An intelligent content generation and accurate distribution method based on AIGC

Abstract

This invention discloses an intelligent content generation and precise distribution method based on AIGC, comprising the following steps: constructing user individual feature vectors; constructing user association graphs; performing community detection processing to obtain user groups and calculate group distribution parameters, and constructing corresponding group semantic resonance fields; calculating group condition vectors, obtaining generation prompt information, and constructing conditional input sequences; using the AIGC model to generate target content consistent with the semantic features of the corresponding user groups; performing semantic projection calculation to obtain content semantic vectors, and calculating group fit and individual matching degree; generating distribution results for different user individuals; collecting user feedback data on the distribution results, and updating user individual feature vectors, user association graphs, and group semantic resonance fields, thereby realizing integrated collaborative processing of content generation and content distribution.

Description

Technical Field

[0001] This invention relates to the field of content generation models and content distribution optimization, and in particular to an intelligent content generation and accurate distribution method based on AIGC. Background Technology

[0002] With the continuous expansion of internet content production, the methods of information dissemination to different user groups are gradually shifting from large-scale distribution of uniform content to differentiated generation and targeted distribution to specific audiences. Especially in application scenarios such as content marketing, news push, knowledge services, intelligent customer service, education and tutoring, and multimedia dissemination, how to quickly generate content that matches user needs based on user characteristics and accurately distribute the generated content to target users has become an important research direction in the field of intelligent information processing. In recent years, generative artificial intelligence technology has developed rapidly. AIGC technology, based on large language models, multimodal generative models, and related neural network architectures, has already possessed strong capabilities in generating text, images, and composite content, thus providing a new technological foundation for intelligent content production.

[0003] In existing technologies, a common approach is to first generate standardized content based on preset templates, manual rules, or general generation models, and then distribute the content to different users through recommendation systems, ranking systems, or tag matching systems. Another approach involves introducing user profiles, behavioral sequences, interest tags, or contextual scenarios into the recommendation system to match and rank existing content, thereby improving the distribution hit rate. Yet another approach attempts to incorporate some user attribute information into the content generation stage, causing the generated results to be biased towards specific groups to some extent. However, from the perspective of the overall technical chain, most existing methods still treat content generation and content distribution as two relatively independent processing stages: the former is responsible for "what content to generate," and the latter for "who to see the content." While this staged processing approach can accomplish basic functions, it still has significant shortcomings in terms of group-level semantic expression, generation constraint mechanisms, and distribution consistency. Summary of the Invention

[0004] One objective of this invention is to propose an AIGC-based intelligent content generation and precise distribution method. This invention fully utilizes user individual feature modeling, graph structure analysis, group semantic resonance field modeling, and generative artificial intelligence technologies to achieve collaborative ranking and individual-level adjustment for user groups and individuals. It also combines feedback data to collaboratively update user individual feature vectors, user association graphs, and group semantic resonance fields. This invention has the advantages of high integration of content generation and distribution, strong group semantic expression capabilities, high controllability of the generation process, and high distribution accuracy.

[0005] According to an embodiment of the present invention, a method for intelligent content generation and precise distribution based on AIGC includes the following steps: Collect and preprocess user behavior data, user attribute data, and contextual environment data to construct individual user feature vectors; Construct a user association graph based on individual user feature vectors; Community detection processing is performed on the user association graph to obtain user groups and calculate group distribution parameters, and construct the corresponding group semantic resonance field; The group condition vector is calculated based on the group semantic resonance field to obtain the generated prompt information, and the generated prompt information is fused with the group condition vector through semantic coupling to construct the condition input sequence; The conditional input sequence is input into the AIGC model, and the generation process is constrained based on the conditional input sequence to generate target content that is consistent with the semantic features of the corresponding user group. Semantic projection is performed on the target content to obtain the content semantic vector. The group fit is calculated based on the content semantic vector and the group semantic resonance field. The individual matching degree is calculated based on the content semantic vector and the user's individual feature vector. Based on group adaptability and individual matching degree, the target content is sorted within the corresponding user group, and the target content is adjusted at the individual level according to the individual matching degree to generate distribution results for different individual users; Collect feedback data from individual users on the distribution results, and update the individual user feature vector, user association graph, and group semantic resonance field based on the feedback data.

[0006] Optionally, the collected user behavior data, user attribute data, and contextual environment data are subjected to unified data alignment processing. Data from different sources are associated and integrated according to user identifiers, missing values ​​are filled, outliers are removed, and numerical normalization is performed. User behavior data is serialized and encoded to extract behavioral pattern features. User attribute data is categorically encoded to form attribute features. Contextual environment data is temporally extracted to form scene features. The behavioral features, attribute features, and scene features are concatenated and fused, and dimensionality reduction and nonlinear mapping processing are performed to obtain a unified dimension user individual feature vector.

[0007] Optionally, the construction of the user association graph specifically includes: Calculate the feature similarity between two users based on their individual feature vectors. Each user is mapped to a node in the user association graph. The connection relationship is determined based on the feature similarity between two users. When the feature similarity between two users is greater than or equal to the preset similarity threshold, an edge is established between the two corresponding nodes. When the feature similarity between two users is less than the preset similarity threshold, no edge is established between the two corresponding nodes. Assign edge weights to the established edges, and generate a user association graph based on all nodes, the established edges, and the edge weights of each edge.

[0008] Optionally, the construction of the collective semantic resonance field specifically includes: Community detection processing is performed on the user association graph, and all individual users are divided according to the edges and edge weights in the user association graph to obtain user groups; The group distribution parameters are calculated based on the individual feature vectors of all users within the user group. The group distribution parameters include the group center vector and the group discrete matrix. Calculate the group center vector; Calculate the discrete matrix of the population; Construct the corresponding group semantic resonance field based on the group center vector and the group discrete matrix.

[0009] Optionally, the construction of the conditional input sequence specifically includes: Calculate the group condition vector based on the group semantic resonance field of the corresponding user group; Obtain the generated prompt information and convert the generated prompt information into an input vector sequence; Calculate the semantic coupling coefficient based on each input vector in the input vector sequence and the group conditional vector; Based on the semantic coupling coefficient, the input vector sequence is semantically coupled with the group condition vector to construct a conditional input sequence.

[0010] Optionally, the generation of the target content specifically includes: The conditional input sequence is input into the AIGC model, and the conditional input sequence is represented as a vector sequence formed by arranging the conditional input vectors in the input order, with each conditional input vector corresponding to an input position; Based on the order of the conditional input vectors in the conditional input sequence, the generation state corresponding to the AIGC model is calculated sequentially. Calculate the probability distribution of the content unit corresponding to each input position based on the generation state corresponding to each input position; The content unit corresponding to each input position is determined based on the probability distribution of the content unit corresponding to each input position. Output a sequence of content units according to the order of the content units corresponding to each input position, and then concatenate the content unit sequence in order to obtain target content that is consistent with the semantic features of the corresponding user group.

[0011] Optionally, the calculation of the group fitness and individual fitness specifically includes: Semantic projection is performed on the target content to obtain the content semantic vector; Calculate group fit based on content semantic vectors and group semantic resonance fields; The individual matching degree is calculated based on the content semantic vector and the user's individual feature vector.

[0012] Optionally, the generation of the distribution result specifically includes: Based on the distribution of target content within the corresponding user group, read the group fit degree corresponding to the target content and the individual matching degree corresponding to each user within the corresponding user group, and calculate the distribution score for each user within the corresponding user group. Based on the distribution score corresponding to each individual user, the target content is sorted within the corresponding user group to obtain the sorting results; Based on the individual matching degree, the target content is adjusted at the individual level to obtain the individual adjustment vector corresponding to each user. Generate individual-level adjusted content semantic vectors based on individual adjustment vectors; Distribution results are generated based on the sorting results and the individual-level adjusted semantic vectors of each user.

[0013] Optionally, the updating of the individual user feature vector, user association graph, and group semantic resonance field specifically includes: For each individual user within the corresponding user group, feedback data on the distribution results is collected. The feedback data includes click feedback data, dwell feedback data, and conversion feedback data. The click feedback data, dwell feedback data, and conversion feedback data corresponding to each individual user are numerically processed to obtain the feedback value corresponding to each individual user. Update the individual user feature vector based on the feedback value corresponding to each individual user; Update the user association graph based on the updated individual user feature vectors; Based on the updated individual user feature vectors, the group distribution parameters corresponding to each user group are recalculated, and the corresponding group semantic resonance field is updated based on the recalculated group distribution parameters. An updated group semantic resonance field is constructed based on the updated group center vector and the updated group discrete matrix.

[0014] The beneficial effects of this invention are: This invention processes user behavior data, user attribute data, and contextual environment data to construct individual user feature vectors and further builds user association graphs. Based on this, community detection processing is performed on the user association graphs to obtain user groups and calculate group distribution parameters. This leads to the construction of a corresponding group semantic resonance field. This allows the content generation process to no longer rely solely on single user features or static tag information, but rather to characterize the relationships and common semantic distribution states between multiple users at the user group level. Therefore, this invention effectively improves the ability to express the common needs of target user groups, overcoming the problems of insufficient utilization of user relationship features, lack of group semantic information, and weak content generation basis in existing technologies.

[0015] This invention calculates a group condition vector based on a group semantic resonance field to obtain generation prompt information. It then fuses the generation prompt information with the group condition vector through semantic coupling to construct a conditional input sequence. This sequence is then input into an AIGC model to constrain the generation process, thereby generating target content consistent with the semantic features of the corresponding user group. Compared to existing technologies that simply concatenate user information into the generation input or perform secondary filtering after generation, this invention introduces a user-group-oriented constraint mechanism at the content generation stage. This ensures that the generated results maintain a higher consistency with the corresponding user group in terms of semantic direction, content expression, and target orientation, thereby improving the controllability, relevance, and generation quality of the target content.

[0016] This invention performs semantic projection calculation on target content to obtain content semantic vectors. It then calculates group fit based on the content semantic vectors and the group semantic resonance field, and individual matching based on the content semantic vectors and individual user feature vectors. This allows content evaluation to move beyond a single user level and simultaneously achieve both group and individual level evaluation capabilities. Furthermore, based on group fit and individual matching, the target content is sorted and adjusted at the individual level. This ensures that the distribution results satisfy both the overall semantic consistency of the corresponding user group and the differentiated needs of individual users. Therefore, it effectively improves the accuracy, stability, and hit rate of content distribution, addressing the problems of existing technologies that rely solely on a single matching indicator, lack sufficient group consistency, and are inadequately matched to individuals.

[0017] After distribution is completed, this invention collects feedback data from individual users regarding the distribution results and updates the individual user feature vectors, user association graphs, and group semantic resonance fields based on this feedback data, thereby establishing a complete closed loop from user modeling, group modeling, content generation, content distribution to feedback updates. This closed-loop mechanism enables the system to continuously correct individual user characteristics, user group structure, and group semantic expression based on actual distribution results, thereby continuously improving the consistency and adaptability of subsequent content generation and distribution. Compared with existing technologies that only update recommendation parameters or local interest tags, this invention has stronger dynamic optimization capabilities and system synergy, maintaining better adaptability to changes in user interests, scenarios, and content requirements.

[0018] This invention achieves organic linkage between user group modeling, group semantic expression, AIGC controlled generation, two-layer adaptive computation, and feedback collaborative updating. It not only improves the integration of content generation and content distribution, but also enhances the adaptability of generated content to target user groups and individual target users. Therefore, it has the beneficial effects of strong content generation targeting, high distribution accuracy, strong ability to characterize group commonalities, complete feedback loop, and strong system sustainable optimization capability. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an AIGC-based intelligent content generation and precise distribution method proposed in this invention. Figure 2 This is a schematic diagram illustrating the construction of a group semantic resonance field for an AIGC-based intelligent content generation and precise distribution method proposed in this invention. Figure 3 This is a schematic diagram illustrating the construction of target content for an AIGC-based intelligent content generation and precise distribution method proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A method for intelligent content generation and precise distribution based on AIGC includes the following steps: Collect and preprocess user behavior data, user attribute data, and contextual environment data to construct individual user feature vectors; Construct a user association graph based on individual user feature vectors; Community detection processing is performed on the user association graph to obtain user groups and calculate group distribution parameters, and construct the corresponding group semantic resonance field; The group condition vector is calculated based on the group semantic resonance field to obtain the generated prompt information, and the generated prompt information is fused with the group condition vector through semantic coupling to construct the condition input sequence; The conditional input sequence is input into the AIGC model, and the generation process is constrained based on the conditional input sequence to generate target content that is consistent with the semantic features of the corresponding user group. Semantic projection is performed on the target content to obtain the content semantic vector. The group fit is calculated based on the content semantic vector and the group semantic resonance field. The individual matching degree is calculated based on the content semantic vector and the user's individual feature vector. Based on group adaptability and individual matching degree, the target content is sorted within the corresponding user group, and the target content is adjusted at the individual level according to the individual matching degree to generate distribution results for different individual users; Collect feedback data from individual users on the distribution results, and update individual user feature vectors, user association graphs, and group semantic resonance fields based on the feedback data.

[0022] In this embodiment, the collected user behavior data, user attribute data, and contextual environment data are uniformly aligned. Data from different sources are associated and integrated according to user identifiers, missing values ​​are filled, outliers are removed, and numerical normalization is performed. User behavior data is serialized and encoded to extract behavioral pattern features. User attribute data is categorically encoded to form attribute features. Contextual environment data is temporally extracted to form scene features. Behavioral features, attribute features, and scene features are concatenated and fused, and dimensionality reduction and nonlinear mapping are performed to obtain a unified dimension user individual feature vector.

[0023] In this embodiment, the construction of the user association graph specifically includes: Calculate the feature similarity between two users based on their individual feature vectors. The specific process for calculating the feature similarity between two individual users is as follows: Read the feature vectors corresponding to the two individual users respectively, and assign each feature value to the corresponding dimension according to the same feature dimension. Multiply the feature values ​​of each corresponding dimension in the feature vectors of the two individual users respectively, and then sum them to obtain the inner product value of the feature vectors of the two individual users. Sum the squares of the feature values ​​of each dimension in the feature vectors of the two individual users and take the square root to obtain the length of each vector. Divide the inner product value by the product of the two vector lengths to obtain the feature similarity between the two individual users. The value of the feature similarity represents the degree of closeness between the two individual users in terms of feature distribution. A high feature similarity indicates a strong feature correlation between the two individual users. Each user is mapped to a node in the user association graph. The connection relationship is determined based on the feature similarity between two users. When the feature similarity between two users is greater than or equal to the preset similarity threshold, an edge is established between the two corresponding nodes. When the feature similarity between two users is less than the preset similarity threshold, no edge is established between the two corresponding nodes. Assign edge weights to the established edges, and generate a user association graph based on all nodes, the established edges, and the edge weights of each edge. The process of generating a user association graph is as follows: After establishing the connection relationship between individual users, each individual user is identified as a node in the user association graph, and the connection between two individual users with established connection relationship is identified as an edge; for each edge, the feature similarity between the corresponding two individual users is read, and the feature similarity is used as the edge weight. All nodes corresponding to all individual users, all established edges, and edge weights corresponding to each edge are uniformly organized to form a user association graph that represents the strength of the association relationship between individual users.

[0024] In this embodiment, the construction of the collective semantic resonance field specifically includes: Community detection processing is performed on the user association graph, and all individual users are divided into user groups based on the edges and edge weights in the user association graph. The partitioning process is as follows: Using individual users in the user association graph as nodes and edges and their weights as connections, traverse all nodes in the graph and initialize each node as an initial community. Calculate the connection strength between nodes based on edge weights, merge pairs of nodes with high connection strength, and gradually aggregate nodes with large and interconnected edge weights into the same candidate community. During node aggregation, compare the changes in the total edge weights within a community and the cross-community edge weights before and after a node is merged, and use a merging method that increases the connection strength within the community to iteratively partition the nodes. Repeat the node merging and community adjustment operations until the community affiliation of each node in the user association graph no longer changes, thus obtaining the user group. The group distribution parameters are calculated based on the individual feature vectors of all users within the user group. The group distribution parameters include the group center vector and the group discrete matrix. The calculation process for the group center vector is as follows: the feature values ​​of all individual user feature vectors in the same user group are accumulated in each corresponding dimension, and the accumulated result of each dimension is divided by the number of individual users in the user group to obtain the average value of the user group in each feature dimension, which is used as the group center vector. The process of calculating the discrete matrix of the group is as follows: For the feature vector of each user in the same user group, calculate the difference vector between the feature vector of the user and the central vector of the group. Perform matrix multiplication on the difference vector and the transpose of the difference vector to obtain the discrete matrix components of the corresponding user. Then, sum up the discrete matrix components corresponding to all users in the user group and divide by the number of users in the user group to obtain the discrete matrix of the group. The corresponding group semantic resonance field is constructed based on the group center vector and the group discrete matrix. The construction process is as follows: for any position vector in the semantic space, calculate the difference vector between the position vector and the group center vector, perform matrix multiplication on the transpose of the difference vector, the inverse matrix of the group discrete matrix and the difference vector in sequence, and multiply the result by negative half and then perform exponential operation to obtain the resonance intensity of the position vector in the corresponding group semantic resonance field, thus forming the group semantic resonance field.

[0025] In this embodiment, the construction of the conditional input sequence specifically includes: Calculate the group condition vector based on the group semantic resonance field of the corresponding user group; The specific process of calculating the group condition vector is as follows: Select semantic position vectors in the semantic space corresponding to the group semantic resonance field, multiply the resonance intensity of each semantic position vector in the group semantic resonance field with the corresponding semantic position vector, sum the results, and divide the sum by the sum of the resonance intensities corresponding to all semantic position vectors to obtain the group condition vector. Obtain the generated prompt information and convert it into an input vector sequence; The generated prompt information includes content theme information, content task information, content format information, and scenario constraint information. Content theme information represents the theme object of the content to be generated, content task information represents the generation task type of the content to be generated, content format information represents the output format of the content to be generated, and scenario constraint information represents the application scenario corresponding to the content to be generated. The process of converting the generated prompt information into an input vector sequence is as follows: the generated prompt information is segmented into text units arranged in order; the corresponding word representations are extracted sequentially according to the arrangement position of each text unit in the generated prompt information, and each text unit is mapped to a corresponding initial vector; position vectors are superimposed on the initial vectors to obtain input vectors representing the text content and positional relationships; all input vectors are arranged according to the original order of each text unit in the generated prompt information to form an input vector sequence. Calculate the semantic coupling coefficient based on each input vector in the input vector sequence and the group conditional vector; The semantic coupling coefficient is calculated as follows: each input vector in the input vector sequence is linearly transformed with the group condition vector to obtain the corresponding transformed vector; the transformed vector corresponding to each input vector is multiplied dimension by dimension with the transformed vector corresponding to the group condition vector and accumulated to obtain the association value between the input vector and the group condition vector; the association values ​​corresponding to all input vectors are exponentially operated on, and the exponential result corresponding to each association value is divided by the sum of the exponential results corresponding to all association values ​​to obtain the semantic coupling coefficient corresponding to each input vector. Based on the semantic coupling coefficient, the input vector sequence is semantically coupled with the group condition vector to construct the condition input sequence. The construction process is as follows: for each input position, the corresponding input vector and the vector obtained by mapping the group condition vector are weighted and combined according to the corresponding semantic coupling coefficient to obtain the condition input vector corresponding to the input position, and the condition input sequence is composed of all the condition input vectors in sequence.

[0026] In this embodiment, the generation of target content specifically includes: The conditional input sequence is input into the AIGC model and represented as a vector sequence formed by arranging the conditional input vectors in the input order, with each conditional input vector corresponding to an input position; Based on the order of the conditional input vectors in the conditional input sequence, the corresponding generation state of the AIGC model is calculated sequentially. The AIGC model consists of an input processing layer, a semantic coupling processing layer, a generation state update layer, and a content output layer. The input processing layer is used to receive the conditional input sequence, the semantic coupling processing layer is used to maintain the fusion relationship between the generation prompt information and the group conditional vector, the generation state update layer is used to calculate the current generation state based on the previous generation state and the current conditional input vector, and the content output layer is used to output content units based on the generation state and form the target content. The specific process of the state update layer of the AIGC model is as follows: When processing the current conditional input vector, the generated state corresponding to the previous conditional input vector is read, and the generated state corresponding to the previous conditional input vector is concatenated with the current conditional input vector to obtain the joint input at the current time step; based on the joint input, the gate values ​​of the input gate, forget gate, and output gate are calculated respectively, and the candidate state at the current time step is calculated based on the product accumulation result of each component in the joint input and the corresponding connection parameter. The product accumulation result is processed with hyperbolic tangent to obtain the value of each component in the candidate state; the input gate is used to weight each component in the candidate state, and the forget gate is used to weight each component in the generated state corresponding to the previous conditional input vector. The weighted candidate state and the weighted generated state corresponding to the previous conditional input vector are added component by component to obtain the intermediate state at the current time step; the output gate is used to weight each component in the intermediate state, and the intermediate state is processed with hyperbolic tangent before the component weighting to obtain the generated state corresponding to the current conditional input vector. Calculate the probability distribution of the content unit corresponding to each input position based on the generation state corresponding to each input position; The calculation process of the content unit probability distribution is as follows: For the generation state corresponding to each input position, read the values ​​of each component in the corresponding generation state, and multiply and accumulate each component value with the connection parameter corresponding to each candidate content unit to obtain the score value corresponding to each candidate content unit; perform exponential operation on the score value corresponding to each candidate content unit, and divide the exponential result corresponding to each candidate content unit by the sum of the exponential results corresponding to all candidate content units to obtain the probability value corresponding to each candidate content unit; the probability values ​​corresponding to all candidate content units form the content unit probability distribution corresponding to the input position. The content unit corresponding to each input position is determined based on the probability distribution of the content unit corresponding to each input position. The specific process is as follows: select the candidate content unit with the highest probability value from the probability distribution of the content unit corresponding to the current input position, and use it as the content unit corresponding to the current input position. Output the content unit sequence according to the arrangement order of the content units corresponding to each input position, and concatenate the content unit sequence in order to obtain the target content that is consistent with the semantic features of the corresponding user group.

[0027] In this embodiment, the calculation of group fitness and individual fitness specifically includes: The semantic projection of the target content is calculated to obtain the content semantic vector. Specifically, the target content is represented as a sequence of content units arranged in the content order, and each content unit in the content unit sequence is represented by a vector to obtain the content unit vector corresponding to each content unit. All content unit vectors are summed and averaged according to the number of content units to obtain the content semantic vector corresponding to the target content. The group fit degree is calculated based on the content semantic vector and the group semantic resonance field. Specifically, the content semantic vector is used as the position vector in the semantic space and input into the group semantic resonance field of the corresponding user group. The resonance intensity of the content semantic vector in the group semantic resonance field is calculated, and the resonance intensity is determined as the group fit degree of the target content for the user group. The individual matching degree is calculated based on the content semantic vector and the user's individual feature vector. Specifically, for each user belonging to the corresponding user group, the user's individual feature vector is read. The components of each corresponding dimension in the content semantic vector and the user's individual feature vector are multiplied and summed to obtain the corresponding inner product value. The square roots of the squares of each dimension component in the content semantic vector and the user's individual feature vector are summed and taken to obtain the respective vector lengths. The inner product value is divided by the product of the vector length of the content semantic vector and the vector length of the user's individual feature vector to obtain the individual matching degree of the target content for that user.

[0028] In this embodiment, the generation of the distribution result specifically includes: Based on the distribution processing of target content within the corresponding user group, the group fit degree corresponding to the target content and the individual matching degree corresponding to each user within the corresponding user group are read. Then, a distribution score is calculated for each user within the corresponding user group. Specifically, the group fit degree is multiplied by the weight coefficient corresponding to the group fit degree to obtain the group score component; the individual matching degree is multiplied by the weight coefficient corresponding to the individual matching degree to obtain the individual score component; the group score component and the individual score component are summed to obtain the distribution score of the target content for the corresponding user. The weight coefficient corresponding to the individual matching degree is one minus the weight coefficient corresponding to the group fit degree. The target content is sorted within the corresponding user group based on the distribution score corresponding to each individual user. Specifically, the distribution scores corresponding to all individual users within the corresponding user group are sorted according to the numerical order to obtain the sorting result of the target content within the corresponding user group. The sorting result is formed by arranging the individual users within the corresponding user group according to their distribution scores. Based on the individual matching degree, the target content is adjusted at the individual level to obtain the individual adjustment vector corresponding to the user. Specifically, for each user in the corresponding user group, the user's individual feature vector and the content semantic vector corresponding to the target content are read, and the difference vector between the user's individual feature vector and the content semantic vector is calculated. The difference vector is multiplied by the individual matching degree of the corresponding user and then multiplied by the adjustment coefficient to obtain the individual adjustment vector corresponding to the user. The content semantic vector is generated based on the individual adjustment vector. Specifically, the content semantic vector corresponding to the target content is added to the individual adjustment vector of the corresponding user to obtain the individual-level adjusted content semantic vector for the user. Distribution results are generated based on the sorting results and the individual-level adjusted content semantic vectors corresponding to each user. Specifically, according to the order of each user in the sorting results, each user is associated with the corresponding individual-level adjusted content semantic vector to form distribution results for each user within the corresponding user group.

[0029] In this embodiment, the updating of individual user feature vectors, user association graphs, and group semantic resonance fields specifically includes: For each individual user within the corresponding user group, feedback data on the distribution results is collected. The feedback data includes click feedback data, dwell feedback data, and conversion feedback data. The click feedback data, dwell feedback data, and conversion feedback data for each individual user are then numerically processed to obtain the feedback value for each individual user. Specifically, the value corresponding to the click feedback data is multiplied by the feedback weight corresponding to the click feedback data, the value corresponding to the dwell feedback data is multiplied by the feedback weight corresponding to the dwell feedback data, and the value corresponding to the conversion feedback data is multiplied by the feedback weight corresponding to the conversion feedback data. The three are then added together to obtain the feedback value for the corresponding individual user. The user individual feature vector is updated based on the feedback value corresponding to each user individual. Specifically, for each user individual, the user individual feature vector before the update is combined with the feedback value corresponding to the user individual and the content semantic vector corresponding to the target content. The content semantic vector is multiplied by the feedback value and the update step size and then superimposed on the user individual feature vector before the update to obtain the user individual feature vector after the update. The user association graph is updated based on the updated user individual feature vectors. Specifically, for any two users, the corresponding updated user individual feature vectors are read, and the components of each corresponding dimension in the two updated user individual feature vectors are multiplied and summed to obtain the corresponding inner product value. The squares of each dimension component in the two updated user individual feature vectors are summed and squared to obtain the length of each vector. The inner product value is divided by the product of the lengths of the two updated user individual feature vectors to obtain the updated feature similarity between the two users. The edge weights of the corresponding edges in the user association graph are updated based on the updated feature similarity. The group distribution parameters for each user group are recalculated based on the updated individual user feature vectors, and the corresponding group semantic resonance field is updated based on the recalculated group distribution parameters. The updated group distribution parameters include the updated group center vector and the updated group discrete matrix. The calculation process of the updated group center vector is as follows: the components of all updated individual user feature vectors in the same user group are accumulated in each corresponding dimension, and the accumulated result in each dimension is divided by the number of individual users in the user group to obtain the updated group center vector. The calculation process of the updated group discrete matrix is ​​as follows: for each updated individual user feature vector in the same user group, the difference vector between the updated individual user feature vector and the updated group center vector is calculated, and the difference vector is multiplied by the transpose of the difference vector to obtain the corresponding updated discrete matrix components. The updated discrete matrix components in the same user group are accumulated and then divided by the number of individual users in the user group to obtain the updated group discrete matrix. An updated group semantic resonance field is constructed based on the updated group center vector and the updated group discrete matrix. Specifically, for any position vector in the semantic space, the difference vector between the position vector and the updated group center vector is calculated. Then, the transpose of the difference vector, the inverse of the updated group discrete matrix, and the difference vector are successively multiplied by matrix multiplication. The result is multiplied by negative half and then exponentially calculated to obtain the resonance intensity of the position vector in the updated group semantic resonance field.

[0030] Example 1: Taking the application of a large e-commerce content operation platform in a "Family Health Management" campaign as an example, this invention provides a detailed explanation of an AIGC-based intelligent content generation and precise distribution method. This platform sells various products including blood pressure monitors, body fat scales, fascia guns, lumbar support products, sleep aids, and nutrition courses. Daily, the platform needs to push content to users with different interests, spending power, access times, and device terminals. Current practices primarily involve operators providing standardized campaign text or using a general AIGC model to generate several sets of generic content, which are then distributed based on simple tags or historical click records. While this approach can generate a large amount of content in a short time, it often results in inconsistencies between the content's wording and the target user group's focus. For example, users interested in "exercise recovery" receive content geared towards "sleep improvement," while users interested in "elderly health monitoring" receive motivational exercise text with a more youthful tone, leading to a disconnect between generated content and distribution results. In actual operation, the platform also found that when relying solely on a single user profile for content matching, although individual users may have high click rates, the overall stability of the same batch of content within a user group is poor, the content style is prone to drift, the click-through rate and add-to-cart rate of subsequent distribution fluctuate significantly, and the cost of manual rewriting remains high.

[0031] In this scenario, the platform first collects user behavior data, user attribute data, and contextual environment data from logs over the past ninety days. User behavior data includes page browsing history, click history, favorites history, add-to-cart history, order history, dwell time, and search term sequences; user attribute data includes age range, consumption level, geographic level, device type, and historical purchase category preferences; contextual environment data includes access time period, activity cycle, content entry location, and holiday tags. After denoising, missing data completion, and normalization, the platform constructs individual user feature vectors. Subsequently, the platform constructs a user association graph based on these feature vectors, using individual users as nodes and similarity relationships between users as edges, with edge weights determined by similarity magnitude. Community detection processing is then applied to the user association graph to obtain several user groups. In this embodiment, the platform identifies twelve main user groups from 186,000 active users during the activity period, including typical groups such as "Sedentary Office Workers' Recovery Group," "Nighttime Sleep Improvement Group," "Elderly Family Monitoring Group," "Exercise Recovery and Improvement Group," and "High-Single-Amount Health Equipment Decision-Making Group."

[0032] After identifying user groups, the platform calculates group distribution parameters for each group and constructs a corresponding group semantic resonance field. This group semantic resonance field is not simply a set of labels, but rather a distributed expression of the common focus directions of a user group in the semantic space. For example, the group semantic resonance field for the "family elderly monitoring group" shows a significantly higher concentration in semantic regions related to "stability, reliability, easy reading, and remote viewing" than in regions such as "trendy, shaping, and challenge." Conversely, the group semantic resonance field for the "exercise recovery and improvement group" is concentrated in semantic regions such as "muscle relaxation, post-training recovery, portability, battery life, and intensity adjustment." The platform then calculates a group condition vector based on the group semantic resonance field to obtain generated prompt information. This prompt information includes the theme, expression form, application scenario, and related text description of the content to be generated, such as "recommended text on the first screen of the family health management special page," "short content for mobile product cards," and "recommendation scenarios suitable for 8 PM to 10 PM." Then, the platform fuses the generated prompt information with the group condition vector through semantic coupling to construct a condition input sequence, and inputs the condition input sequence into the AIGC model to generate target content that is consistent with the semantic features of the corresponding user group.

[0033] In this embodiment, the platform does not generate a single set of uniform content, but rather generates different content versions for different user groups. For example, for the "family elderly monitoring group," the target content emphasizes "clear visibility, stable measurement, viewability by children, and minimal operation steps"; for the "nighttime sleep improvement group," the target content emphasizes "improved sleep environment, low nighttime noise, relaxing companionship, and long-term habit establishment"; and for the "exercise recovery and improvement group," it emphasizes "post-training relaxation, zoned relaxation, deep massage, and portability with long battery life." After generation, the platform performs semantic projection calculation on the target content to obtain content semantic vectors. Then, it calculates the group fit based on the content semantic vectors and the group semantic resonance field, and calculates the individual matching degree based on the content semantic vectors and the user's individual feature vectors. The result obtained in this way is not a simple "good or bad," but rather a two-tiered evaluation at both the group and individual levels. For content with high group fit but average individual fit, the system will retain its exposure qualification at the group level; for content with both high group fit and high individual fit, it will be given higher distribution priority during sorting; for content with average group fit but high individual fit, it will be adjusted at the individual level before being released, thereby reducing the disturbance to the entire user group.

[0034] During the official launch phase of the campaign, the platform ran continuously for fourteen days, generating and distributing approximately 2,400 pieces of target content daily, covering homepage recommendation slots, product cards on search results pages, special event venues, and message delivery slots. To verify the effectiveness of this invention, the platform simultaneously set up three control schemes. The first scheme used a traditional rule template plus tag distribution method, where content was generated by a manual template library and then distributed based on category tags and coarse-grained audience tags. The second scheme used a general AIGC plus collaborative filtering method, where content was generated by a general model and then distributed using a conventional recommendation model. The third scheme used a single user profile plus AIGC targeted distribution method, which did not construct user association graphs and group semantic resonance fields, but directly drove content generation and distribution based on individual user feature vectors. The fourth scheme is the scheme of this invention, which adopts a complete process of user association graphs, group semantic resonance fields, semantic coupling, conditional input sequences, semantic projection calculation, and collaborative ranking of group fit and individual matching degree. The four schemes were compared under the same time period, the same product pool, and the same exposure scale, with the total exposure controlled at around 4.8 million times to ensure the comparability of experimental data. The experimental results are shown in Table 1 below.

[0035] Table 1. Comparison of the effects of different content generation and distribution schemes in the family health management topic.

[0036] As shown in Table 1, the method of this invention achieves a click-through rate of 6.34% while maintaining a relatively consistent exposure scale. This is significantly higher than the 3.21% of rule-based template plus tag distribution, the 4.07% of general AIGC plus collaborative filtering, and the 4.86% of single-user profile plus AIGC targeted distribution. This indicates that this invention does not simply rely on AIGC to increase content attractiveness, but rather, through the combined effects of user association graphs, user group segmentation, and group semantic resonance fields, makes the target content more consistent with the semantic characteristics of the target user group during the generation stage, thus achieving a higher click-through rate after actual exposure.

[0037] In terms of add-to-cart conversion rate and average dwell time, the method of this invention also has significant advantages. The add-to-cart conversion rate reached 2.21%, which is 1.25 percentage points higher than the traditional rule template solution and 0.58 percentage points higher than the single user profile plus AIGC targeted distribution solution. The average dwell time reached 37.6 seconds, indicating that users are not only willing to click, but also willing to further read and browse related content. This result shows that after the content is generated, the present invention obtains the content semantic vector through semantic projection calculation, and then combines group adaptability and individual matching degree for sorting and individual-level adjustment, which effectively enhances the fit between content and user needs, so that content distribution no longer stops at coarse-grained recommendation, but forms a two-layer precise matching for user groups and individual users.

[0038] The technical value of this invention can be further demonstrated by two indicators: the manual rewriting rate and the standard deviation of the distribution score. The manual rewriting rate decreased from 31.2% in the rule-based template scheme to 9.1%, indicating that the target content generated by this invention already has high usability upon initial output, significantly reducing the content revision burden on the operations team. The standard deviation of the distribution score decreased from 0.184 to 0.103, indicating that the distribution results of this invention are more stable within the same activity cycle, and it is less likely to result in one group performing well while another group is significantly unbalanced. This is closely related to the invention's modeling of the common semantics of the user group through a group semantic resonance field, and its continuous updating of individual user feature vectors, user association graphs, and the group semantic resonance field through feedback data.

[0039] As can be seen from this embodiment, the present invention effectively solves the problems in the prior art, such as the disconnect between content generation and content distribution, insufficient utilization of common features of user groups, lack of stable coupling between generated input and group semantics, and lack of group-level closed-loop updates after distribution. In real business scenarios, the present invention not only improves the integration of content generation and accurate distribution, but also achieves quantifiable improvements in click-through rate, conversion rate, dwell time, manual rewriting cost, and distribution stability, proving that the present invention has good engineering application value and practical promotion significance.

[0040] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AIGC-based intelligent content generation and accurate distribution method, characterized in that, Includes the following steps: Collect and preprocess user behavior data, user attribute data, and contextual environment data to construct individual user feature vectors; Construct a user association graph based on individual user feature vectors; Community detection processing is performed on the user association graph to obtain user groups and calculate group distribution parameters, and construct the corresponding group semantic resonance field; The group condition vector is calculated based on the group semantic resonance field to obtain the generated prompt information, and the generated prompt information is fused with the group condition vector through semantic coupling to construct the condition input sequence; The conditional input sequence is input into the AIGC model, and the generation process is constrained based on the conditional input sequence to generate target content that is consistent with the semantic features of the corresponding user group. Semantic projection is performed on the target content to obtain the content semantic vector. The group fit is calculated based on the content semantic vector and the group semantic resonance field. The individual matching degree is calculated based on the content semantic vector and the user's individual feature vector. Based on group adaptability and individual matching degree, the target content is sorted within the corresponding user group, and the target content is adjusted at the individual level according to the individual matching degree to generate distribution results for different individual users; Collect feedback data from individual users on the distribution results, and update the individual user feature vector, user association graph, and group semantic resonance field based on the feedback data. 2.The AIGC-based intelligent content generation and accurate distribution method of claim 1, wherein, The collected user behavior data, user attribute data, and contextual environment data are uniformly aligned. Data from different sources are associated and integrated according to user identifiers, missing values ​​are filled, outliers are removed, and numerical normalization is performed. User behavior data is serialized and encoded to extract behavioral pattern features. User attribute data is categorically encoded to form attribute features. Contextual environment data is temporally extracted to form scene features. The behavioral features, attribute features, and scene features are concatenated and fused, and then dimensionality reduction and nonlinear mapping are performed to obtain a unified-dimensional user individual feature vector. 3.The AIGC-based intelligent content generation and accurate distribution method of claim 1, wherein, The construction of the user association graph specifically includes: Calculate the feature similarity between two users based on their individual feature vectors. Each user is mapped to a node in the user association graph. The connection relationship is determined based on the feature similarity between two users. When the feature similarity between two users is greater than or equal to the preset similarity threshold, an edge is established between the two corresponding nodes. When the feature similarity between two users is less than the preset similarity threshold, no edge is established between the two corresponding nodes. Assign edge weights to the established edges, and generate a user association graph based on all nodes, the established edges, and the edge weights of each edge.

4. The intelligent content generation and accurate distribution method based on AIGC of claim 1, wherein, The construction of the collective semantic resonance field specifically includes: Community detection processing is performed on the user association graph, and all individual users are divided according to the edges and edge weights in the user association graph to obtain user groups; The group distribution parameters are calculated based on the individual feature vectors of all users within the user group. The group distribution parameters include the group center vector and the group discrete matrix. Calculate the group center vector; Calculate the discrete matrix of the population; Construct the corresponding group semantic resonance field based on the group center vector and the group discrete matrix.

5. The method for intelligent content generation and precise distribution based on AIGC according to claim 1, characterized in that, The construction of the conditional input sequence specifically includes: Calculate the group condition vector based on the group semantic resonance field of the corresponding user group; Obtain the generated prompt information and convert the generated prompt information into an input vector sequence; Calculate the semantic coupling coefficient based on each input vector in the input vector sequence and the group conditional vector; Based on the semantic coupling coefficient, the input vector sequence is semantically coupled with the group condition vector to construct a conditional input sequence.

6. The method for intelligent content generation and precise distribution based on AIGC according to claim 1, characterized in that, The generation of the target content specifically includes: The conditional input sequence is input into the AIGC model, and the conditional input sequence is represented as a vector sequence formed by arranging the conditional input vectors in the input order, with each conditional input vector corresponding to an input position; Based on the order of the conditional input vectors in the conditional input sequence, the generation state corresponding to the AIGC model is calculated sequentially. Calculate the probability distribution of the content unit corresponding to each input position based on the generation state corresponding to each input position; The content unit corresponding to each input position is determined based on the probability distribution of the content unit corresponding to each input position. Output a sequence of content units according to the order of the content units corresponding to each input position, and then concatenate the content unit sequence in order to obtain target content that is consistent with the semantic features of the corresponding user group.

7. The method for intelligent content generation and precise distribution based on AIGC according to claim 1, characterized in that, The calculation of the group fitness and individual fitness specifically includes: Semantic projection is performed on the target content to obtain the content semantic vector; Calculate group fit based on content semantic vectors and group semantic resonance fields; The individual matching degree is calculated based on the content semantic vector and the user's individual feature vector.

8. The method for intelligent content generation and precise distribution based on AIGC according to claim 1, characterized in that, The generation of the distribution results specifically includes: Based on the distribution of target content within the corresponding user group, read the group fit degree corresponding to the target content and the individual matching degree corresponding to each user within the corresponding user group, and calculate the distribution score for each user within the corresponding user group. Based on the distribution score corresponding to each individual user, the target content is sorted within the corresponding user group to obtain the sorting results; Based on the individual matching degree, the target content is adjusted at the individual level to obtain the individual adjustment vector corresponding to each user. Generate individual-level adjusted content semantic vectors based on individual adjustment vectors; Distribution results are generated based on the sorting results and the individual-level adjusted semantic vectors of each user.

9. The method for intelligent content generation and precise distribution based on AIGC according to claim 1, characterized in that, The updates to the individual user feature vectors, user association graphs, and group semantic resonance fields specifically include: For each individual user within the corresponding user group, feedback data on the distribution results is collected. The feedback data includes click feedback data, dwell feedback data, and conversion feedback data. The click feedback data, dwell feedback data, and conversion feedback data corresponding to each individual user are numerically processed to obtain the feedback value corresponding to each individual user. Update the individual user feature vector based on the feedback value corresponding to each individual user; Update the user association graph based on the updated individual user feature vectors; Based on the updated individual user feature vectors, the group distribution parameters corresponding to each user group are recalculated, and the corresponding group semantic resonance field is updated based on the recalculated group distribution parameters. An updated group semantic resonance field is constructed based on the updated group center vector and the updated group discrete matrix.

Images