A corporate image brand promotion system
By collecting data from multiple sources and using deep reinforcement learning, fine-grained user profiles are constructed, solving the problem of the lack of multi-dimensional profiles in existing brand promotion systems and achieving precise and efficient brand promotion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGXI INST OF FASHION TECH
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-05
AI Technical Summary
Existing brand promotion systems lack the ability to deeply mine and personalize user profiles across multiple dimensions, such as user behavior paths and interest preferences, resulting in insufficient traffic statistics to achieve precise targeting.
By collecting data from multiple sources, processing data in a unified manner, constructing heterogeneous graph neural network profiles, and conducting real-time monitoring and deep reinforcement learning, we can generate fine-grained user profiles and deliver brand promotion materials in real time with precision.
It achieves high-precision, fine-grained user profiling, and can adaptively adjust the material version, push time and frequency in millisecond-level feedback to improve the effectiveness of the campaign and avoid overexposure and user churn.
Smart Images
Figure CN122155794A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of enterprise informatization and digital marketing technology, specifically to a corporate image and brand promotion system. Background Technology
[0002] Currently, various online brand promotion and marketing tools have emerged in the market, including social media marketing platforms, search engine optimization and paid search promotion tools, and multi-channel operation systems for official websites and mobile applications. These systems help businesses publish brand content across different channels, track basic traffic data, and manage materials and perform simple performance statistics through their respective platform backends, thereby increasing brand online exposure to a certain extent.
[0003] However, existing technologies still have significant shortcomings: traditional statistics mostly focus on traffic volume and lack the ability to deeply mine and personalize user profiles across multiple dimensions such as user behavior paths and interests. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a corporate image brand promotion system that solves the problem of how to achieve real-time and accurate delivery of brand promotion materials through fine-grained profile generation and deep reinforcement learning.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a corporate image and brand promotion system, comprising: The multi-source data acquisition module is used to collect multi-source heterogeneous data related to users from multiple sources in real time. The data preprocessing and unified management module is used to unify the multi-source heterogeneous data and store it in a centralized data warehouse according to a unified data model. A heterogeneous graph neural network profile building module is used to generate refined user profiles in real time through graph neural network embedding learning. The heterogeneous graph neural network profile building module includes a graph structure generation unit and a graph representation learning unit. The personalized content recommendation and push module is used to match the relevance and prioritize brand promotion materials based on the refined user profile, and push them to target users in real time through the multi-source channels. The real-time monitoring and deep reinforcement learning delivery optimization module is used to perceive user responses in real time during the push process and adaptively adjust materials, time periods and frequencies through a deep reinforcement learning model to dynamically optimize the brand content delivery strategy. The real-time monitoring and deep reinforcement learning delivery optimization module includes a state perception and reward mapping unit, a reinforcement learning strategy decision-making unit and a strategy execution and online update unit.
[0006] Preferably, the multiple sources include the company's official website, mobile client, social media platform and offline interactive terminal, and the multiple heterogeneous data includes access records, interaction behavior data and social relationship data.
[0007] Preferably, the unification includes the following steps: 3.1 Remove missing values and perform outlier detection based on standard scores; 3.2 Map each continuous feature to the [0,1] interval; 3.3 Records are uniquely identified based on user and device identifiers.
[0008] Preferably, the graph structure generation unit is used to construct a heterogeneous graph with multiple types of nodes and multiple types of edges from user nodes, content nodes, interest tag nodes and social relationship nodes through the centralized data warehouse.
[0009] Preferably, the graph representation learning unit iteratively calculates the node embedding according to the following model formula:
[0010] in, Indicates the first Layered user nodes eigenvectors, Represents nodes The set of directly connected neighbor nodes. Represents a set The number of elements, and The weight matrix is a learnable matrix. It is a non-linear activation function. For the updated node Eigenvectors.
[0011] Preferably, the state awareness and reward mapping unit maps user responses to reinforcement learning rewards according to the following instant reward model formula:
[0012] in, For the first Instant rewards for each moment For the first Click-through rate at any given moment For the first Conversion rate at any time For the first User fatigue index at any time and These are the preset weighting coefficients.
[0013] Preferably, the reinforcement learning strategy decision-making unit uses a near-end strategy optimization algorithm to iterate strategies within a multi-action space that includes the version of the push material, the push time period, and the push frequency, and outputs the optimal delivery action.
[0014] Preferably, the strategy execution and online update unit is used to synchronize the optimal delivery action to the personalized content recommendation and push module for execution, and to update the parameters of the reinforcement learning strategy decision unit online according to the latest user response, so as to form a real-time closed-loop optimization.
[0015] This invention provides a corporate image and brand promotion system. It has the following beneficial effects: This corporate image and brand promotion system, through a multi-source data collection module and a unified data preprocessing workflow, can aggregate heterogeneous data from the official website, mobile devices, social media, and offline interactive terminals in real time. The data is then cleaned, standardized, and deduplicated in a centralized data warehouse. The heterogeneous graph neural network profiling module further utilizes graph structure generation and graph representation learning to effectively uncover potential connections between users across multiple dimensions, including behavior, interests, and social relationships. This achieves a high-precision, fine-grained user profile that far surpasses traditional click / browse statistics, thus laying a solid data foundation for subsequent accurate recommendations.
[0016] In the personalized content recommendation and push module, the system intelligently matches and prioritizes content based on fine-grained user profiles. In the real-time monitoring and deep reinforcement learning-based delivery optimization module, it constructs a self-learning closed loop based on the PPO / DQN algorithm through the organic synergy of state awareness and reward mapping, reinforcement learning strategy decision-making, and online updates. This mechanism can adaptively adjust the material version, push time, and frequency in millisecond-level feedback, significantly improving delivery effectiveness. Furthermore, by introducing a fatigue penalty factor into the rewards, it effectively avoids overexposure and user churn, ensuring that brand promotion is both precise and efficient. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating the process of realizing the invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Example 1 like Figure 1As shown, this embodiment of the invention provides a corporate image and brand promotion system, including a multi-source data acquisition module for real-time acquisition of multi-source heterogeneous data related to users from multiple sources. These multiple sources include the company's official website, mobile clients, social media platforms, and offline interactive terminals. The multi-source heterogeneous data includes access records, interaction behavior data, and social relationship data.
[0020] The data preprocessing and unified management module is used to unify multi-source heterogeneous data and store it in a centralized data warehouse according to a unified data model.
[0021] Unification includes the following steps: 3.1 Remove missing values and perform outlier detection based on standard scores.
[0022] 3.2 Map each continuous feature to the interval [0,1].
[0023] 3.3 Records are uniquely identified based on user and device identifiers.
[0024] The heterogeneous graph neural network profile building module is used to generate refined user profiles in real time through graph neural network embedding learning. The heterogeneous graph neural network profile building module includes a graph structure generation unit and a graph representation learning unit.
[0025] The graph structure generation unit is used to construct a heterogeneous graph with multiple types of nodes and edges from user nodes, content nodes, interest tag nodes, and social relationship nodes through a centralized data warehouse.
[0026] The graph representation shows that the learning unit iteratively calculates node embeddings according to the following model formula:
[0027] in, Indicates the first Layered user nodes eigenvectors, Represents nodes The set of directly connected neighbor nodes. Represents a set The number of elements, and The weight matrix is a learnable matrix. It is a non-linear activation function. For the updated node Eigenvectors.
[0028] The specific implementation method is as follows: Data preparation: The system loads cleaned and standardized multi-source heterogeneous data from a centralized data warehouse, including: Number of user nodes: 1000.
[0029] The number of content nodes is 500.
[0030] The number of interest tag nodes is 30.
[0031] Social relationships edge 2000.
[0032] User behavior edge 12000.
[0033] User - tag edge 8000.
[0034] Graph structure generation: The graph structure generation unit organizes all the nodes and edges mentioned above into a heterogeneous graph according to three node types and multiple edge types. User nodes, content nodes, and interest tag nodes are each of the three types of nodes in the graph; social relationships, behavioral relationships, and interest relationships are each of the three types of edges.
[0035] First round of embedded learning: Taking user node U3 as an example, its processing flow is as follows: The initial feature vector of U3 in the first layer contains sixteen values, with the first four dimensions being 0.30, 0.20, 0.10, and 0.05 respectively, and the remaining dimensions taking values in the same way.
[0036] Node U3 connects to twenty neighbors in the graph, including social friends, clicked content nodes, and interest tag nodes. The system automatically calculates the average vector of the features of these twenty neighbors, with the first four dimensions being 0.25, 0.15, 0.05, and 0.02, and the remaining dimensions being calculated similarly.
[0037] The system uses pre-trained weights to weight and sum the features of U3 itself and the average features of its neighbors, resulting in a set of real number vectors. The first four dimensions are approximately 1.20, 0.80, 0.50, and 0.10, and the remaining dimensions are also calculated.
[0038] A non-linear activation process is performed on the real number vector, setting all negative values to zero, and finally generating a new feature vector of the first layer, whose value is the same as or zero as the weighted result.
[0039] Second round of embedded learning: Using the new feature vector from the first layer as input to the second layer, the weighted summation and activation processes described above are repeated to obtain a new feature vector from the second layer. The first four dimensions of this vector are approximately 1.50, 1.00, 0.70, and 0.20, with the remaining dimensions output sequentially. This second-layer vector is the final fine-grained user profile of U3.
[0040] Image application: The system inputs the final profile vectors of all one thousand users into a clustering algorithm, which automatically divides them into eight user groups. Node U3 is assigned to the second group, whose user profiles have the highest vector similarity to the promotional content C102. The system then accurately pushes material C102 to all users in the second group.
[0041] Through the above embodiments, it is clear how this module can utilize specific data to construct heterogeneous graphs and iteratively generate high-precision user profiles in actual operations, providing a solid foundation for personalized recommendations and precise targeting.
[0042] The personalized content recommendation and push module is used to match and prioritize brand promotion materials based on refined user profiles, and push them to target users in real time through multiple sources.
[0043] The real-time monitoring and deep reinforcement learning-based delivery optimization module is used to perceive user responses in real time during the push process and adaptively adjust materials, time periods, and frequencies through a deep reinforcement learning model to dynamically optimize the brand's content delivery strategy. The real-time monitoring and deep reinforcement learning-based delivery optimization module includes a state perception and reward mapping unit, a reinforcement learning strategy decision-making unit, and a strategy execution and online update unit.
[0044] The state awareness and reward mapping unit maps user responses to reinforcement learning rewards based on the following instant reward model formula:
[0045] in, For the first Instant rewards for each moment For the first Click-through rate at any given moment For the first Conversion rate at any time For the first User fatigue index at any time and These are the preset weighting coefficients.
[0046] The reinforcement learning strategy decision-making unit uses a proximal policy optimization algorithm to iterate strategies within a multi-action space that includes push material version, push time period, and push frequency, and outputs the optimal delivery action.
[0047] The strategy execution and online update unit is used to synchronize the optimal delivery action to the personalized content recommendation and push module for execution, and to update the parameters of the reinforcement learning strategy decision unit online based on the latest user response, so as to form a real-time closed-loop optimization.
[0048] The specific implementation method is as follows: Application scenario: Cold start scenario for new creative materials. When a new creative material is launched for the first time, the system has no historical feedback data for the material, and it is necessary to quickly explore the optimal combination of advertising.
[0049] After launching 100 new promotional materials at midnight, the system has yet to provide any historical feedback. The Status Awareness and Reward Mapping unit collected 120 clicks and 8 conversions within 10 seconds, calculating a click-through rate of 10.7%, a conversion rate of 6.7%, and a user fatigue level of 0.1. This unit mapped these values to a low-instantaneous reward signal of strength 0.05 to encourage the exploration of more campaign combinations.
[0050] The reinforcement learning strategy decision-making unit initiates a high-exploration mode within the three-dimensional action space of material version, push time period, and push frequency, setting the exploration rate to 30% to try three different time periods. After three rounds of iteration, it was found that the instant reward for pushing the first type of material between 20:00 and 22:00 reached the highest level of 0.20, and the system established the optimal delivery strategy based on this.
[0051] The strategy execution and online update unit then distributed the first type of material and the push configuration from 20:00 to 22:00. Within 5 minutes, it collected another 230 clicks and 15 conversions. The model automatically lowered the exploration rate to 10% and updated the parameters online with the new data. After 10 minutes, the strategy converged to the current optimal solution.
[0052] Example 2 Unlike Example 1, this example applies to a scenario where user fatigue is increasing.
[0053] After continuously pushing the first type of material for 24 hours, the system captured 400 negative feedbacks and recorded a large number of swiping behaviors. User fatigue level climbed to 70%, click-through rate plummeted from 8% to 3%, and conversion rate dropped from 5% to 2%. The reward mapping unit mapped this negative change into a negative instant reward signal with a strength of -0.15, indicating that the current campaign was at risk of overexposure.
[0054] The reinforcement learning strategy decision unit prioritizes two actions in the action space: reducing the frequency of push notifications and changing the creative version. It reduces the number of daily push notifications from 3 to 1 and replaces the creative with the second type of material. The fatigue penalty weight is increased from 1.0 to 2.0 to enhance the inhibition effect.
[0055] The strategy execution and online update unit immediately implemented the new plan. Within 3 hours, the click-through rate rebounded to 6%, the conversion rate rebounded to 4%, and the fatigue level dropped to 45%. Based on the latest feedback, the model updated the parameters online and gradually restored the regular campaign frequency while continuously monitoring the results.
[0056] Example 3 Unlike Example 1, this example is applied to a scenario of in-depth maintenance of high-value users.
[0057] The system identified 150 high-value users with an average order value exceeding 300 yuan and at least 3 cumulative conversions. The status awareness and reward mapping unit collected feedback from this group separately, statistically analyzing click-through rate (15%), conversion rate (10%), and user fatigue level (20%), mapping continuous positive reward signals to maintain campaign intensity.
[0058] The reinforcement learning strategy decision unit prioritizes "exclusive offers" and "high-value creative ideas" to the highest level, reduces the exploration rate to 5%, and selects to push the third type of material to the user group at 20:00 on Friday night within the three-dimensional action space, setting the daily push frequency to twice.
[0059] After the strategy execution and online update unit immediately distributed the plan, the conversion rate of high-value users rose to 18% within 2 hours. The model updated the parameters in real time based on the new round of feedback and continued to maintain the campaign combination.
[0060] Example 4 Unlike Example 1, this example applies to a time-slot optimization scenario during a major promotional event.
[0061] During the Double Eleven promotion, the system collected user response data in two time slots: 10:00 AM to 12:00 PM and 2:00 PM to 4:00 PM. The click-through rate was 12% and the conversion rate was 7% in the morning, with a fatigue level of 25%. In the afternoon, the click-through rate was 18%, the conversion rate was 12%, and the fatigue level was 28%. The instant reward signal indicates higher returns in the afternoon.
[0062] The reinforcement learning strategy decision unit automatically adjusts the action space to prioritize the afternoon time slot for pushing the fourth type of material, and optimizes the push frequency from once per hour to once every two hours to balance coverage and user tolerance.
[0063] After the strategy execution and online update unit implemented the new plan, the conversion rate was monitored to further increase to 15% within three hours that afternoon. Based on this, the model updated the parameters and solidified the combination of this time period and creative materials into the regular campaign plan for the big promotion.
[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A corporate image and brand promotion system, characterized in that, include: The multi-source data acquisition module is used to collect multi-source heterogeneous data related to users from multiple sources in real time. The data preprocessing and unified management module is used to unify the multi-source heterogeneous data and store it in a centralized data warehouse according to a unified data model. A heterogeneous graph neural network profile building module is used to generate refined user profiles in real time through graph neural network embedding learning. The heterogeneous graph neural network profile building module includes a graph structure generation unit and a graph representation learning unit. The personalized content recommendation and push module is used to match the relevance and prioritize brand promotion materials based on the refined user profile, and push them to target users in real time through the multi-source channels. The real-time monitoring and deep reinforcement learning delivery optimization module is used to perceive user responses in real time during the push process and adaptively adjust the materials, time periods and frequencies through a deep reinforcement learning model. The real-time monitoring and deep reinforcement learning delivery optimization module includes a state perception and reward mapping unit, a reinforcement learning strategy decision-making unit and a strategy execution and online update unit.
2. The corporate image brand promotion system according to claim 1, characterized in that: The multiple sources include the company's official website, mobile client, social media platform and offline interactive terminal, and the multi-source heterogeneous data includes access records, interaction behavior data and social relationship data.
3. The corporate image brand promotion system according to claim 1, characterized in that: The unification includes the following steps: 3.1 Remove missing values and perform outlier detection based on standard scores; 3.2 Map each continuous feature to the [0,1] interval; 3.3 Records are uniquely identified based on user and device identifiers.
4. The corporate image brand promotion system according to claim 1, characterized in that: The graph structure generation unit is used to construct a heterogeneous graph with multiple types of nodes and multiple types of edges from user nodes, content nodes, interest tag nodes, and social relationship nodes through the centralized data warehouse.
5. A corporate image brand promotion system according to claim 1, characterized in that: The graph representation learning unit iteratively calculates node embeddings according to the following model formula: in, Indicates the first Layered user nodes eigenvectors, Represents nodes The set of directly connected neighbor nodes. Represents a set The number of elements, and The weight matrix is a learnable matrix. It is a non-linear activation function. For the updated node Eigenvectors.
6. The corporate image brand promotion system according to claim 1, characterized in that: The state awareness and reward mapping unit maps user responses to reinforcement learning rewards according to the following instant reward model formula: in, For the first Instant rewards for each moment For the first Click-through rate at any given moment For the first Conversion rate at any time For the first User fatigue index at any time and These are the preset weighting coefficients.
7. A corporate image brand promotion system according to claim 1, characterized in that: The reinforcement learning strategy decision-making unit uses a near-end policy optimization algorithm to iterate the strategy within a multi-action space that includes the version of the push material, the push time period, and the push frequency, and outputs the optimal delivery action.
8. A corporate image brand promotion system according to claim 7, characterized in that: The strategy execution and online update unit is used to synchronize the optimal delivery action to the personalized content recommendation and push module for execution, and to update the parameters of the reinforcement learning strategy decision unit online according to the latest user response.