A marketing content intelligent distribution method and system based on user behavior data

Abstract

The present application relates to the technical field of data analysis, and specifically discloses a marketing content intelligent distribution method and system based on user behavior data. The present application obtains user content interaction sequence and feedback data, constructs a content intention graph in combination with historical behavior data, extracts a recommendation fatigue characteristic coefficient, further divides explicit and implicit feedback data, accurately extracts content labels and monitors distribution density, quantizes content concentration characteristic values of excessive concentration, simultaneously mines implicit feedback data to obtain cross-circle layer flow tendency characteristic values, and finally combines candidate content freshness and interest matching degree, performs weighted fusion sorting through multi-target adjustment of a weight vector, and screens high-quality content for intelligent distribution, which can relieve user fatigue caused by excessive push of similar content, and improve marketing content matching precision and conversion efficiency.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a method and system for intelligent distribution of marketing content based on user behavior data. Background Technology

[0002] With the deepening development of digital marketing, marketing platforms generally adopt personalized content distribution strategies to improve user activity and conversion efficiency. Current mainstream technical solutions mostly build interest profiles based on explicit behavioral data such as user history clicks and browsing time, and then match and push content accordingly. However, existing systems rely too heavily on short-term, high-frequency user interactions and continuously push similar tagged content to users (such as pushing "sneaker" related content for several consecutive days), leading to user fatigue and resistance, and thus a decline in marketing conversion. Therefore, a smart marketing content distribution method based on user behavior data is needed to solve the above problems. Summary of the Invention

[0003] The purpose of this invention is to provide a method for intelligent distribution of marketing content based on user behavior data, comprising: Acquire user behavior data on the marketing platform, including content interaction sequences and content feedback data; Obtain historical behavior data of users on the marketing platform, and obtain a content intent graph based on the historical behavior data and the content interaction sequence, and obtain a recommendation fatigue feature coefficient based on the content intent graph; Explicit feedback data and implicit feedback data are obtained based on the content feedback data, and multiple content tags are obtained based on the explicit feedback data and the content intent graph. The first distribution density corresponding to each content tag is obtained, and it is determined in turn whether each first distribution density exceeds a preset threshold. If the first distribution density exceeds a preset threshold, it is determined that the content tag corresponding to the first distribution density has excessive concentration of marketing content distribution, and all second content tags exceeding the threshold and their second distribution densities are counted, and content concentration feature values ​​are obtained based on multiple second content tags and multiple second distribution densities. The implicit feedback data is used to obtain the duration of user ignoring distributed content and the behavior of quickly skipping it. Based on the ignoring duration, the behavior of quickly skipping it and the content interaction sequence, the cross-circle flow tendency feature value is obtained. A marketing content set is constructed based on the recommended fatigue characteristic coefficient, the content set characteristic value, and the cross-circle flow tendency characteristic value. The marketing content is then intelligently distributed to users based on the marketing content set to obtain the intelligent distribution result.

[0004] Preferably, the step of obtaining a content intent graph based on the historical behavior data and the content interaction sequence, and obtaining a recommendation fatigue feature coefficient based on the content intent graph, includes: Based on the historical behavior data, historical content interaction sequences and historical feedback tags are obtained, and the historical content interaction sequences and historical feedback tags are time-series aligned and spliced ​​to obtain user long and short-term behavior sequences. Obtain the user-content interaction relationship in the user's long and short-term behavior sequences, construct a user-content bipartite graph based on the user-content interaction relationship, obtain user node vectors and content node vectors based on the user-content bipartite graph, and obtain a content intent graph based on the user node vectors, the content node vectors and the user-content interaction relationship; The frequency and average weight of user connections to similar content are obtained based on the content intent graph, and the interaction intensity saturation value is obtained based on the connection frequency and the average weight. Obtain the user's most recent interaction timestamp for similar content, and calculate the time decay factor based on the most recent interaction timestamp; The recommended fatigue characteristic coefficients are obtained based on the interaction intensity saturation value and the time decay factor.

[0005] Preferably, the step of obtaining explicit feedback data and implicit feedback data based on the content feedback data, and obtaining multiple content tags based on the explicit feedback data and the content intent graph, includes: The content feedback data is categorized into explicit feedback data and implicit feedback data. The content identifier and behavior intensity value are obtained based on the explicit feedback data, and the content node embedding vector is obtained from the content intent graph based on the content identifier. Obtain the semantic vectors of tags from a preset content tag library, and calculate the cosine similarity based on the semantic vectors of the tags and the embedding vectors of the content nodes; Multiple initial content tags are obtained based on the cosine similarity and the behavior intensity value, and multiple initial weights are obtained corresponding to the multiple initial content tags. Multiple initial content tags with initial weights greater than a preset threshold are defined as multiple content tags.

[0006] Preferably, the step of obtaining feature values ​​from the content set based on multiple second content tags and multiple second distribution densities includes: Obtain the frequency of occurrence and the total number of distributed content for each second content tag within a preset statistical period, and obtain the original concentration based on the frequency of occurrence and the total number of distributed content; Obtain the historical average click-through rate of the content corresponding to each second content tag, map the historical average click-through rate to an adjustment factor based on the sigmoid function, and obtain the corrected concentration based on the adjustment factor and the original concentration. Obtain multiple second distribution densities and multiple distribution times for multiple second content tags, and perform time weighting on the multiple second distribution densities according to the multiple distribution times to obtain multiple time-weighted densities; Multiple concentration contribution values ​​are calculated based on the corrected concentration and multiple time-weighted densities, and content concentration feature values ​​are calculated based on the multiple concentration contribution values.

[0007] Preferably, the step of obtaining the user's ignore duration and quick skip behavior for distributed content based on the implicit feedback data, and obtaining cross-circle flow tendency feature values ​​based on the ignore duration, the quick skip behavior, and the content interaction sequence, includes: Extract the complete exposure duration of each distributed content from the implicit feedback data, and treat the exposure duration that is less than the first time threshold as the ignore duration. In particular, treat the complete exposure duration that is less than the second time threshold and is accompanied by rapid scrolling or page closing behavior as rapid skip behavior. Based on the fast skip behavior, obtain multiple content tags that are skipped within a preset continuous time window, and obtain the frequency of fast skip behavior based on the multiple content tags and the preset continuous time window; The first positive interaction content tag after the quick skip action is obtained according to the content interaction sequence, and the first positive interaction content tag is used as the new interaction tag; Each content tag and the new interaction tag are respectively input into the pre-trained tag semantic model to obtain multiple content tag semantic vectors and new interaction tag semantic vectors. Multiple vector distances between the multiple content tag semantic vectors and the new interaction tag semantic vectors are calculated, and the vector distances are used as semantic distances. Cross-sphere flow tendency feature values ​​are obtained based on multiple semantic distances, the ignoring duration, and the frequency of fast skipping behavior.

[0008] Preferably, the step of constructing a marketing content set based on the recommended fatigue characteristic coefficient, the content concentration characteristic value, and the cross-circle flow tendency characteristic value, and then intelligently distributing marketing content to users based on the marketing content set to obtain intelligent distribution results, includes: Get multiple candidate marketing content to be distributed, and get the release time and current time of each candidate marketing content, and get the real-time freshness score based on the release time and the current time; Based on the content intent graph, obtain the target user node vector and the candidate content node vector corresponding to each candidate marketing content, and calculate the inner product of each candidate content node vector and the target user node vector, and use the inner product as the graph relevance score; Based on the recommended fatigue feature coefficients, the feature values ​​of the content set, and the feature values ​​of the cross-sphere flow tendency, a multi-objective adjustment weight vector is constructed. The multiple graph relevance scores, multiple freshness scores, and the multi-objective adjustment weight vector are weighted and fused to obtain the comprehensive ranking score of the multiple candidate marketing content; Based on the comprehensive ranking score, the candidate marketing content is sorted in descending order, and the top N content is selected to form a marketing content set. The content in the marketing content set is then packaged into push tasks in sequence, the distribution operation is executed, and the distribution log is recorded. The distribution log is used as the intelligent distribution result.

[0009] This application also provides an intelligent marketing content distribution system based on user behavior data, including: The behavior data collection module is used to acquire user behavior data on the marketing platform, including content interaction sequences and content feedback data. The intent graph and fatigue analysis module is used to acquire users' historical behavior data on the marketing platform, and to acquire a content intent graph based on the historical behavior data and the content interaction sequence, and to acquire recommendation fatigue feature coefficients based on the content intent graph. The content tag and density monitoring module is used to obtain explicit feedback data and implicit feedback data based on the content feedback data, obtain multiple content tags based on the explicit feedback data and the content intent graph, obtain the first distribution density corresponding to each content tag, and sequentially determine whether each first distribution density exceeds a preset threshold. The content concentration measurement module is used to determine that the content tag corresponding to the first distribution density has excessive concentration of marketing content distribution if the first distribution density exceeds a preset threshold, and to count all second content tags that exceed the threshold and their second distribution densities, and to obtain content concentration feature values ​​based on multiple second content tags and multiple second distribution densities. The cross-circle flow analysis module is used to obtain the duration of user ignoring distributed content and the behavior of quickly skipping content based on the implicit feedback data, and to obtain cross-circle flow tendency characteristic values ​​based on the ignoring duration, the behavior of quickly skipping content and the content interaction sequence. The intelligent distribution decision module is used to construct a marketing content set based on the recommendation fatigue feature coefficient, the content set feature value, and the cross-circle flow tendency feature value, and to intelligently distribute marketing content to users based on the marketing content set to obtain intelligent distribution results.

[0010] Preferably, the intent mapping and fatigue analysis module includes: The behavior sequence splicing unit is used to obtain historical content interaction sequences and historical feedback tags based on the historical behavior data, and to perform time-series alignment and splicing of the historical content interaction sequences and historical feedback tags to obtain user long and short-term behavior sequences. The intent graph construction unit is used to obtain the user-content interaction relationship in the user's long and short-term behavior sequences, construct a user-content bipartite graph based on the user-content interaction relationship, obtain user node vectors and content node vectors based on the user-content bipartite graph, and obtain a content intent graph based on the user node vectors, the content node vectors and the user-content interaction relationship. An interaction intensity calculation unit is used to obtain the frequency and average weight of user connections to similar content based on the content intent graph, and to obtain the interaction intensity saturation value based on the connection frequency and the average weight. The time decay calculation unit is used to obtain the user's most recent interaction timestamp with similar content, and calculate the time decay factor based on the most recent interaction timestamp. The fatigue feature calculation unit is used to obtain recommended fatigue feature coefficients based on the interaction intensity saturation value and the time decay factor.

[0011] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0012] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0013] The beneficial effects of this application are as follows: This invention obtains user content interaction sequences and feedback data, constructs a content intent graph by combining historical behavior data, extracts recommendation fatigue feature coefficients, further divides explicit and implicit feedback data, accurately extracts content tags and monitors distribution density, quantifies content concentration feature values ​​for overly concentrated tags, and simultaneously mines implicit feedback data to obtain cross-circle flow tendency feature values. Finally, it combines the freshness and interest matching degree of candidate content, and uses multi-objective adjustment weight vector weighted fusion sorting to select high-quality content for intelligent distribution. This can alleviate the user fatigue problem caused by excessive push of similar content and improve the matching accuracy and conversion efficiency of marketing content. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of a method flow according to an embodiment of this application.

[0015] Figure 2 This is a schematic diagram of the system structure according to an embodiment of this application.

[0016] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0018] like Figures 1-2 As shown, this application provides a method for intelligent distribution of marketing content based on user behavior data, including: S1. Obtain user behavior data on the marketing platform, including content interaction sequences and content feedback data; S2. Obtain historical behavior data of users on the marketing platform, and obtain a content intent graph based on the historical behavior data and the content interaction sequence, and obtain a recommendation fatigue feature coefficient based on the content intent graph; S3. Obtain explicit feedback data and implicit feedback data based on the content feedback data, obtain multiple content tags based on the explicit feedback data and the content intent graph, obtain the first distribution density corresponding to each content tag, and sequentially determine whether each first distribution density exceeds a preset threshold. S4. If the first distribution density exceeds the preset threshold, it is determined that the content tag corresponding to the first distribution density has excessive concentration of marketing content distribution, and all second content tags exceeding the threshold and their second distribution densities are counted, and content concentration feature values ​​are obtained based on multiple second content tags and multiple second distribution densities. S5. Obtain the user's ignore duration and quick skip behavior for the distributed content based on the implicit feedback data, and obtain the cross-circle flow tendency feature value based on the ignore duration, the quick skip behavior and the content interaction sequence; S6. Based on the recommended fatigue feature coefficient, the content set feature value, and the cross-circle flow tendency feature value, a marketing content set is constructed, and the marketing content is intelligently distributed to users based on the marketing content set to obtain the intelligent distribution result.

[0019] As described in steps S1-S6 above, this invention first acquires user behavior data on the marketing platform, including content interaction sequences and content feedback data. The acquisition of this data relies on the marketing platform's behavior data collection module, which records user actions within the platform in real time. The content interaction sequence specifically covers the chronological record of user interactions with marketing content, such as browsing, clicking, saving, and sharing. For example, it could be a sequence of a user's actions within a week: browsing sports shoe reviews, clicking running gear promotional links, and saving sportswear recommendation pages. Content feedback data represents the user's response to the marketing content they encounter. This data also originates from the platform's real-time monitoring and recording of user behavior, providing fundamental data support for subsequent feature extraction and distribution decisions. Secondly, historical user behavior data on the marketing platform is acquired, and a content intent graph is obtained based on this historical behavior data and content interaction sequences, thereby obtaining a recommendation fatigue characteristic coefficient. The historical behavior data is stored long-term in the marketing platform's database, containing all past user interaction records within the platform's operational cycle, covering earlier content browsing, feedback, and other behaviors. Specifically, historical content interaction sequences and historical feedback tags are extracted from the historical behavior data, and these are aligned and concatenated chronologically to form long- and short-term user behavior sequences. For example, the short-term interaction sequence of the past month and the long-term interaction sequence of the past six months are integrated chronologically to fully present the user's behavior trajectory. Then, a user-content bipartite graph is constructed based on the user-content interaction relationships in this sequence. User node vectors and content node vectors are obtained through graph embedding technology. Combined with the weight distribution of user-content interaction relationships, a content intent graph reflecting user interest associations and content attribute associations is constructed. This graph clearly presents the strength of association between different users and various types of content, as well as the aggregation relationships of similar content. Subsequently, the frequency and average weight of user connections to similar content are extracted from the content intent graph. A weighted calculation yields the interaction intensity saturation value, which quantifies the depth and saturation of user interaction with a particular type of content. Simultaneously, the timestamps of the user's most recent interactions with similar content are obtained, and a time decay factor is calculated based on a time decay function. This factor reflects the degree to which user interest decays over time; for example, content frequently interacted with three months ago has a significantly higher time decay factor than content interacted with a week ago. Finally, the recommendation fatigue feature coefficient is obtained by multiplying the interaction intensity saturation value and the time decay factor. This coefficient directly reflects the user's fatigue level with current similar content pushes. A higher interaction intensity saturation value and a smaller time decay factor result in a larger recommendation fatigue feature coefficient, indicating that users are more prone to fatigue with similar content. The core value of this step lies in constructing a content intent graph by fusing long-term and short-term behavioral data. This breaks the limitations of relying solely on short-term data and allows for a more comprehensive understanding of users' long-term interest patterns. The extraction of recommendation fatigue feature coefficients enables a quantitative assessment of user fatigue levels, providing a crucial basis for avoiding excessive push of similar content. For example, when a user's recommendation fatigue feature coefficient for "outdoor camping equipment" reaches a high value, the system can detect that the user may have become fatigued with this type of content and adjust the distribution strategy accordingly.

[0020] Next, explicit and implicit feedback data are obtained based on the content feedback data. Multiple content tags are then acquired based on the explicit feedback data and the content intent graph, along with the first distribution density corresponding to each content tag. It is then determined whether each first distribution density exceeds a preset threshold. The classification of content feedback data is based on the directness of the feedback. Explicit feedback data is explicit feedback initiated by users, such as clicks, favorites, comments, and purchases; this data is directly recorded by the platform as the result of users' active actions. Implicit feedback data is indirect feedback generated by users during use, such as content exposure duration, scrolling speed, and page closing timing; this is collected through the platform's behavior monitoring tools. When acquiring content tags, the process first extracts content identifiers and behavior intensity values ​​from explicit feedback data. The behavior intensity value is set according to the weight of the feedback type; for example, the intensity value of purchasing behavior is higher than that of browsing behavior. Then, based on the content identifier, the corresponding content node embedding vector is retrieved from the content intent graph. This vector is a mathematical expression of the content attributes. Subsequently, the tag semantic vectors in the preset content tag library are retrieved. The preset content tag library is a standardized tag set pre-built by the platform based on industry attributes and content classification, containing core attribute tags of various marketing content. By calculating the cosine similarity between the tag semantic vector and the content node embedding vector, and combining it with the behavior intensity value for weighting, multiple initial content tags and their corresponding initial weights are obtained. The initial content tags with initial weights greater than a preset threshold are determined as the final multiple content tags. This process ensures a strong correlation between content tags and user explicit feedback. For example, a user's click behavior on a certain sports watch (explicit feedback) is combined with the high similarity between the node vector of the watch in the content intent graph and the semantic vectors of tags such as "smart wearables" and "sports monitoring," ultimately determining these two tags as core content tags. The first distribution density calculation involves the ratio of the number of times marketing content corresponding to each content tag is pushed to that user within a preset period to the total number of pushes. The preset threshold is determined based on the platform's historical marketing data and user acceptance tests, and is used to determine whether there is an excessive concentration trend in the distribution of content corresponding to that tag. This step accurately extracts content tags corresponding to user interests by distinguishing between explicit and implicit feedback data. At the same time, by monitoring the first distribution density, a quantitative standard is provided for subsequent judgments on the concentration of content distribution, avoiding the blindness of tag extraction and the disorder of distribution density, ensuring that content tags can truly reflect user interests and that the distribution density is within a reasonable range.

[0021] Subsequently, if the first distribution density exceeds a preset threshold, it is determined that the corresponding content tag has excessive concentration of marketing content distribution. All second content tags exceeding the threshold and their second distribution densities are statistically analyzed to obtain the content concentration feature value. When the first distribution density of a content tag exceeds the preset threshold, it indicates that the content corresponding to that tag has an excessively high proportion in recent pushes, which may cause user fatigue. This type of tag is the second content tag, and its second distribution density is the actual distribution density value exceeding the threshold for that tag. When calculating the content concentration feature value, firstly, the frequency of occurrence and the total number of distributed content for each second content tag within a preset statistical period are obtained. The ratio of these two is the original concentration degree, which directly reflects the proportion of that tag in distribution. Next, the historical average click-through rate (CTR) of the content corresponding to that tag is obtained, and it is mapped to an adjustment factor using the sigmoid function. The historical average CTR is calculated from the platform database based on the click data of the tag's past pushes. The application of the sigmoid function keeps the adjustment factor between 0 and 1. For tags with high CTRs, the adjustment factor is closer to 1, and its calibration set is appropriately increased. For tags with low click-through rates, the adjustment factor is closer to 0, reducing their corrected concentration and thus reasonably correcting the original concentration to obtain the corrected concentration. Then, combining the second distribution density and distribution time of multiple second content tags, a time-weighted density is obtained, reflecting the impact of recent distribution density on user perception, based on the principle that the more recent the distribution, the higher the weight. Finally, the concentration contribution value of each second content tag is obtained by multiplying the corrected concentration and the time-weighted density. The average of all concentration contribution values ​​is the content concentration feature value, which quantifies the degree of over-concentration in overall marketing content distribution. This step, through precise identification and quantitative evaluation of over-concentrated content tags, breaks the traditional vague judgment of content distribution concentration, objectively reflecting the distribution saturation state of various tag contents, and providing a clear quantitative basis for subsequent adjustments to distribution strategies. For example, when the content concentration feature value of the "leisure snacks" tag is high, the system can clearly know that this type of content is over-concentrated and its proportion needs to be reduced.

[0022] Furthermore, based on implicit feedback data, the platform obtains the duration of user ignoring distributed content and the behavior of quickly skipping content. Combined with content interaction sequences, it acquires cross-platform flow tendency characteristics. The platform extracts the complete exposure duration of each distributed piece of content from the implicit feedback data. This data is obtained by monitoring the content display time of users browsing the page in real time. Exposure durations shorter than a first time threshold (set based on the platform's average effective content browsing time, such as 5 seconds) are defined as ignoring durations, indicating that users have not shown effective interest in the content. Complete exposure durations shorter than a second time threshold (such as 2 seconds) accompanied by rapid scrolling or page closing are defined as quickly skipping behavior. This type of behavior directly reflects user resistance or disinterest in the content. Subsequently, multiple content tags corresponding to quickly skipping behavior are statistically analyzed within a preset continuous time window (such as 1 hour), calculating the frequency of quickly skipping behavior. Higher frequencies indicate stronger user resistance to the currently pushed content. From the content interaction sequence, the first positive interaction content tag after quickly skipping behavior (such as tags corresponding to clicks, favorites, etc.) is selected as a new interaction tag. This tag reflects the new interest direction that users truly focus on after skipping uninteresting content. Each content tag and a new interaction tag are input into a pre-trained tag semantic model (trained on a large amount of tag semantic data, capable of converting tags into semantically related vectors) to obtain corresponding semantic vectors. The vector distance between the two is calculated as the semantic distance; a larger semantic distance indicates a greater difference in the interest circles to which the two tags belong. Finally, a cross-circle flow tendency feature value is obtained by weighted summing of semantic distance, ignore duration, and frequency of rapid skipping behavior. A higher value indicates that the user is more likely to leave the current interest circle and seek new content of interest. This step delves into the latent user demand signals in implicit feedback data, breaking the limitations of traditional reliance solely on explicit feedback. It can accurately capture the dynamic changing trends of user interests. For example, when a user frequently skips content in the "home furnishings" category, and the first positive interaction tag is "outdoor backpack," the semantic distance between the two is large, and the cross-circle flow tendency feature value is high. The system can perceive that the user's interest has shifted from the home furnishing circle to the outdoor circle, providing a basis for subsequently pushing new circle content.

[0023] Next, a marketing content set is constructed based on the recommendation fatigue feature coefficient, content concentration feature value, and cross-circle flow tendency feature value. This set is then used for intelligent distribution of marketing content to users, yielding intelligent distribution results. First, multiple candidate marketing content sets to be distributed are acquired. These sets originate from the marketing platform's content library, covering marketing materials across various themes and tags. The publication time of each candidate marketing content is recorded. By calculating the time difference between the publication time and the current time, and combining this with a freshness decay function, a real-time freshness score is obtained. The more recent the publication time, the higher the freshness score, ensuring the timeliness of the pushed content. The target user node vector and the candidate content node vector of each candidate marketing content are extracted from the content intent graph. The inner product of these two vectors is calculated as the graph relevance score; a larger inner product indicates a higher degree of interest matching between the user and the content. A multi-objective adjustment weight vector is constructed based on recommendation fatigue feature coefficients, content concentration feature values, and cross-circle flow tendency feature values. The weight corresponding to the recommendation fatigue feature coefficient is used to reduce the push priority of content with high fatigue levels; the weight corresponding to the content concentration feature value is used to reduce the push proportion of content with excessively concentrated tags; and the weight corresponding to the cross-circle flow tendency feature value is used to increase the push weight of content from new interest circles. The weight allocation is dynamically adjusted based on platform marketing goals and user feedback data. The graph relevance score, freshness score, and multi-objective adjustment weight vector are weighted and fused to obtain a comprehensive ranking score for each candidate marketing content. After sorting by score in descending order, the top N (N is set according to the platform's push capacity and user acceptance, such as 10 items) of content are selected to form a marketing content set. These contents are then packaged into push tasks, and distribution operations are performed, recording distribution logs. These logs include push content, push time, and subsequent user feedback, serving as the intelligent distribution result for subsequent data iteration. This step, through comprehensive weighting of multi-dimensional features, enables precise and dynamic adjustment of marketing content delivery. It ensures the matching degree between content and user interests, effectively avoids the over-pushing of similar content, and adapts to the cross-circle flow of user interests. For example, when the user's recommendation fatigue feature coefficient is high and the cross-circle flow tendency feature value is high, the system will reduce the weight of high fatigue tag content and increase the comprehensive ranking score of new interest circle content, thereby pushing more content that matches the user's new interests.

[0024] In summary, compared to existing marketing content distribution technologies that rely solely on short-term explicit feedback data, this invention comprehensively integrates user long-term and short-term behavioral data, as well as explicit and implicit feedback data, to construct a content intent graph and extract multi-dimensional feature values, forming a systematic intelligent distribution logic. Its core advantage lies in its ability to accurately capture dynamic changes in user interests, quantitatively assess the concentration of content distribution and user fatigue, and dynamically optimize content delivery through multi-objective weight vector adjustment. This effectively alleviates user fatigue caused by excessive push of similar content, while simultaneously improving the matching accuracy between marketing content and user needs, ultimately enhancing marketing conversion efficiency.

[0025] In one embodiment, step S2, which involves obtaining a content intent graph based on the historical behavior data and the content interaction sequence, and obtaining a recommendation fatigue feature coefficient based on the content intent graph, includes: S21. Obtain historical content interaction sequences and historical feedback tags based on the historical behavior data, and perform time-series alignment and splicing of the historical content interaction sequences and the historical feedback tags to obtain user long and short-term behavior sequences. S22. Obtain the user-content interaction relationship in the user's long and short-term behavior sequence, construct a user-content bipartite graph based on the user-content interaction relationship, obtain user node vectors and content node vectors based on the user-content bipartite graph, and obtain a content intent graph based on the user node vectors, the content node vectors and the user-content interaction relationship; S23. Obtain the frequency and average weight of user connections to similar content based on the content intent graph, and obtain the interaction intensity saturation value based on the connection frequency and the average weight; S24. Obtain the most recent interaction timestamp of the user on similar content, and calculate the time decay factor based on the most recent interaction timestamp; S25. Obtain recommended fatigue characteristic coefficients based on the interaction intensity saturation value and the time decay factor.

[0026] As described in steps S21-S25 above, user behavior on marketing platforms exhibits temporal correlation and interest evolution characteristics. Short-term high-frequency interactions may only represent momentary needs, while long-term behavioral sequences better reflect stable interest tendencies. Furthermore, users experience saturation in receiving similar content; exceeding a certain threshold leads to fatigue and negatively impacts marketing effectiveness. Existing technologies rely heavily on short-term behavioral data for interest assessment, failing to fully integrate the correlation between short- and long-term behaviors and lacking a quantitative evaluation mechanism for user fatigue. This makes it difficult to accurately grasp the boundaries of user reception of similar content, easily leading to over-push issues. This step first integrates short- and long-term behavioral data to construct a content intent graph, then extracts interaction intensity saturation values ​​and time decay factors based on the graph, ultimately forming a recommendation fatigue feature coefficient. This specifically addresses the problems of incomplete interest intent mining and ambiguous fatigue assessment in existing technologies, providing quantitative support for dynamic adjustments to distribution strategies. Specifically, it obtains historical content interaction sequences and historical feedback tags based on historical behavioral data, aligns and concatenates them temporally to obtain user short- and long-term behavioral sequences. Historical behavioral data originates from the marketing platform's long-term data storage module, encompassing all past user interactions during platform operation. This includes early browsing, clicks, favorites, comments, and corresponding feedback tags (such as positive reviews, negative reviews, no clear feedback, etc.). This data is continuously collected and stored through the platform's behavioral monitoring tools. Historical content interaction sequences are sequences formed by arranging content-interaction-related behaviors from historical behavioral data in chronological order. Historical feedback tags are identifiers of user attitudes or behavioral outcomes for each interaction. Time-series alignment involves mapping historical content interaction sequences to historical feedback tags according to the timestamps of the behaviors, ensuring each interaction has a corresponding feedback identifier. Stitching integrates the aligned historical data with the current content interaction sequence (i.e., the time-series record of the user's recent content interaction behaviors), forming a long-term and short-term user behavior sequence encompassing both long-term stable behaviors and short-term instantaneous behaviors. For example, a user might browse outdoor camping equipment 2-3 times per month for the past six months (historical content interaction sequence) and repeatedly favorite a particular brand of camping tents (historical feedback tags). Combined with a recent week's browsing of camping cooking equipment, time-series alignment and stitching create a complete long-term and short-term user behavior sequence. The physical significance of this step lies in breaking the limitations of short-term behavioral data, comprehensively capturing the temporal evolution trajectory of user interests, and laying a data foundation for the subsequent construction of accurate content intent maps. Its technical effect is that subsequent interest analysis is no longer limited to instantaneous needs, but can take into account both stable interests and recent dynamics, thereby improving the comprehensiveness of interest intent mining.

[0027] This process involves obtaining user-content interaction relationships from users' long-term and short-term behavior sequences, constructing a user-content bipartite graph based on these relationships, and then obtaining user node vectors and content node vectors from the bipartite graph. Finally, it combines these user-content interaction relationships to obtain a content intent graph. User-content interaction relationships are the associations and association strengths between users and various types of marketing content extracted from users' long-term and short-term behavior sequences. For example, a user's click on content A corresponds to a strong association, while a brief browsing of content B corresponds to a weak association. The association strength is calculated by assigning different weights to behavior types (clicks, favorites, browsing, etc.). When constructing the user-content bipartite graph, users are treated as one set of nodes, marketing content as another set, and the interaction relationship between users and content as the edges connecting the two sets. The weight of each edge represents the corresponding association strength. When obtaining node vectors from this bipartite graph, graph embedding algorithms (such as DeepWalk and Node2Vec) are used to generate context sequences for nodes through random walks. Then, the Word2Vec model is used to map user nodes and content nodes into a low-dimensional vector space, obtaining user node vectors and content node vectors. These vectors can quantitatively represent user interest characteristics and content attribute characteristics. Finally, by combining user node vectors, content node vectors, and the weight distribution of user-content interaction relationships, a content intent graph is constructed. This graph presents the interest relationships between users and content, and between content itself, in a visualized or structured data format. For example, content with similar attributes will form clusters in the graph, and user nodes will form strong connections with content node clusters corresponding to their interests. The core technical implementation of this step lies in transforming unstructured interaction relationships into quantifiable vector features through graph embedding technology. Its physical meaning is to structurally associate user interests with content attributes, making the interest intent hidden in behavioral data explicit. The technical effect is that the constructed content intent graph can clearly present the association logic between users' short-term and long-term interests, providing a structured analytical carrier for subsequent extraction of fatigue-related features.

[0028] The interaction intensity saturation value is obtained by acquiring the frequency and average weight of user connections to similar content based on the content intent graph. In the content intent graph, similar content has been clustered into distinct content clusters through vector clustering. The frequency of user connections to similar content is the sum of the number of connections between the user node and each content node in that cluster. The average weight is the arithmetic mean of the weights corresponding to these connections. Both of these values ​​are calculated by traversing the connection relationships between user nodes and similar content clusters in the content intent graph. The interaction intensity saturation value is calculated using a weighted summation and normalization method. The connection frequency is multiplied by the average weight to obtain the total interaction intensity, which is then divided by a preset upper limit value for interaction intensity (determined based on the average maximum effective interaction intensity of users to similar content according to historical data from the platform). This yields the normalized interaction intensity saturation value, which ranges from 0 to 1. The closer the value is to 1, the closer the user's interaction intensity with that type of content is to saturation. For example, in a content intent graph, content related to "running shoes" forms a cluster. A user node is connected to 10 content nodes within this cluster, with a connection frequency of 10 and an average weight of 0.8. The preset upper limit for interaction intensity is 12. Therefore, the total interaction intensity is 8, and the interaction intensity saturation value is 8 / 12 ≈ 0.67, indicating that the user's interaction intensity with "running shoes" content has reached 67% saturation. The physical meaning of this step is to quantify the user's interaction depth and reception saturation for similar content. The technical effect is to clearly define the user's level of interest and investment in a particular type of content through the interaction intensity saturation value, providing a core quantitative indicator for judging whether user fatigue is imminent.

[0029] Obtain the most recent interaction timestamp of a user with similar content, and calculate the time decay factor based on this timestamp. The most recent interaction timestamp is the time record of the last interaction between a user and a cluster of similar content, extracted from the content intent graph. This data comes from the timestamp information in the user's long and short-term behavior sequences. The time decay factor is calculated using an exponential decay function, with the formula: Time decay factor = e^(-k*(current time - most recent interaction timestamp) / T), where k is the decay coefficient (set according to the type and characteristics of the platform's marketing content; for example, fast-moving consumer goods content has a larger k value and a faster decay rate; durable goods content has a smaller k value and a slower decay rate), and T is the time constant (usually set as the average interest duration period of similar content, obtained from historical data statistics of the platform). For example, if a user's most recent interaction timestamp for "outdoor backpack" content is 30 days ago, the current time is the statistical day, k=0.05, and T=60 days, then the time decay factor = e^(-0.05*(30) / 60) = e^(-0.025)≈0.975, indicating that the user's interest in "outdoor backpack" content has a low degree of decay. If the most recent interaction timestamp is 120 days ago, then the time decay factor = e^(-0.05*(120) / 60) = e^(-0.1)≈0.905, and the degree of interest decay has increased significantly. The physical meaning of this step is to quantify the natural decay of user interest over time. The technical effect is to distinguish the activity level of user interest in similar content under different time dimensions, avoid fatigue judgment bias caused by ignoring the time factor based solely on interaction intensity, and make the subsequent recommendation fatigue feature coefficient more timely.

[0030] The recommendation fatigue characteristic coefficient is obtained based on the interaction intensity saturation value and the time decay factor. It is calculated as the product of the two: Recommendation Fatigue Characteristic Coefficient = Interaction Intensity Saturation Value × Time Decay Factor. This coefficient ranges from 0 to 1. A higher value indicates a higher level of user fatigue with similar content. When the coefficient reaches a preset threshold (determined by the platform based on historical marketing data and user feedback, such as 0.7), it indicates that the user has approached or reached a fatigue threshold, and the push of this type of content should be reduced. For example, if a user's interaction intensity saturation value for "smart home" content is 0.8 and their time decay factor is 0.9, then the recommendation fatigue characteristic coefficient is 0.72, exceeding the preset threshold of 0.7. This indicates that the user has developed significant fatigue with this type of content, and the push priority of "smart home" content should be reduced. The physical significance of this step is to combine the user's interaction saturation with the degree of interest decay for similar content to form a final quantitative assessment of user fatigue. The technical effect is to provide a clear fatigue judgment standard for marketing content distribution, enabling the distribution strategy to dynamically adjust the proportion of similar content pushed based on this coefficient, thereby avoiding user resistance caused by excessive pushes from the source.

[0031] In one embodiment, step S3, which involves obtaining explicit feedback data and implicit feedback data based on the content feedback data, and obtaining multiple content tags based on the explicit feedback data and the content intent graph, includes: S31. The content feedback data is classified into explicit feedback data and implicit feedback data. S32. Obtain the content identifier and behavior intensity value based on the explicit feedback data, and obtain the content node embedding vector from the content intent graph based on the content identifier; S33. Obtain the semantic vector of the tag in the preset content tag library, and calculate the cosine similarity based on the semantic vector of the tag and the embedding vector of the content node; S34. Obtain multiple initial content tags based on the cosine similarity and the behavior intensity value, obtain multiple initial weights corresponding to the multiple initial content tags, and define multiple initial content tags with initial weights greater than a preset threshold as multiple content tags.

[0032] As described in steps S31-S34 above, user feedback on marketing content includes both explicit and implicit forms. Explicit feedback directly reflects the user's active attitude towards the content, while implicit feedback indirectly reflects their interests. Content tags are the extraction of the core attributes of marketing content and serve as a crucial bridge connecting user interests and content distribution. In marketing content distribution scenarios, only by accurately obtaining content tags corresponding to user interests can the distribution density of similar content be effectively monitored, avoiding excessive concentrated pushes. In existing technologies, the extraction of content tags often relies on single-dimensional feedback data or lacks a deep correlation with user behavioral intentions, resulting in insufficient tag extraction accuracy and difficulty in truly reflecting user interests. This, in turn, affects the accuracy of subsequent distribution density monitoring and fails to effectively avoid the problem of excessive pushes of similar content. Therefore, this invention classifies content feedback data into explicit and implicit feedback data. The content feedback data originates from the marketing platform's behavioral data collection module, which records various response data generated by users during their exposure to marketing content, including active operational behaviors and passive browsing behaviors. The classification is based on the initiative and explicitness of user feedback. Explicit feedback data consists of behavioral data initiated by users that directly reflects their attitude towards content, such as clicking to view marketing content details, saving content of interest, commenting on or liking content, or purchasing products promoted by the content. This data is directly collected and recorded through the platform's interaction behavior monitoring tools. Implicit feedback data consists of behavioral data generated by users during browsing that indirectly reflects their interests, such as the full exposure time of content on the user's page, the user's scrolling speed, and whether the user quickly closes the content page. This type of data is collected through the platform's page behavior monitoring tools. The physical significance of this step is to distinguish the attribute characteristics of different feedback data, clarifying which data can directly reflect user interests and which data needs further analysis and mining. Its technical effect is to provide accurate data source filtering for subsequent tag extraction, avoiding tag extraction bias caused by the mixed use of different types of feedback data, ensuring that subsequent tag extraction is based on explicit user interest feedback, and improving the correlation between tags and user interests.

[0033] The platform obtains content identifiers and behavior intensity values ​​from explicit feedback data, and retrieves content node embedding vectors from the content intent graph based on the content identifiers. The content identifier is a unique identification code assigned to each piece of marketing content by the marketing platform to distinguish different marketing content. This identifier corresponds one-to-one with each piece of content and is stored in the platform's content database, directly accessible through content interaction records in the explicit feedback data. The behavior intensity value is a quantitative representation of explicit feedback behavior, used to distinguish the degree to which different explicit feedback behaviors reflect user interest. Based on historical marketing data and behavior conversion value, the platform presets weight values ​​for different types of explicit feedback behaviors. For example, the purchase behavior has the highest weight value (e.g., set to 5), followed by the collection behavior (e.g., set to 3), then the click behavior (e.g., set to 2), and the comment and like behavior has a lower weight value (e.g., set to 1). The behavior intensity value is the preset weight value extracted based on the specific type of explicit feedback behavior of the user. Content node embedding vectors are quantitative representations of the attributes of each marketing content in a content intent graph. A content intent graph is a structured graph built based on users' long-term and short-term behavior sequences, where each content node corresponds to a unique embedding vector. This vector is obtained by transforming the content's attribute features and user interaction features into a low-dimensional vector using graph embedding algorithms (such as Node2Vec). Based on the acquired content identifier, the corresponding content node can be accurately located in the content intent graph, and its embedding vector can then be extracted. The physical significance of this step is to establish a correlation between explicit feedback behavior and content attribute features, quantifying user interest intensity through behavior intensity values, and making content attributes computable through content node embedding vectors. Its technical effect is to provide quantitative foundational data for subsequent tag matching, enabling tag matching to combine user interest intensity and content attribute features, thus improving the accuracy of tag matching. For example, the content node embedding vector corresponding to a user's purchase behavior of a certain marketing content (behavior intensity value 5) can more accurately match tags that the user is truly interested in.

[0034] The system retrieves the semantic vectors of tags from a pre-defined content tag library and calculates cosine similarity based on the tag semantic vectors and content node embedding vectors. The pre-defined content tag library is a standardized set of tags pre-built by the marketing platform based on industry attributes, product categories, and user interests. It covers core attribute tags that all marketing content on the platform may involve, such as "outdoor camping," "smart wearables," and "snacks." Each tag is transformed into a tag semantic vector through a pre-trained semantic model (such as Word2Vec or BERT). This vector quantifies the semantic attributes of the tag. All tags and their corresponding semantic vectors are stored in the platform's tag database and can be directly accessed. Cosine similarity is a calculation method used to measure the angle between two vectors. Its value ranges from -1 to 1. The closer the value is to 1, the higher the semantic similarity between the two vectors, meaning the content attributes corresponding to the content node embedding vector and the tag attributes corresponding to the tag semantic vector are more closely matched. During the calculation, the cosine similarity is calculated one by one between the content node embedding vectors obtained from the content intent graph and the semantic vectors of each tag in the pre-defined content tag library to obtain the semantic matching degree between the content and each tag. The physical meaning of this step is to achieve precise matching between content attributes and tag semantics through vector similarity calculation. Its technical effect is to upgrade the matching of content and tags from traditional keyword matching to deep matching at the semantic level, avoiding the limitations of keyword matching and improving the accuracy of tag matching. For example, for a marketing content about "smart camping lights", the cosine similarity between the embedded vector of its content node and the semantic vector of the tag "outdoor camping" will be much higher than the similarity with the tag "leisure snacks", ensuring that the matched tags conform to the core attributes of the content.

[0035] Multiple initial content tags are obtained based on cosine similarity and behavioral intensity values, and multiple initial weights are corresponding to these tags. Initial content tags with weights greater than a preset threshold are defined as multiple content tags. The initial content tags are selected based on their cosine similarity being greater than a preset similarity threshold (set according to the platform's tag matching accuracy requirements, such as 0.6). These tags have a high semantic match with the current content attributes. The initial weights are calculated by multiplying the cosine similarity by the behavioral intensity value, i.e., Initial Weight = Cosine Similarity × Behavioral Intensity Value. This weight comprehensively considers the semantic match between the tag and the content, as well as the intensity of user feedback behavior, and can more comprehensively reflect the correlation between the tag and user interests. The preset threshold is a critical value used to filter effective content tags. This threshold is determined through statistical analysis of historical platform data. Based on a correlation test between tag weights and actual user interests, a threshold value (such as 0.8) is set to distinguish between effective and invalid tags. Initial content tags with initial weights greater than the preset threshold are selected as the final content tags. This means that these tags not only closely match the content attributes but also reflect genuine user interest through strong explicit user feedback. For example, if a user makes a purchase based on marketing content about a "smart sports watch" (behavior intensity value 5), the cosine similarity between the node embedding vector of this content and the semantic vector of the "smart wearable" tag is 0.9, and the cosine similarity with the "sports equipment" tag is 0.85, with corresponding initial weights of 4.5 and 4.25 respectively, both greater than the preset threshold of 0.8. Therefore, "smart wearable" and "sports equipment" are both selected as content tags. However, the cosine similarity with the "outdoor backpack" tag is 0.5, with an initial weight of 2.5, which is less than the preset threshold, and is therefore excluded. The physical significance of this step is to further improve the effectiveness and relevance of tags through weight calculation and threshold screening, and to eliminate tags that are not closely related to user interests. The technical effect is that the final content tags can accurately and centrally reflect the user's core interests, providing a high-quality tag basis for subsequent distribution density monitoring, avoiding statistical deviations in distribution density due to invalid tags, and ensuring that it is possible to accurately determine whether there is excessive push of similar content.

[0036] In one embodiment, step S4, which involves obtaining feature values ​​from the content set based on a plurality of second content tags and a plurality of second distribution densities, includes: S41. Obtain the frequency of occurrence and the total number of distributed content for each second content tag within a preset statistical period, and obtain the original concentration based on the frequency of occurrence and the total number of distributed content; S42. Obtain the historical average click-through rate of the content corresponding to each second content tag, map the historical average click-through rate to an adjustment factor based on the sigmoid function, and obtain the corrected concentration based on the adjustment factor and the original concentration. S43. Obtain multiple second distribution densities and multiple distribution times of multiple second content tags, and perform time weighting on the multiple second distribution densities according to the multiple distribution times to obtain multiple time-weighted densities; S44. Calculate multiple concentration contribution values ​​based on the corrected concentration degree and multiple time-weighted densities, and calculate content concentration feature values ​​based on the multiple concentration contribution values.

[0037] As described in steps S41-S44 above, when the initial distribution density of some content tags exceeds a preset threshold, it means that the marketing content corresponding to these tags is being pushed to users too frequently, exhibiting an over-concentration trend. This over-concentration is one of the important reasons for user fatigue and reduced marketing conversion efficiency. To solve this problem, it is necessary to accurately quantify the degree of this over-concentration and clarify the impact weight of different over-concentrated tags on the overall distribution effect in order to adjust the distribution strategy accordingly. In existing technologies, the evaluation of content distribution concentration often uses single-dimensional statistical indicators, such as only considering the frequency of tag appearance or distribution density, without comprehensively considering factors such as the popularity of the content itself and the distribution characteristics of distribution time. This results in an incomplete and inaccurate evaluation of concentration, making it difficult to accurately reflect the actual impact of over-concentration on user experience and marketing effectiveness. This step introduces quantitative indicators such as original concentration, corrected concentration, and time-weighted density in stages to gradually correct and improve the evaluation of concentration, ultimately forming a comprehensive content concentration characteristic value. This specifically solves the problem of single-dimensional and insufficiently accurate concentration evaluation in existing technologies, achieving a comprehensive and objective quantification of the degree of over-concentration in marketing content distribution.

[0038] Specifically, the frequency of occurrence and the total number of distributed content for each second content tag within a preset statistical period are obtained, and the raw concentration is calculated based on these two data points. The second content tag is a content tag whose distribution density exceeds a preset threshold, as determined in the preliminary assessment. Its related distribution data is stored in the marketing platform's distribution log database and can be directly retrieved. The preset statistical period is a time interval set based on the marketing platform's distribution frequency and user behavior feedback cycle, such as 7 days or 15 days, used to limit the statistical time range and ensure the targeted and timely nature of the data statistics. Frequency of occurrence refers to the total number of times the marketing content corresponding to the second content tag is pushed to target users within the preset statistical period, which can be calculated by iterating through the relevant records for that tag in the distribution log. The total number of distributed content refers to the total number of all marketing content items pushed to target users by the marketing platform within the same preset statistical period, also obtained by statistically analyzing all push records in the distribution log. The initial concentration ratio is calculated as the ratio of the frequency of occurrence of the second content tag to the total number of distributed content, i.e., Initial Concentration Ratio = Frequency / Total Number of Distributed Content. This ratio directly reflects the proportion of a particular second content tag in the total distributed content. The higher the ratio, the greater the proportion of content corresponding to that tag in the distribution, initially reflecting the degree of distribution concentration. For example, within a preset statistical period of 15 days, if the frequency of occurrence of "leisure snacks" as the second content tag is 30 times, and the total number of distributed content during the same period is 100 times, then its initial concentration ratio is 30 / 100 = 0.3, indicating that the content with this tag accounts for 30% of the total distributed content, initially showing a certain degree of concentration. The physical significance of this step is to establish a preliminary quantitative understanding of distribution concentration through the statistics of basic data and ratio calculation. Its technical effect is to provide basic reference data for subsequent more accurate concentration assessments, giving subsequent corrections and optimizations a clear starting point and avoiding ambiguous judgments about concentration.

[0039] The historical average click-through rate (CTR) of each content corresponding to the second content tag is obtained. This CTR is then mapped to an adjustment factor using the sigmoid function, and finally, a corrected concentration is obtained by combining the adjustment factor with the original concentration. The historical average CTR of the content corresponding to the second content tag refers to the average ratio of user clicks to impressions for all marketing content under that tag in past pushes. This data comes from the marketing platform's historical marketing data statistics database and is calculated by aggregating the historical click and impression records of all content under that tag. It reflects the popularity and user acceptance of the content itself. The sigmoid function is a commonly used non-linear mapping function, expressed as σ(x)=1 / (1+e^(-x)). This function can map the input historical average CTR (typically ranging from 0 to 1) to an adjustment factor between 0 and 1, and has the characteristics of high sensitivity in the middle region and a flattening trend at both ends, effectively distinguishing the adjustment weights corresponding to different CTRs. In the specific mapping process, the historical average click-through rate (CTR) is first standardized (e.g., multiplied by 10), and then input into the sigmoid function to obtain the adjustment factor. For tags with a high historical average CTR, the adjustment factor is closer to 1, indicating that the content of this tag is popular, and even with concentrated distribution, user resistance is relatively low, so the correction weight for its concentration can be appropriately reduced. For tags with a low historical average CTR, the adjustment factor is closer to 0, indicating that the content of this tag is unpopular, and excessive concentrated distribution will exacerbate user resistance, so the correction weight for its concentration needs to be increased. The calculation method for the corrected concentration is the original concentration multiplied by the adjustment factor, i.e., corrected concentration = original concentration × adjustment factor. This calculation corrects the original concentration, ensuring that the concentration assessment considers not only the distribution ratio but also the user acceptance of the content itself. For example, the original concentration of a second content tag, "smart home appliances," is 0.25, with a historical average click-through rate of 0.15. After standardization, this becomes 1.5. Inputting it into the sigmoid function yields an adjustment factor of approximately 0.817, corresponding to a corrected concentration of 0.25 × 0.817 ≈ 0.204. Similarly, the original concentration of another second content tag, "niche books," is also 0.25, with a historical average click-through rate of 0.03. After standardization, this becomes 0.3. Inputting it into the sigmoid function yields an adjustment factor of approximately 0.575, corresponding to a corrected concentration of 0.25 × 0.575 ≈ 0.144. This demonstrates that even with the same original concentration, tags with lower popularity have lower corrected concentrations, better reflecting the negative impact of excessive concentration. The physical significance of this step is to incorporate the user acceptance of the content itself into the concentration assessment, avoiding the bias caused by judging concentration solely based on distribution share. Its technical effect is to make the concentration assessment more in line with actual marketing scenarios, distinguish the differences in excessive concentration of different popularity tags, and improve the rationality and accuracy of the concentration assessment.

[0040] Multiple distribution densities and distribution times for multiple second content tags are obtained. The second distribution densities are then time-weighted based on their distribution times to obtain a time-weighted density. The second distribution density is the actual distribution density value corresponding to the second content tag that exceeds a preset threshold. This data was already calculated and stored in the relevant data module when determining whether the first distribution density exceeded the preset threshold. The distribution time refers to the specific time point when the marketing content was pushed for each second distribution density, sourced from the marketing platform's distribution log database, and corresponds one-to-one with the second distribution density. The core logic of time weighting is to consider the impact of the proximity of the distribution time on user perception. Recent distribution activities have a greater impact on users' current fatigue and resistance, and therefore are given higher weights, while distant distribution activities have a smaller impact and are given lower weights. In practice, an exponential weighting method is used. A time decay coefficient is set (based on the update frequency of the platform's marketing content and the user's memory cycle, such as 0.1). Using the current time as the benchmark, the time difference between each distribution time and the current time is calculated. The time weight is e^(-time decay coefficient × time difference). The second distribution density is then multiplied by the corresponding time weight to obtain the time-weighted density, i.e., time-weighted density = second distribution density × time weight. For example, the two distribution densities of a certain second content tag, "sports apparel," are 0.4 (distributed 3 days ago) and 0.35 (distributed 10 days ago), respectively. With a time decay coefficient set to 0.1, the time weight for 3 days ago is e^(-0.1×3)≈0.741, and the time weight for 10 days ago is e^(-0.1×10)≈0.368. Therefore, the corresponding time-weighted densities are 0.4×0.741≈0.296 and 0.35×0.368≈0.129, respectively. The recent distribution density, after weighting, contributes more to the overall evaluation. The physical meaning of this step is to quantify the impact of distribution time on concentration perception, reflecting the objective law that users are more sensitive to recent over-distribution. Its technical effect is to make density evaluation more timely, accurately capturing the impact of recently over-distributed tags on overall concentration, and avoiding lag in concentration evaluation caused by ignoring the time factor.

[0041] Multiple concentration contribution values ​​are calculated based on the corrected concentration degree and multiple time-weighted densities, and then a content concentration feature value is calculated based on these multiple concentration contribution values. The concentration contribution value is calculated as the product of the corrected concentration degree and the corresponding time-weighted density for each second content tag, i.e., Concentration Contribution Value = Corrected Concentration Degree × Time-Weighted Density. This value comprehensively considers the corrected concentration degree of the tag, content user acceptance, and the impact of distribution time, quantifying the actual contribution of each second content tag to the overall over-concentration of marketing content distribution. The content concentration feature value is the average of the concentration contribution values ​​of all second content tags, i.e., Content Concentration Feature Value = Sum of All Concentration Contribution Values ​​ / Number of Second Content Tags. This averaging calculation yields a comprehensive quantitative indicator reflecting the degree of over-concentration in overall marketing content distribution. This indicator ranges from 0 to 1; a higher value indicates a more severe degree of over-concentration, and vice versa. For example, in a marketing scenario, there are three secondary content tags with concentration contribution values ​​of 0.18, 0.15, and 0.12, respectively. The content concentration feature value is (0.18 + 0.15 + 0.12) / 3 = 0.15, indicating a certain degree of over-concentration in the overall marketing content distribution. The physical meaning of this step is to integrate the multi-dimensional evaluation data of all over-concentrated tags to form a final quantitative conclusion on the overall distribution concentration. Its technical effect is to provide a clear and unified decision-making basis for subsequent distribution strategy adjustments, enabling the marketing platform to accurately determine whether and to what extent the distribution strategy needs to be adjusted based on the content concentration feature value. For example, when the content concentration feature value exceeds a preset threshold (such as 0.2), the platform can significantly reduce the proportion of push notifications for over-concentrated tags, thereby alleviating user fatigue.

[0042] In one embodiment, step S5, which involves obtaining the user's ignore duration and quick skip behavior for distributed content based on the implicit feedback data, and obtaining cross-circle flow tendency characteristic values ​​based on the ignore duration, the quick skip behavior, and the content interaction sequence, includes: S51. Extract the complete exposure duration of each distributed content from the implicit feedback data, and use the exposure duration that is less than the first time threshold as the ignore duration. In this case, the complete exposure duration that is less than the second time threshold and is accompanied by fast scrolling or page closing behavior is used as fast skip behavior. S52. Obtain multiple content tags that are skipped within a preset continuous time window based on the fast skip behavior, and obtain the frequency of fast skip behavior based on the multiple content tags and the preset continuous time window. S53. Obtain the first positive interaction content tag after the quick skip action according to the content interaction sequence, and use the first positive interaction content tag as the new interaction tag; S54. Input each content tag and the new interaction tag into the pre-trained tag semantic model to obtain multiple content tag semantic vectors and new interaction tag semantic vectors, and calculate multiple vector distances between the multiple content tag semantic vectors and the new interaction tag semantic vectors, and use the vector distances as semantic distances; S55. Obtain cross-sphere flow tendency feature values ​​based on multiple semantic distances, the ignoring duration, and the frequency of fast skipping behavior.

[0043] As described in steps S51-S55 above, implicit feedback data from users on marketing platforms (such as exposure duration, scrolling behavior, page closing actions, etc.) can reflect changes in user interests. When users lose interest in a certain type of content currently being pushed, they often exhibit this through implicit behaviors such as quickly skipping or ignoring it, and this may be accompanied by a shift to a new interest group. In marketing content distribution scenarios, if these implicit signals and interest shift trends are not captured in time, and content from the original interest group continues to be pushed, it will not only exacerbate user fatigue but also miss the opportunity to push content from new interest groups, affecting marketing effectiveness. In existing technologies, the capture of changes in user interests largely relies on explicit feedback data, and the mining of implicit feedback data is not deep enough. Furthermore, there is a lack of quantitative evaluation methods for interest group shift trends, making it difficult to accurately determine when users need to leave their current interest group and turn to a new interest direction. This step extracts key behavioral indicators from implicit feedback in stages and combines them with semantic association analysis of content tags to construct a quantitative evaluation system for cross-circle flow trends. This specifically addresses the problems of insufficient implicit feedback mining and ambiguous judgment of interest shift trends in existing technologies, and achieves accurate perception of dynamic changes in user interests.

[0044] Specifically, the complete exposure duration of each distributed piece of content is extracted from implicit feedback data. Exposure durations shorter than a first time threshold are ignored, while those shorter than a second time threshold accompanied by rapid scrolling or page closing are defined as "quick skipping." Implicit feedback data originates from the marketing platform's behavior data collection module. This module monitors user behavior in real-time, collecting data such as content exposure duration, scrolling speed, and page actions, which are stored in a database and can be directly retrieved. Complete exposure duration refers to the length of time the content is fully displayed on the user's page, starting from when the content finishes loading and ending when the user scrolls, closes the page, or switches pages. The first and second time thresholds are thresholds set based on historical user browsing time statistics from the platform. The first time threshold (e.g., 8 seconds) is used to determine if the user has basic attention to the content, and the second time threshold (e.g., 3 seconds) is used to determine if the user exhibits significant resistance, with the second time threshold being shorter than the first time threshold. When the total exposure time is less than the first time threshold, it indicates that the user has not paid effective attention to the content, and this time is considered an ignored duration. When the total exposure time is less than the second time threshold, and the user is simultaneously detected to be scrolling the page rapidly (the scrolling speed exceeds a preset speed threshold, such as 500 pixels per second) or closing the page directly, it indicates that the user has a clear aversion to the content, and this is judged as a rapid skipping behavior. For example, when a user browses a marketing piece of "traditional home appliances," the total exposure time is only 2 seconds, accompanied by rapid page scrolling. Since 2 seconds is less than the second time threshold of 3 seconds, this behavior is judged as a rapid skipping behavior. However, when browsing another piece of "smart home" content, the total exposure time is 6 seconds, which is less than the first time threshold of 8 seconds but greater than the second time threshold of 3 seconds, and this 6 seconds is considered an ignored duration. The physical significance of this step is to transform the duration information and operational behavior in implicit feedback data into quantifiable and measurable user attitude indicators. Its technical effect is to accurately identify the user's state of ignoring or resisting the distributed content, providing a basic behavioral basis for subsequent analysis of changes in user interests and avoiding misjudgments of user interests due to ignoring implicit behavioral signals.

[0045] The algorithm retrieves multiple content tags skipped within a preset continuous time window based on the quick-skip behavior, and combines these content tags with the preset continuous time window to obtain the frequency of quick-skip behavior. The content tags corresponding to quick-skip behavior refer to the core tags associated with the marketing content quickly skipped by users. These tags originate from multiple content tags extracted in the early stages through explicit feedback data combined with content intent mapping. Each distributed piece of content is pre-associated with corresponding content tags and stored in a content-tag association database, which can be retrieved by associating the content with the quick-skip behavior. The preset continuous time window is a time interval (e.g., 30 minutes) set based on the continuity characteristics of user behavior, used to statistically analyze the concentration of quick-skip behavior within a short period, avoiding weakened behavioral correlation due to excessively long time spans. The frequency of quick-skip behavior refers to the total number of times a user performs a quick-skip behavior within the preset continuous time window, obtained by counting the number of records corresponding to all quick-skip behaviors within that time window. For example, within a preset 30-minute continuous time window, a user quickly skipped three items: "Traditional Home Appliances," "Kitchenware," and "Home Building Materials," with corresponding content tags of "Traditional Home Appliances," "Kitchenware," and "Home Building Materials," respectively. This quick-skipping behavior occurred three times. The physical meaning of this step is to capture the concentrated aversion trend of users to certain categories of content by statistically analyzing the quick-skipping behavior and corresponding tags within a short period. The technical effect is to transform a single quick-skipping behavior into correlated behavioral trend data, providing a basis for determining whether users have a general aversion to the content of their current interest group. When the frequency of quick-skipping behavior is high and the corresponding tags are concentrated in a certain interest group, it indicates that the user may have become tired of the content in that group.

[0046] The first positive interaction content tag following a quick skip action is obtained from the content interaction sequence and used as the new interaction tag. The content interaction sequence is a chronological record of user interactions with all marketing content on the marketing platform, including various interactive behaviors such as quick skip, ignore, click, favorite, and purchase. This sequence is continuously recorded and stored by the behavior data collection module and arranged in chronological order. Positive interaction behaviors refer to positive interactions that reflect user interests, including clicks, favorites, purchases, comments, and likes, corresponding to negative behaviors such as quick skip and ignore. In the content interaction sequence, after locating the time node corresponding to the quick skip action, the first positive interaction behavior is retrieved. The content tag associated with the marketing content corresponding to this positive interaction behavior is the first positive interaction content tag, or the new interaction tag. For example, if a user's content interaction sequence is: quickly skipping "traditional home appliances" content → quickly skipping "kitchen utensils" content → clicking "smart wearables" content → favorited "sports equipment" content, the first positive interaction behavior after the quick skip action is clicking "smart wearables" content, and the corresponding content tag "smart wearables" is the new interaction tag. The physical significance of this step is to capture the user's new interest after resisting the current content, and to clarify the specific direction of the user's interest shift. Its technical effect is to associate the user's negative behavior with subsequent positive behavior, and to discover the user's new interest tendencies from the comparison of behaviors, providing a clear basis for judging the target circle of cross-circle flow.

[0047] The process involves inputting each content tag and new interaction tag into a pre-trained tag semantic model, resulting in multiple semantic vectors for both content tags and new interaction tags. The vector distance between these vectors is then calculated and used as the semantic distance. Each content tag refers to all core content tags extracted earlier (including tags quickly skipped by the user and other related tags), while the new interaction tag corresponds to the user's new interests determined in step S53. The pre-trained tag semantic model is a deep learning model (such as BERT or Word2Vec) trained on a large amount of tag semantic data. This model can convert text-based tags into low-dimensional vectors with semantic relationships. The model parameters are fixed after being trained and optimized using a large-scale tag corpus and deployed in the algorithm module of the marketing platform, allowing for direct vector conversion. Each content tag and new interaction tag is input into this model, which outputs corresponding high-dimensional feature vectors. After dimensionality reduction, low-dimensional semantic vectors for the content tags and new interaction tags are obtained. Vector distance, calculated using common methods such as Euclidean distance or cosine distance, measures the similarity between the semantic vectors of two tags. A larger vector distance indicates a greater difference in the semantic categories to which the two tags belong, i.e., a greater difference in their corresponding interest circles. Conversely, a smaller vector distance indicates a stronger semantic correlation between the two tags, suggesting they belong to the same or similar interest circles. For example, when the content tag "traditional home appliances" and the new interactive tag "smart wearables" are input into the model, the resulting semantic vector distance is large, indicating a significant difference in their interest circles. Conversely, the semantic vector distance between the content tag "smartwatch" and the new interactive tag "smart wearables" is small, indicating they belong to similar interest circles. The physical significance of this step is that it quantifies the differences in interest circles between different tags through semantic vector transformation and distance calculation. Its technical effect is to transform the semantic correlation of tags into a calculable quantitative indicator, providing a core quantitative basis for judging whether users move across interest circles and avoiding subjective judgment biases regarding interest circles.

[0048] Cross-circle flow tendency feature values ​​are obtained based on multiple semantic distances, ignore durations, and the frequency of quick skipping behaviors. During calculation, semantic distance, ignore duration, and the frequency of quick skipping behaviors are first standardized, transforming them into normalized values ​​between 0 and 1 to eliminate dimensional differences between different indicators. The standardization process uses the Min-Max normalization method, where for each indicator, the normalized value = (original value - minimum value) / (maximum value - minimum value), where the minimum and maximum values ​​are determined based on the indicator value ranges statistically analyzed from historical platform data. Subsequently, preset weights are assigned to the three indicators. The weight allocation is based on the degree of influence of each indicator on the cross-circle flow tendency, determined through correlation analysis between historical platform behavior data and interest transfer results. For example, the weight of quick skipping behavior frequency is set to 0.4, the weight of semantic distance is set to 0.35, and the weight of ignore duration is set to 0.25 (the sum of the weights is 1). Finally, the three normalized indicators are multiplied by their corresponding weights and summed to obtain the cross-circle flow tendency feature value, i.e., Cross-circle flow tendency feature value = (Normalized semantic distance × weight 1) + (Normalized ignore duration × weight 2) + (Normalized fast skip behavior frequency × weight 3). This feature value ranges from 0 to 1. The higher the value, the more obvious the user's tendency to flow across circles, and the more necessary it is to adjust the distribution strategy and push content from the interest circle to which the new interaction tag belongs. For example, if a user's normalized semantic distance is 0.8, normalized ignore duration is 0.7, and normalized fast skip behavior frequency is 0.9, with corresponding weights of 0.35, 0.25, and 0.4 respectively, then the cross-circle flow tendency feature value = 0.8 × 0.35 + 0.7 × 0.25 + 0.9 × 0.4 = 0.28 + 0.175 + 0.36 = 0.815, indicating that the user has a strong tendency to flow across circles. The physical significance of this step is to comprehensively and quantitatively assess users' tendency to move across different interest groups by integrating multiple key indicators. Its technical effect is to integrate scattered behavioral and semantic data into unified quantitative indicators, providing a clear and actionable basis for subsequent distribution strategy adjustments. This enables marketing platforms to accurately determine whether to push content from new interest groups to users based on the magnitude of this feature value.

[0049] In one embodiment, step S6, which involves constructing a marketing content set based on the recommendation fatigue characteristic coefficient, the content set characteristic value, and the cross-circle flow tendency characteristic value, and then intelligently distributing marketing content to users based on the marketing content set to obtain an intelligent distribution result, includes: S61. Obtain multiple candidate marketing content to be distributed, and obtain the release time and current time of each candidate marketing content, and obtain a real-time freshness score based on the release time and the current time; S62. Obtain the target user node vector and the candidate content node vector corresponding to each candidate marketing content based on the content intent graph, and calculate the inner product of each candidate content node vector and the target user node vector, and use the inner product as the graph relevance score. S63. Based on the recommended fatigue feature coefficient, the content set feature value and the cross-sphere flow tendency feature value, construct a multi-objective adjustment weight vector; S64. The multiple graph relevance scores, multiple freshness scores and the multi-objective adjustment weight vector are weighted and fused to obtain the comprehensive ranking score of the multiple candidate marketing content; S65. Arrange the multiple candidate marketing contents in descending order according to the comprehensive ranking score, select the top N contents to form a marketing content set, and encapsulate the contents in the marketing content set into push tasks in order, perform the distribution operation and record the distribution log, and use the distribution log as the intelligent distribution result.

[0050] As described in steps S61-S65 above, the core objective of marketing content distribution is to meet user interests while avoiding excessive push of similar content, and to maintain user attention by considering the timeliness of the content. The three key features extracted in the previous steps comprehensively reflect the user's current content reception status and changing needs from three dimensions: user fatigue, content distribution concentration, and user interest shift trends. The freshness of candidate marketing content directly affects users' willingness to click, while interest matching determines the degree to which the content aligns with user needs. Only by comprehensively considering these dimensions can a scientific and reasonable distribution strategy be formulated, avoiding distribution bias caused by single-dimensional decisions. In existing technologies, the ranking of marketing content distribution often focuses on interest matching or a single feedback indicator, lacking a comprehensive integration of user fatigue, content concentration, and interest shift trends. This leads to rigid distribution strategies that struggle to balance user interests, content diversity, and timeliness, and fail to effectively address the dynamic changes in user needs. This step constructs a multi-objective adjustment weight vector to weight and fuse the adjustment requirements corresponding to interest matching degree, freshness and the three major feature values ​​to form a comprehensive ranking basis. It specifically solves the problems of single distribution decision dimension and lack of dynamic adjustment mechanism in the existing technology, and achieves a balance of multi-dimensional demands.

[0051] Specifically, the process involves acquiring multiple candidate marketing content pieces to be distributed, along with the publication time and current time for each piece, and calculating a real-time freshness score based on these two data points. The candidate marketing content originates from the marketing platform's content library, which stores all marketing materials to be pushed, covering various tags and themes, and is regularly uploaded and updated by operations staff. The publication time of each candidate marketing content piece is recorded synchronously upon uploading to the content library and stored in the content attribute database, accessible through content identifiers. The current time is the real-time time at the time the system executes the distribution operation, provided by the platform's server clock. The real-time freshness score is calculated using a freshness decay function, with the formula: Real-time Freshness Score = e^(-k × (Current Time - Publication Time) / T), where k is the decay coefficient (set according to content type; for example, news content has a larger k value and decays faster, while product introduction content has a smaller k value and decays slower), and T is a time constant (set as the average effective freshness period for similar content, obtained from historical platform data). The score ranges from 0 to 1; the closer the publication time is to the current time, the higher the freshness score, indicating stronger timeliness and a higher likelihood of user clicks. For example, if a news-type candidate marketing content was published 1 hour ago, and the current time is the statistical time, k=0.2, T=24 hours, then the real-time freshness score = e^(-0.2×(1) / 24)≈0.992, which is extremely timely; while another product introduction content was published 72 hours ago, k=0.1, T=72 hours, then the real-time freshness score = e^(-0.1×(72) / 72)=e^(-0.1)≈0.905, which still maintains a certain level of freshness. The physical meaning of this step is to quantify the timeliness value of candidate marketing content, and its technical effect is to provide a timeliness basis for distribution and sorting, ensuring that the pushed content can meet the user's demand for fresh information, avoiding the decline in user attention due to the push of outdated content, and at the same time, the setting of decay coefficients for different types of content ensures the pertinence of freshness evaluation.

[0052] The content intent graph obtains the target user node vector and the candidate content node vector corresponding to each candidate marketing content, and calculates their inner product as the graph relevance score. The content intent graph is a structured graph constructed based on users' long-term and short-term behavior sequences. The target user node vector is a quantitative expression of the target user's interest features in the graph, while the candidate content node vector is a quantitative expression of the candidate marketing content's attribute features. Both are transformed using graph embedding algorithms (such as Node2Vec) and stored in the graph database, and can be retrieved using user and content identifiers respectively. The inner product calculation is a common method for measuring the similarity between two vectors. The larger the result, the higher the fit between the target user node vector and the candidate content node vector, i.e., the stronger the match between the candidate marketing content and the user's interests. For example, if the target user node vector emphasizes the interest feature of "outdoor camping," and the candidate content node vector corresponding to a certain candidate marketing content emphasizes the attribute of "camping tent," their inner product is high, resulting in a high graph relevance score, indicating a high match between the content and the user's interests. Conversely, the inner product between the candidate content node vector emphasizing the attribute of "kitchenware" and the target user node vector is low, resulting in a low graph relevance score. The physical meaning of this step is to quantify the degree of matching between candidate marketing content and user interests. Its technical effect is to provide a core interest matching basis for distribution ranking, ensuring that the pushed content can meet the user's basic interest needs, avoiding the push of content that is irrelevant to the user's interests, and improving the accuracy of content distribution.

[0053] Based on recommendation fatigue feature coefficients, content concentration feature values, and cross-circle flow tendency feature values, a multi-objective adjustment weight vector is constructed. The recommendation fatigue feature coefficient reflects the degree of user fatigue with similar content; the content concentration feature value reflects the degree of over-concentration in content distribution; and the cross-circle flow tendency feature value reflects the trend of users shifting to new interest circles. These three feature values ​​are calculated in previous steps, stored in the feature database, and can be directly accessed. The construction process of the multi-objective adjustment weight vector involves assigning a corresponding adjustment weight to each feature value, which, together with the weights of the graph relevance score and freshness score, forms a complete weight vector. The weight allocation is dynamically adjusted based on the influence of the feature value on the distribution strategy, and is trained and optimized by the platform based on historical marketing data and user feedback. Specifically, the weight corresponding to the fatigue feature coefficient is used to reduce the ranking priority of content with high fatigue levels; the higher the coefficient, the lower the weight of the corresponding content. The weight corresponding to the content concentration feature value is used to reduce the ranking priority of overly concentrated content; the higher the feature value, the lower the weight of the corresponding content. The weight corresponding to the cross-circle flow tendency feature value is used to increase the ranking priority of content in new interest circles; the higher the feature value, the higher the weight of content in new interest circles. Simultaneously, the graph relevance score and freshness score are assigned basic weights to ensure the core status of interest matching and timeliness. For example, when the recommendation fatigue feature coefficient is 0.8 (high fatigue), the content concentration feature value is 0.7 (high concentration), and the cross-circle flow tendency feature value is 0.9 (strong transfer tendency), the constructed multi-objective adjustment weight vector might be: graph relevance score weight 0.3, freshness score weight 0.2, recommendation fatigue adjustment weight -0.2 (negative weight indicates reduced priority), content concentration adjustment weight -0.15 (negative weight), and cross-circle flow adjustment weight 0.15 (positive weight), achieving a balanced adjustment of multiple objectives through weight allocation. The physical significance of this step is to transform the adjustment requirements corresponding to the three major feature values ​​into quantifiable weight indicators. Its technical effect is to provide a dynamic adjustment basis for subsequent comprehensive ranking, so that the ranking results can respond to changes in user fatigue, content concentration, and interest shift trends, avoiding over-pushing or interest disconnect caused by single-dimensional ranking.

[0054] Multiple graph relevance scores, multiple freshness scores, and a multi-objective adjustment weight vector are weighted and fused to obtain a comprehensive ranking score for multiple candidate marketing content. The weighted fusion calculation method is as follows: Comprehensive Ranking Score = (Graph Relevance Score × Graph Relevance Weight) + (Freshness Score × Freshness Weight) + (Recommendation Fatigue Feature Coefficient of Candidate Content's Corresponding Tags × Recommendation Fatigue Adjustment Weight) + (Content Set Feature Value of Candidate Content's Corresponding Tags × Content Set Adjustment Weight) + (Cross-Circle Flow Tendency Feature Value × Cross-Circle Flow Adjustment Weight). The recommendation fatigue feature coefficient and content set feature value of the candidate content's corresponding tags are obtained by associating the candidate content's tags with the previously calculated feature values. That is, each candidate content has been pre-associated with corresponding content tags, and the corresponding feature value can be obtained through tag matching. For example, if a candidate marketing content has a graph relevance score of 0.8, a freshness score of 0.75, a corresponding recommendation fatigue feature coefficient of 0.8, a content concentration feature value of 0.7, a cross-circle flow tendency feature value of 0.9, and a multi-objective adjustment weight vector of [0.3, 0.2, -0.2, -0.15, 0.15], then the comprehensive ranking score = 0.8 × 0.3 + 0.75 × 0.2 + 0.8 × (-0.2) + 0.7 × (-0.15) + 0.9 × 0.15 = 0.24 + 0.15 - 0.16 - 0.105 + 0.135 = 0.26. The physical significance of this step is to integrate multi-dimensional evaluation indicators to form a comprehensive evaluation of candidate marketing content. Its technical effect is to organically combine interest matching degree, timeliness and user status, and the adjustment needs of content distribution status, so that the comprehensive ranking score can fully reflect the distribution value of candidate content, avoid the advantages of a single indicator from covering the shortcomings of other dimensions, and ensure the scientificity and rationality of the ranking results.

[0055] Multiple candidate marketing content items are sorted in descending order based on their comprehensive ranking scores. The top N items are selected to form a marketing content set. The content in this set is then packaged into push tasks in sequence, distributed, and a distribution log is recorded. This log serves as the intelligent distribution result. Descending order means ranking all candidate marketing content from highest to lowest comprehensive ranking score; higher-scoring content has higher distribution priority. N is a preset push quantity threshold, set by the platform based on user receiving habits and push capacity (e.g., 10 or 15 items per push). Selecting the top N items to form the marketing content set ensures that all pushed content is high-quality content with the highest comprehensive value. When the content is packaged into push tasks in sequence, the push sequence is allocated according to the sorted order, prioritizing the push of higher-ranking content to ensure users first encounter the content most relevant to their needs. The distribution operation is executed by the marketing platform's push module, pushing the content to the user's terminal according to the user's receiving channel (e.g., app push, SMS push, in-app message push, etc.). The distribution log is recorded by the log module, including information such as push content identifier, push time, push channel, and user identifier. This information is stored in the distribution log database as the result of intelligent distribution, providing a basis for subsequent distribution strategy optimization and data iteration. For example, if there are 50 candidate marketing content items, and N is set to 10, after comprehensive sorting, the top 10 items with the highest scores are selected to form a set of marketing content. These are then packaged into push tasks in sorted order and pushed to users through the app. The push details for each item are recorded, forming a distribution log. The physical meaning of this step is to transform the comprehensively evaluated high-quality content into actual distribution actions and record the distribution process. Its technical effect is to achieve accurate and orderly distribution of marketing content, ensuring that the pushed content not only matches users' interests and timeliness needs but also alleviates fatigue and avoids excessive concentration.

[0056] This application also provides an intelligent marketing content distribution system based on user behavior data, including: The behavior data collection module is used to acquire user behavior data on the marketing platform, including content interaction sequences and content feedback data. The intent graph and fatigue analysis module is used to acquire users' historical behavior data on the marketing platform, and to acquire a content intent graph based on the historical behavior data and the content interaction sequence, and to acquire recommendation fatigue feature coefficients based on the content intent graph. The content tag and density monitoring module is used to obtain explicit feedback data and implicit feedback data based on the content feedback data, obtain multiple content tags based on the explicit feedback data and the content intent graph, obtain the first distribution density corresponding to each content tag, and sequentially determine whether each first distribution density exceeds a preset threshold. The content concentration measurement module is used to determine that the content tag corresponding to the first distribution density has excessive concentration of marketing content distribution if the first distribution density exceeds a preset threshold, and to count all second content tags that exceed the threshold and their second distribution densities, and to obtain content concentration feature values ​​based on multiple second content tags and multiple second distribution densities. The cross-circle flow analysis module is used to obtain the duration of user ignoring distributed content and the behavior of quickly skipping content based on the implicit feedback data, and to obtain cross-circle flow tendency characteristic values ​​based on the ignoring duration, the behavior of quickly skipping content and the content interaction sequence. The intelligent distribution decision module is used to construct a marketing content set based on the recommendation fatigue feature coefficient, the content set feature value, and the cross-circle flow tendency feature value, and to intelligently distribute marketing content to users based on the marketing content set to obtain intelligent distribution results.

[0057] In one embodiment, the intent mapping and fatigue analysis module includes: The behavior sequence splicing unit is used to obtain historical content interaction sequences and historical feedback tags based on the historical behavior data, and to perform time-series alignment and splicing of the historical content interaction sequences and historical feedback tags to obtain user long and short-term behavior sequences. The intent graph construction unit is used to obtain the user-content interaction relationship in the user's long and short-term behavior sequences, construct a user-content bipartite graph based on the user-content interaction relationship, obtain user node vectors and content node vectors based on the user-content bipartite graph, and obtain a content intent graph based on the user node vectors, the content node vectors and the user-content interaction relationship. An interaction intensity calculation unit is used to obtain the frequency and average weight of user connections to similar content based on the content intent graph, and to obtain the interaction intensity saturation value based on the connection frequency and the average weight. The time decay calculation unit is used to obtain the user's most recent interaction timestamp with similar content, and calculate the time decay factor based on the most recent interaction timestamp. The fatigue feature calculation unit is used to obtain recommended fatigue feature coefficients based on the interaction intensity saturation value and the time decay factor.

[0058] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0059] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0060] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0061] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0062] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for intelligent distribution of marketing content based on user behavior data, characterized in that, include: Acquire user behavior data on the marketing platform, including content interaction sequences and content feedback data; Obtain historical behavior data of users on the marketing platform, and obtain a content intent graph based on the historical behavior data and the content interaction sequence, and obtain a recommendation fatigue feature coefficient based on the content intent graph; Explicit feedback data and implicit feedback data are obtained based on the content feedback data, and multiple content tags are obtained based on the explicit feedback data and the content intent graph. The first distribution density corresponding to each content tag is obtained, and it is determined in turn whether each first distribution density exceeds a preset threshold. If the first distribution density exceeds a preset threshold, it is determined that the content tag corresponding to the first distribution density has excessive concentration of marketing content distribution, and all second content tags exceeding the threshold and their second distribution densities are counted, and content concentration feature values ​​are obtained based on multiple second content tags and multiple second distribution densities. The implicit feedback data is used to obtain the duration of user ignoring distributed content and the behavior of quickly skipping it. Based on the ignoring duration, the behavior of quickly skipping it and the content interaction sequence, the cross-circle flow tendency feature value is obtained. A marketing content set is constructed based on the recommended fatigue characteristic coefficient, the content set characteristic value, and the cross-circle flow tendency characteristic value. The marketing content is then intelligently distributed to users based on the marketing content set to obtain the intelligent distribution result.

2. The intelligent distribution method for marketing content based on user behavior data according to claim 1, characterized in that, The steps of obtaining a content intent graph based on the historical behavior data and the content interaction sequence, and obtaining a recommendation fatigue feature coefficient based on the content intent graph, include: Based on the historical behavior data, historical content interaction sequences and historical feedback tags are obtained, and the historical content interaction sequences and historical feedback tags are time-series aligned and spliced ​​to obtain user long and short-term behavior sequences. Obtain the user-content interaction relationship in the user's long and short-term behavior sequences, construct a user-content bipartite graph based on the user-content interaction relationship, obtain user node vectors and content node vectors based on the user-content bipartite graph, and obtain a content intent graph based on the user node vectors, the content node vectors and the user-content interaction relationship; The frequency and average weight of user connections to similar content are obtained based on the content intent graph, and the interaction intensity saturation value is obtained based on the connection frequency and the average weight. Obtain the user's most recent interaction timestamp for similar content, and calculate the time decay factor based on the most recent interaction timestamp; The recommended fatigue characteristic coefficients are obtained based on the interaction intensity saturation value and the time decay factor.

3. The intelligent distribution method for marketing content based on user behavior data according to claim 1, characterized in that, The step of obtaining explicit feedback data and implicit feedback data based on the content feedback data, and obtaining multiple content tags based on the explicit feedback data and the content intent graph, includes: The content feedback data is categorized into explicit feedback data and implicit feedback data. The content identifier and behavior intensity value are obtained based on the explicit feedback data, and the content node embedding vector is obtained from the content intent graph based on the content identifier. Obtain the semantic vectors of tags from a preset content tag library, and calculate the cosine similarity based on the semantic vectors of the tags and the embedding vectors of the content nodes; Multiple initial content tags are obtained based on the cosine similarity and the behavior intensity value, and multiple initial weights are obtained corresponding to the multiple initial content tags. Multiple initial content tags with initial weights greater than a preset threshold are defined as multiple content tags.

4. The intelligent distribution method for marketing content based on user behavior data according to claim 1, characterized in that, The step of obtaining feature values ​​in the content set based on multiple second content tags and multiple second distribution densities includes: Obtain the frequency of occurrence and the total number of distributed content for each second content tag within a preset statistical period, and obtain the original concentration based on the frequency of occurrence and the total number of distributed content; Obtain the historical average click-through rate of the content corresponding to each second content tag, map the historical average click-through rate to an adjustment factor based on the sigmoid function, and obtain the corrected concentration based on the adjustment factor and the original concentration. Obtain multiple second distribution densities and multiple distribution times for multiple second content tags, and perform time weighting on the multiple second distribution densities according to the multiple distribution times to obtain multiple time-weighted densities; Multiple concentration contribution values ​​are calculated based on the corrected concentration and multiple time-weighted densities, and content concentration feature values ​​are calculated based on the multiple concentration contribution values.

5. The intelligent distribution method for marketing content based on user behavior data according to claim 1, characterized in that, The step of obtaining the user's ignore duration and quick skip behavior for distributed content based on the implicit feedback data, and obtaining cross-circle flow tendency feature values ​​based on the ignore duration, the quick skip behavior, and the content interaction sequence, includes: Extract the complete exposure duration of each distributed content from the implicit feedback data, and treat the exposure duration that is less than the first time threshold as the ignore duration. In particular, treat the complete exposure duration that is less than the second time threshold and is accompanied by rapid scrolling or page closing behavior as rapid skip behavior. Based on the fast skip behavior, obtain multiple content tags that are skipped within a preset continuous time window, and obtain the frequency of fast skip behavior based on the multiple content tags and the preset continuous time window; The first positive interaction content tag after the quick skip action is obtained according to the content interaction sequence, and the first positive interaction content tag is used as the new interaction tag; Each content tag and the new interaction tag are respectively input into the pre-trained tag semantic model to obtain multiple content tag semantic vectors and new interaction tag semantic vectors. Multiple vector distances between the multiple content tag semantic vectors and the new interaction tag semantic vectors are calculated, and the vector distances are used as semantic distances. Cross-sphere flow tendency feature values ​​are obtained based on multiple semantic distances, the ignoring duration, and the frequency of fast skipping behavior.

6. The intelligent distribution method for marketing content based on user behavior data according to claim 1, characterized in that, The step of constructing a marketing content set based on the recommended fatigue characteristic coefficient, the content set characteristic value, and the cross-circle flow tendency characteristic value, and then intelligently distributing marketing content to users based on the marketing content set to obtain intelligent distribution results includes: Get multiple candidate marketing content to be distributed, and get the release time and current time of each candidate marketing content, and get the real-time freshness score based on the release time and the current time; Based on the content intent graph, obtain the target user node vector and the candidate content node vector corresponding to each candidate marketing content, and calculate the inner product of each candidate content node vector and the target user node vector, and use the inner product as the graph relevance score; Based on the recommended fatigue feature coefficients, the feature values ​​of the content set, and the feature values ​​of the cross-sphere flow tendency, a multi-objective adjustment weight vector is constructed. The multiple graph relevance scores, multiple freshness scores, and the multi-objective adjustment weight vector are weighted and fused to obtain the comprehensive ranking score of the multiple candidate marketing content; Based on the comprehensive ranking score, the candidate marketing content is sorted in descending order, and the top N content is selected to form a marketing content set. The content in the marketing content set is then packaged into push tasks in sequence, the distribution operation is executed, and the distribution log is recorded. The distribution log is used as the intelligent distribution result.

7. A marketing content intelligent distribution system based on user behavior data, characterized in that, include: The behavior data collection module is used to acquire user behavior data on the marketing platform, including content interaction sequences and content feedback data. The intent graph and fatigue analysis module is used to acquire users' historical behavior data on the marketing platform, and to acquire a content intent graph based on the historical behavior data and the content interaction sequence, and to acquire recommendation fatigue feature coefficients based on the content intent graph. The content tag and density monitoring module is used to obtain explicit feedback data and implicit feedback data based on the content feedback data, obtain multiple content tags based on the explicit feedback data and the content intent graph, obtain the first distribution density corresponding to each content tag, and sequentially determine whether each first distribution density exceeds a preset threshold. The content concentration measurement module is used to determine that the content tag corresponding to the first distribution density has excessive concentration of marketing content distribution if the first distribution density exceeds a preset threshold, and to count all second content tags that exceed the threshold and their second distribution densities, and to obtain content concentration feature values ​​based on multiple second content tags and multiple second distribution densities. The cross-circle flow analysis module is used to obtain the duration of user ignoring distributed content and the behavior of quickly skipping content based on the implicit feedback data, and to obtain cross-circle flow tendency characteristic values ​​based on the ignoring duration, the behavior of quickly skipping content and the content interaction sequence. The intelligent distribution decision module is used to construct a marketing content set based on the recommendation fatigue feature coefficient, the content set feature value, and the cross-circle flow tendency feature value, and to intelligently distribute marketing content to users based on the marketing content set to obtain intelligent distribution results.

8. The intelligent marketing content distribution system based on user behavior data according to claim 7, characterized in that, The intent mapping and fatigue analysis module includes: The behavior sequence splicing unit is used to obtain historical content interaction sequences and historical feedback tags based on the historical behavior data, and to perform time-series alignment and splicing of the historical content interaction sequences and historical feedback tags to obtain user long and short-term behavior sequences. The intent graph construction unit is used to obtain the user-content interaction relationship in the user's long and short-term behavior sequences, construct a user-content bipartite graph based on the user-content interaction relationship, obtain user node vectors and content node vectors based on the user-content bipartite graph, and obtain a content intent graph based on the user node vectors, the content node vectors and the user-content interaction relationship. An interaction intensity calculation unit is used to obtain the frequency and average weight of user connections to similar content based on the content intent graph, and to obtain the interaction intensity saturation value based on the connection frequency and the average weight. The time decay calculation unit is used to obtain the most recent interaction timestamp of the user on similar content, and calculate the time decay factor based on the most recent interaction timestamp. The fatigue feature calculation unit is used to obtain recommended fatigue feature coefficients based on the interaction intensity saturation value and the time decay factor.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images