention heads is dynamically adjusted according to the dimension of the peer user group association vector, and the basic setting is 8 attention heads. The multi-head self-attention layer further extracts the interaction features with 2 layers, and the number of attention heads is the same as that of the cross-modal interaction layer. The model also includes an adaptive feature fusion module, and the fusion depth is determined by the user association state score. The score range of 0-30 corresponds to 1-layer fusion, 31-70 corresponds to 2-layer fusion, and 71-100 corresponds to 3-layer fusion. The model adopts a multi-task learning framework to optimize two objectives, namely individual conversion prediction and group influence prediction, simultaneously. The loss function is the weighted sum of the losses of the two tasks, and the weight ratio is 0.7 for individual conversion prediction and 0.3 for group influence prediction. The total number of model parameters is about 5 million, and it is reduced to less than 1 million through model compression techniques such as knowledge distillation and weight quantization to meet the real-time inference requirements.

[0061] In this embodiment, the detailed steps for establishing the training data set of the interaction perception model are as follows: first, collect multi-scenario user interaction behavior data, including browsing trajectory data, click stream data, dwell time data, interactive operation data and conversion event data for at least 6 months, with a data volume of no less than 1 million interaction records and covering no less than 100,000 users; then clean the original data, remove invalid records, outliers and duplicate data, and retain more than 90% of valid data after cleaning; then extract individual user characteristics, including demographic characteristics, behavioral habit characteristics, consumption capacity characteristics, etc., with a characteristic dimension of no less than 50 dimensions; then construct a user relationship network, and calculate based on indicators such as interaction co-occurrence, content overlap, and spatiotemporal consistency. The strength of association between users is used to screen user pairs with association strength greater than the threshold of 0.3 to form relationship edges. Then, positive and negative sample sets are constructed according to the interaction results. The positive samples are successful equity conversion cases and corresponding pre-interaction sequences, and the negative samples are failed equity push cases and corresponding pre-interaction sequences. The ratio of positive and negative samples is controlled at 1:3. Then, the data set is divided into training set, validation set and test set in chronological order, with a ratio of 70%, 15% and 15% respectively. The training data is then enhanced, including sequence truncation, noise addition, feature masking and other methods to increase sample diversity. Finally, a multimodal training sample library containing sequence features, graph structure features and label information is generated. The sample format is a structured tensor, which is convenient for batch training of models. The training process adopts a hierarchical learning rate strategy, with different learning rates set for different network layers. The learning rate of the underlying feature extraction layer is relatively low, at 0.0005, and the learning rate of the high-level interaction layer is relatively high, at 0.001. An early stopping strategy is implemented, and training is stopped if there is no improvement in the performance of the validation set for five consecutive rounds. The validation set performance indicators are periodically evaluated, including an individual conversion rate prediction accuracy target of no less than 85% and a group response consistency target of no less than 80%. Finally, the model parameters with the best overall performance on the validation set are selected as the pre-trained model for online deployment.

[0062] To better understand and implement the present invention, an embodiment 2 of a specific application scenario of the present invention is provided below: A research and development team developed a writing interaction application that integrates psychological analysis and health services. On the surface, this application is for calligraphy practice, but in fact, it infers the user's mental state and living habits by analyzing the behavioral patterns during the user's writing process, and then configures personalized health product rights and interests. The team selected 50,000 active users for a 3-month test. The ages of the users ranged from 18 to 65 years old, covering students, working professionals, housewives, and retirees. The system recorded multi-dimensional behavioral data of the users during the writing process, such as stroke characteristics, pause duration, writing speed, modification frequency, completion rate, and font style preference. These data have been proven to be highly correlated with the user's stress level, concentration state, emotional fluctuations, and decision-making style. Each user completed at least 100 writing exercises, and each exercise lasted from 5 to 15 minutes. After the original behavioral data collected by the system was cleaned, step S01 of the application was used to construct a 128-dimensional user behavioral feature vector. Different dimensions in the vector represent different mental state characteristics. For example, the 1st - 20th dimensions represent the stress level, the 21st - 40th dimensions represent the concentration ability, the 41st - 60th dimensions represent the emotional stability, the 61st - 80th dimensions represent the decision-making style, the 81st - 100th dimensions represent the living regularity, and the 101st - 128th dimensions represent the social preference. Table 1 shows some key features extracted from the user writing behavior data and their mental health implications: Table 1 Correspondence Table between Writing Behavior Features and Mental Health

[0063] Applying step S02 of the application, the system combines the user's geographical location data, usage time period data, and device information data, and calculates the association status score between users through an interaction state neural network. In actual applications, the association status scores between users with similar living and working rhythms and similar mental health conditions are generally relatively high, with an average score of 72.8, while the average association score between random user pairs is 25.3. The system constructed a user relationship graph containing 50,000 nodes and approximately 350,000 edges based on these scores, and calculated a 32-dimensional peer user group association vector. Table 2 shows the average values of the association vector components of user groups with different mental states: Table 2 Main Components of the Association Vector of User Groups with Different Mental States

[0064] Applying step S03, the system identifies key influential nodes in each user group through the minimum spanning tree algorithm. In the high-stress group, the key influential nodes are mostly users with strong health awareness who actively seek improvement; in the mood swing group, the key influential nodes are mainly composed of users who are socially active and willing to share experiences. The system calculates the influence index for each user. The results show that about 8.3% of the users have an influence index exceeding 75, and these users play the role of leaders in promoting a healthy lifestyle in their respective groups. Applying steps S04 and S05, the system performs weighted processing on the user behavior characteristic data and establishes an intention intensity scoring system. The system infers the degree of demand for different types of health products based on the user's writing behavior patterns, including stress relief products, sleep improvement products, attention enhancement products, mood regulation products, etc. The calculation results of the intention intensity matrix show that the average intention score of the high-stress group for stress relief products is 8.9, and the average intention score for sleep improvement products is 7.8; while the average intention score of the mood swing group for mood regulation products is 9.2, and the average intention score for nutritional supplement products is 6.5. Applying steps S06 and S07, the system generates a personalized list of health product benefits recommendations for different users based on the user intention intensity matrix and the influence index, and implements a multi-objective optimization allocation strategy. Table 3 shows the comparison of the benefit delivery effects for users with different psychological states: Table 3 Comparison of Health Product Benefit Delivery Effects for Users with Different Psychological States

[0065] Compared with the traditional health product recommendation method based on direct expression of user needs and historical purchase analysis, the present invention has achieved a significant improvement through the indirect method of analyzing psychological states by writing behavior. The traditional method mainly relies on users' explicit search, browsing records, and questionnaires. Users often are unwilling or unable to accurately express their mental health needs, resulting in limited recommendation accuracy, with an average benefit conversion rate of 31.5% and a group influence diffusion rate of only 18.3%. While the present invention, through the non-invasive means of analyzing writing behavior, combined with user relationship graphs and identification of key influential nodes, not only accurately captures users' potential psychological needs, but also fully considers the influence propagation characteristics of user groups. The average benefit conversion rate has been increased to 47.9%, the group influence diffusion rate has been increased to 38.3%, and at the same time, the user health improvement feedback rate is as high as 64.9%. The present invention avoids the limitations of user self-reporting in the traditional method, realizes a deep understanding of users' health needs through behavior analysis, effectively activates the group dissemination of the concept of a healthy lifestyle while meeting the needs of individual users, and achieves the precise and efficient allocation of health product benefit resources.

[0066] Example 3: A research and development team applied the method of the present invention to configure rights and interests products based on the interaction behaviors of users in group chats. The team selected 2,000 active shopping communication groups on the platform, with a total of approximately 150,000 users. The group sizes varied from small interest groups of 20 people to large category groups of 5,000 people. The system collected behavioral data such as message sending, content interaction, product sharing, and evaluation feedback in group chats in real time through the API interface. At the same time, combined with the browsing records, search histories, and purchase behaviors of users, a multi-dimensional user behavior feature vector was constructed. The data collection period was 6 months, and each user generated approximately 800 interaction records on average. Table 4 shows the main group chat behavior characteristics collected by the system: Table 4 Group Chat Interaction Behavior Characteristics Table

[0067] Applying step S01, the system constructed a 128-dimensional behavior feature vector based on the group chat behavior sequence of users, where each dimension represents the characteristics of different interaction types and interaction objects. Applying step S02, the system combined the user's geographical location data, active time data, and device information data, and calculated the correlation status score between users through the interaction status neural network. The experimental data showed that the average correlation status score between users who often interacted in the same group chat was 83.2, while the average correlation score for random user pairs across groups was 12.7. Based on these scores, the system constructed a user relationship graph and calculated the group association vector of peer users. Table 5 shows the comparison of the main components of the group association vectors of different types of shopping groups: Table 5 Comparison Table of Group Association Vector Components of Different Types of Shopping Groups

[0068] Applying step S03, the system analyzes the group structure through the minimum spanning tree algorithm, identifies key influencing nodes, and calculates the user influence index. The results show that in the mother and baby products group, active senior mom users usually become key influencing nodes, and the influence index is generally higher than 85; in the digital product group, professional evaluation sharers become key influencing nodes, and the average influence index is 79.3. Applying steps S04 and S05, the system performs weighted processing on the user behavior characteristic data and establishes an intention intensity scoring system. During the experiment, it is found that the speech content of users in group chats has a strong correlation with the shopping intention. For example, users who frequently ask about the usage experience of a certain type of product usually have an intention score of more than 7.5 for that type of product. Applying step S06, the system calculates the similarity between the association vectors of each type of group and the standard association vector, and determines the input weight coefficient through the similarity weight function. The similarity of the mother and baby products group is 0.87, and the corresponding weight coefficient is 1.38; while the similarity of the fresh food group is 0.62, and the corresponding weight coefficient is 0.95. The system generates a rights and interests recommendation list based on the intention intensity matrix and user influence index after weight adjustment, and makes dynamic adjustments according to the real-time responses of users. Table 6 shows some examples of rights and interests recommendations generated by the system: Table 6 Example Table of Rights and Interests Recommendations Based on Group Chat Behavior

[0069] Applying step S07, the system executes a multi-objective optimization and allocation strategy, comprehensively considering the individual conversion probability and group influence, and determines the final rights and interests delivery plan. During the actual execution process, the system sets different weight ratios for each group type. For example, the group influence weights of the mother and baby group and the beauty group are relatively high (0.5), while the individual conversion weight of the digital group is relatively high (0.7). After a 3-month experiment, the comparison results show that the rights and interests delivery plan generated by the method of the present invention has achieved remarkable results in various groups. Table 7 shows the effect comparison with the traditional method: Table 7 Effect Comparison Table of Rights and Interests Delivery

[0070] Traditional methods for allocating rights and interests products mainly rely on the behavioral data of individual users and simple demographic characteristics for recommendation, ignoring the interaction relationships and influence differences of users in social networks. This method has obvious deficiencies in dealing with group chat scenarios and is difficult to identify and utilize the opinion leader effect and information dissemination path in groups. The method of the present invention effectively captures the complex user interaction patterns and opinion dissemination mechanisms in the group chat environment by constructing a user relationship graph, identifying key influencing nodes, and calculating the associated vectors of peer user groups. Experimental results show that the method of the present invention has significantly improved compared with traditional methods in terms of rights and interests utilization rate, driven conversion rate, group activity improvement, and return on investment (ROI), with average improvement amplitudes reaching 58.7%, 74.3%, 84.6%, and 63.2% respectively. The method of the present invention not only improves the allocation efficiency of rights and interests products, but also amplifies the group dissemination effect of rights and interests investment by accurately identifying and motivating key influencing users, providing a powerful tool for precision marketing of social e-commerce platforms.

[0071] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Tables 8, 9, 10, and 11 below.

[0072] Table 8 Variable Explanation Table (First Part)

[0073] Table 9 Variable Explanation Table (Second Part)

[0074] Table 10 Variable Explanation Table (Third Part)

[0075] Table 11 Variable Explanation Table (Fourth Part)

[0076] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention.

Claims

1. A method for configuring rights and interests commodities based on users' real-time interaction information, characterized in that, Including: Construct a multi-dimensional user behavior feature vector; input user behavior feature data into an interaction state neural network to generate a user association state score, construct a user relationship graph, and calculate the peer user group association vector; input the user relationship graph into a minimum spanning tree algorithm to perform a structural analysis on the peer user group, identify key influencing nodes and calculate the internal equity propagation path within the group, and output the user influence index; Introduce a user intention recognition module to assign differential weight values to user behavior feature data to form weighted behavior feature data; establish an intention intensity scoring system to generate an intention intensity matrix; Calculate the similarity between the peer user group association vector and the standard association vector, determine the input weight coefficient, and generate an equity recommendation list; Execute a multi-objective optimization allocation strategy to determine the final equity placement plan.

2. The method according to claim 1, wherein The steps for constructing a multi-dimensional user behavior feature vector are specifically to form a behavior sequence by capturing user interaction data in real time and combining historical interaction records to obtain user behavior feature data.

3. The method according to claim 2, wherein The steps for inputting user behavior feature data into an interaction state neural network are specifically to generate a user association state score by combining user geographical location data, timestamp data, and device information data, construct a user relationship graph based on the user association state score, and calculate the peer user group association vector.

4. The method according to claim 3, wherein The interaction state neural network refers to a feedforward neural network structure used to analyze the interaction associations between users. The input layer receives user behavior feature data, geographical location data, timestamp data, and device information data. The hidden layer uses a fully connected structure for feature extraction, and the output layer generates a user association state score representing the user association strength.

5. The method according to claim 4, wherein The interaction state neural network is trained using a cosine similarity loss function.

6. The method according to claim 3, wherein The peer user group association vector refers to a multi-dimensional feature vector representing the association pattern between users, including geographical distance components, time synchronization components, interaction frequency components, interest overlap components, and social relationship strength components. The vector dimension is 32, and the value range of each component is from 0 to 1.

7. The method according to claim 6, characterized in that, The standard association vector refers to an ideal association state vector statistically obtained from the association patterns of historical high-conversion user groups, serving as a reference benchmark for evaluating the actual user group association state. The vector dimension is the same as that of the peer user group association vector.

8. The method according to claim 7, characterized in that, The similarity weight function refers to a non-linear function that maps the cosine similarity between the peer user group association vector and the standard association vector to the input weight coefficient. The function expression is a sigmoid function, and the steepness parameter is optimized and determined by historical conversion data. The output range is from 0.5 to 1.

5.

9. The method according to claim 1, characterized in that, The minimum spanning tree algorithm refers to an algorithm for finding the minimum-weight connected subgraph in a user relationship graph, used to identify the most cost-effective equity propagation path in a user group. The key influencing node is the node with the highest degree centrality in the minimum spanning tree.

10. The method according to claim 9, wherein The user influence index is calculated by weighted summation of the weights of the connecting edges of the key influencing nodes in the minimum spanning tree.

Images