AI intelligent marketing content publishing subject matching recommendation method

Abstract

The application discloses an AI intelligent marketing content publishing subject matching recommendation method, relates to the technical field of marketing content publishing subject matching recommendation, and comprises the following steps: performing keyword density mutation detection, brand semantic overlap detection and content publishing cycle anomaly detection on the basis of a behavior content fusion graph, and determining whether the content publishing subject has a case that cooperation information is missing but behavior characteristics show that there is a cooperation history; based on the determination result, performing semantic co-occurrence path construction, transmission style similarity analysis and behavior residual discrimination operation, and obtaining a cooperation frequency implicit signal in the case that cooperation information is missing but behavior characteristics show that there is a cooperation history of the content publishing subject. The application solves the problem that the real cooperation frequency of the content publishing subject cannot be identified and the exposure risk cannot be accurately judged in the case that cooperation information is missing but behavior characteristics show that there is a cooperation history, and realizes implicit cooperation frequency extraction and risk level dynamic regulation based on multiple source behavior characteristics.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marketing content publisher matching recommendation, and specifically relates to an AI intelligent marketing content publisher matching recommendation method. BACKGROUND

[0002] The AI intelligent marketing content publisher matching recommendation is a technical method based on artificial intelligence technology, facing brand parties, advertisers or content marketing platforms, automatically identifying and recommending the most suitable "content publisher" (such as social media bloggers, KOLs, live streaming sellers, etc.) for their marketing goals. This method builds content publisher portraits and marketing demand portraits, uses natural language processing (NLP), machine learning, knowledge graph or graph neural network technology, and excavates the matching degree of both parties in content style, audience group, interest label, and communication influence, to achieve accurate, efficient and interpretable recommendation decisions. The current mainstream AI intelligent marketing content publisher matching recommendation system usually includes the following key links: first, the data collection and cleaning module automatically collects and organizes the behavior data and content data of content publishers on social platforms, e-commerce platforms, etc.; second, the content publisher modeling module extracts content publisher features to form multi-dimensional portraits based on text analysis, image recognition and social network relationship analysis; third, the marketing target analysis module generates a marketing demand portrait by structurally analyzing the product information, brand tone, target audience, etc. input by the advertiser; fourth, the matching algorithm engine combines collaborative filtering, deep learning or knowledge graph methods to calculate the similarity and matching degree of the two portraits; finally, the recommendation result generation and feedback optimization module recommends the optimal content publisher combination to the user and continuously optimizes the algorithm weight and feature selection according to the subsequent cooperation effect. Overall, the AI intelligent marketing content publisher matching recommendation method not only improves the efficiency and accuracy of brand selection, but also promotes the dataization and intelligentization process in the field of content marketing.

[0003] The prior art has the following disadvantages:

[0004] In the AI intelligent marketing content publishing subject matching recommendation process, the brand party usually hopes that the system can avoid recommending content publishing subjects that have been frequently cooperated and may have the risk of "excessive exposure" to ensure the freshness and user acceptance of the content. When the system performs content publishing subject screening and matching recommendation, it relies on the historical cooperation records of the content publishing subject to evaluate its cooperation frequency. However, in some actual scenarios, some content publishing subjects will actively delete or clear their past cooperation contents for the purpose of replacing platform identity, cleaning historical contents, or hiding cooperation traces, etc., which causes the system to be unable to obtain the complete historical cooperation information of the content publishing subject when performing data crawling and structured analysis. At this time, although the content publishing subject still retains obvious cooperation behavior characteristics in terms of content style, language structure, topic label, etc., since these characteristics are not modeled or recognized by the system, the existing AI intelligent marketing content publishing subject matching recommendation technology cannot accurately identify whether the content publishing subject has the risk of exposure according to the implicit signals of the cooperation frequency in the case that the cooperation information of the content publishing subject is missing but the behavior characteristics show the existence of cooperation history, thereby causing the system to misjudge the content publishing subject as a "first cooperation object" and recommend it to the brand party. This problem causes the distortion of the recommendation result, the repeated allocation of brand investment resources, the decrease of marketing conversion rate, and further affects the trust and use effect of the brand party on the recommendation system.

[0005] The above information disclosed in the background section is only intended to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide an AI intelligent marketing content publishing subject matching recommendation method to solve the problems in the background.

[0007] In order to achieve the above-mentioned purpose, the present application provides the following technical solution: an AI intelligent marketing content publishing subject matching recommendation method, specifically comprising the following steps:

[0008] S1, generating a behavior content fusion graph for detecting the cooperation information missing state by performing content behavior data fusion processing on the text content, image content, voice content, publishing time information and attention user group distribution data published by the content publishing subject;

[0009] S2, performing keyword density mutation detection, brand semantic coincidence detection and content publishing cycle anomaly detection based on the behavior content fusion graph to determine whether the content publishing subject has the condition of cooperation information missing but behavior characteristics showing the existence of cooperation history;

[0010] S3, based on the determination result, performing semantic co-occurrence path construction, propagation pattern similarity analysis and behavior residual discrimination operation, obtaining the cooperation frequency implicit signal in the case that the content publishing subject cooperation information is missing but the behavior characteristics show the existence of cooperation history;

[0011] S4, the cooperation frequency implicit signal is fused with the interactive heat change index, the topic linkage coincidence index and the content publishing time distribution index, and is input into the residual learning network for feature training, and a risk level result for identifying whether the content publishing subject exists exposure risk is output;

[0012] S5, based on the risk level result, performing content publishing subject matching priority updating, participation state classification, recommendation weight adjustment processing, and combining the risk change trend result output by the content update tracking to perform dynamic regulation operation, so as to realize continuous judgment of exposure risk and optimization of recommendation accuracy.

[0013] Preferably, S1 specifically comprises:

[0014] The text content, image content, voice content, publishing time information and attention user group distribution data published by the content publishing subject are respectively executed for deep feature extraction, the text content obtains a text feature vector through a context semantic encoding network, the image content obtains an image feature vector through a multi-layer visual perception network, the voice content obtains a voice feature vector through a time-frequency convolution encoder, the publishing time information obtains a time feature vector through a time series encoding structure, and the attention user group distribution data obtains an attention user group feature vector through a statistical normalization structure. All feature vectors are uniformly scaled and standardized to form a comparable feature set;

[0015] Based on the comparable feature set, a multi-modal sequence alignment is performed according to the time sequence of the content publishing of the content publishing subject, the text feature vector, the image feature vector, the voice feature vector, the time feature vector and the attention user group feature vector are mapped into a multi-dimensional node sequence, the semantic similarity and the time affinity between nodes are calculated through a cross-modal fusion unit and a weighted edge value is assigned, and an initial multi-modal association structure containing nodes and edges is formed;

[0016] The initial multi-modal association structure is input into a heterogeneous graph embedding network, the nodes and edges are subjected to feature interaction and spectral smoothing processing by using multi-head attention operation, and a behavior content fusion graph containing multi-dimensional behavior characteristics of text content, image content, voice content, publishing time information and attention user group distribution data is output, which is used for detecting the cooperation information missing state.

[0017] Preferably, S2 specifically comprises the following steps:

[0018] S201, extracting text content nodes in the behavior content fusion graph, constructing a time series word frequency matrix, using a sliding window algorithm to segment and statistically analyze the density change rate of keywords in continuous time periods, identifying the mutation point based on the mutation amplitude, marking the mutation node to establish the keyword density mutation detection sequence;

[0019] S202, based on the keyword density mutation detection sequence, extracting the corresponding text content nodes and image content nodes in the graph, performing semantic similarity matching with the pre-defined brand semantic vector, and extracting the keyword semantic link based on the connection relationship between nodes to construct the brand semantic overlap path graph, and realizing brand semantic overlap detection;

[0020] S203, based on the brand semantic overlap path graph, extracting the publication time information node to construct the content publication cycle sequence, using the cycle change recognition model to extract the variation point section of the publication time, and converting the keyword density mutation detection sequence, the brand semantic overlap path graph and the content publication cycle sequence into event node sets on the corresponding time axis respectively, using the event alignment algorithm to calculate the intersection number of the three types of event nodes in the sliding time window, when the intersection number exceeds the preset threshold and there is no corresponding cooperation record data in the content publication subject portrait, it is determined that the content publication subject has missing cooperation information but the behavior characteristics show that there is cooperation history.

[0021] Preferably, S201 specifically comprises:

[0022] extracting the complete keyword set contained in the text content node in the behavior content fusion graph, constructing a time series word frequency matrix segmented at equal time intervals according to the content publication time information, and each matrix unit represents the word frequency count value of the corresponding keyword in the specified time period;

[0023] Based on the time series word frequency matrix, the local interval of each keyword is processed based on the sliding window algorithm, the word frequency change rate of each keyword in adjacent time periods is calculated, and the change rate sequence is compared with the mutation amplitude threshold to identify the time period with a word frequency change rate greater than the mutation amplitude threshold as a mutation candidate segment;

[0024] Performing alignment analysis on the keyword appearance time involved in the mutation candidate segment and the content publication time information, screening out the keywords that have word frequency mutation in the content intensive publication section, marking the corresponding text content node as a mutation node, and sequentially arranging to form a keyword density mutation detection sequence.

[0025] Preferably, S202 specifically comprises:

[0026] The corresponding text content nodes and image content nodes are extracted based on keyword density mutation detection sequence, and semantic vector representations are obtained through a text semantic encoding network and an image semantic decoding network, so as to uniformly map each node representation to a shared semantic space;

[0027] In the shared semantic space, brand semantic vectors contained in a brand semantic vector library are called to perform cosine similarity calculation on each text content node and image content node, and nodes with a similarity value higher than a set semantic similarity threshold are marked as brand semantic candidate nodes;

[0028] Based on the connection relationship between nodes in the behavior content fusion graph, keyword semantic links are constructed from the brand semantic candidate nodes as starting points, semantic paths with a similarity higher than a set value and an appearance frequency meeting statistical conditions are screened between consecutive nodes, brand semantic overlap path graphs are aggregated to complete brand semantic overlap detection.

[0029] Preferably, S3 specifically comprises the following steps:

[0030] S301, a semantic connection graph structure is established based on the sequence and the brand semantic overlap path graph, continuous node tracks are generated in the graph structure according to semantic correlation strength and time proximity, and semantic co-occurrence paths are sequentially connected by calling a node set determined to have missing cooperation information but having a cooperation history displayed by behavior characteristics;

[0031] S302, text content, image content, publication time information and interactive behavior characteristics corresponding to nodes in each path are extracted based on the semantic co-occurrence paths, content style encoding, publication time interval encoding and interactive rhythm encoding are generated by using a propagation style feature extraction network, and a propagation style similarity score of each pair of adjacent nodes is calculated to complete propagation style similarity analysis;

[0032] S303, the propagation style similarity score and the semantic connection strength score of each semantic co-occurrence path are input into a residual difference discrimination structure, residual discrimination operation is performed to calculate path credibility residuals, and paths with a residual value lower than a preset discrimination threshold are selected as a high credibility path set, the appearance frequency of each node in the global semantic co-occurrence path graph is counted, and a cooperation frequency implicit signal is output.

[0033] Preferably, S302 specifically comprises:

[0034] Based on the node sequence in the semantic co-occurrence path, the text content, image content, publication time information and interactive behavior characteristics corresponding to each node are sequentially extracted, and semantic encoding operation is performed on the content data, interval discretization processing is performed on the time data, and sequence normalization operation is performed on the interactive data to generate standard input formats respectively;

[0035] The generated standard input format is input into the propagation style feature extraction network, and the contextual relationship between the text content and the image content is extracted using a multi-scale convolutional structure to form a content style code. The release time interval code is generated by modeling the release time information using a time recursive unit, and the interaction rhythm code is generated by modeling the interaction behavior changes using a frequency-aware network.

[0036] The content style encoding, publication time interval encoding, and interaction rhythm encoding of any two adjacent nodes in the same semantic co-occurrence path are fused according to a preset weighting ratio to construct a composite propagation style vector. The cosine similarity between the composite propagation style vectors is calculated to generate the corresponding propagation style similarity score. The similarity scores between consecutive adjacent nodes are combined into a propagation style similarity score sequence, which is used for subsequent behavior residual discrimination operation structure.

[0037] Preferably, S303 is as follows:

[0038] The propagation style similarity score calculated in each semantic co-occurrence path is fused with the semantic connection strength score between adjacent nodes in the path. The semantic connection strength score is calculated by weighting the cosine similarity of the semantic vectors of adjacent nodes and combining it with the content homology probability to form a comprehensive index reflecting the semantic continuity of the path. After fusion, a multi-dimensional feature input sequence is constructed to maintain the consistency of the node order.

[0039] The fused feature input sequence is input into the residual difference discrimination structure composed of stacked difference normalization units and attention adjustment units. The path residual discrimination operation is performed to extract the degree of deviation between the propagation features and the semantic structure, and the credibility residual score of each semantic co-occurrence path is output.

[0040] Semantic co-occurrence paths with a credibility residual score lower than a preset discrimination threshold are selected as a set of high-credibility paths. The cumulative frequency of each node in the set of high-credibility paths in the global semantic co-occurrence path graph is counted, and the implicit signal of cooperation frequency is output when the cooperation information of the content publishing entity is missing but the behavioral characteristics show that there is a history of cooperation.

[0041] Preferably, S4 is as follows:

[0042] The system acquires implicit signals of cooperation frequency, indicators of interaction popularity changes, indicators of topic linkage overlap, and indicators of content release time distribution. After acquisition, it performs normalization processing and time series synchronization respectively. It constructs feature alignment sequences through a sliding window mechanism, calculates fusion weights based on the correlation matrix between feature groups, and generates fusion feature vectors by superimposing them according to the weight ratio, forming training input data containing multi-source behavioral pattern features.

[0043] The training input data is input into a residual learning network composed of a residual connection structure, a feature channel compression unit and a nonlinear activation structure. The residual connection structure retains cross-layer information flow to improve convergence ability. The feature channel compression unit maintains the core feature dimension. The nonlinear activation structure performs mode enhancement processing on high-order features. A stable and separable feature expression vector is output for risk level discrimination.

[0044] The feature expression vector output by the residual learning network is input into a classification discrimination structure. A risk prediction score is generated through full connection transformation and probability mapping. The risk prediction score is a continuous value in the interval [0, 1], representing the fitting strength between the behavior characteristics of the content publishing subject and the abnormal risk. Then, the risk prediction score is matched with the preset risk division interval. The interval range is determined according to the abnormal sample distribution statistics in the model training stage. The risk prediction score below the first threshold is mapped to low risk. The risk prediction score between the first threshold and the second threshold is mapped to medium risk. The risk prediction score above the second threshold is mapped to high risk. Finally, the risk level result of whether the content publishing subject exists exposure risk is output.

[0045] Preferably, S5 specifically is:

[0046] According to the risk level result, the matching priority of the content publishing subject in the candidate set is adjusted. The content publishing subject with a low risk level result is assigned a first priority, the content publishing subject with a medium risk level result is assigned a second priority, and the content publishing subject with a high risk level result is assigned a third priority. The updated priority is written into the content publishing subject recommendation scheduling cache table in synchronization, which is used to control the recommendation trigger order in task allocation;

[0047] The current risk level result of the content publishing subject and the historical recommendation participation record are called to establish a participation state label mapping in the portrait data. The content publishing subject corresponding to the low risk is marked as a recommended participation state. The content publishing subject corresponding to the medium risk is marked as a to-be-confirmed state. The content publishing subject corresponding to the high risk is marked as a suspended participation state. The participation state is used to control the filtering logic of the recommendation trigger in scheduling;

[0048] Time sequence samples are extracted based on the content publishing subject content update log, and the risk level result after each update is associated in time sequence to construct a risk change trend sequence. The moving average method is used to calculate the short-term trend change direction and fluctuation amplitude. When the trend sequence presents a continuous downward trend and the risk level continuously maintains in the low risk interval, the content publishing subject matching priority and the recommendation weight are dynamically recovered. If the trend sequence presents an upward direction and the fluctuation amplitude exceeds the risk threshold, the current recommendation weight is reduced and the participation state is updated to suspended observation, so as to complete the dynamic regulation and control operation on the risk change trend result.

[0049] In the above technical solution, the present application provides technical effects and advantages:

[0050] 1、The present application constructs a behavior content fusion graph based on text content, image content, voice content, publication time information and attention user group distribution data, combines keyword density mutation detection, brand semantic overlap path graph generation and content publication cycle anomaly identification, and for the first time realizes that in the scene where the content publication subject actively cleans up or hides cooperation information, the system can accurately restore its potential cooperation history track through residual content publication behavior characteristics and semantic co-occurrence relationship. Further, through the construction of a semantic co-occurrence path credibility evaluation system by transmission style similarity analysis and behavior residual discrimination mechanism, the implicit signal of cooperation frequency is extracted, the problem that the content publication subject is misjudged as a "first cooperation object" due to missing cooperation data in the prior art is effectively solved, and the robustness and identification accuracy of the system in the abnormal data scene are significantly improved.

[0051] 2、On the basis of outputting the implicit signal of cooperation frequency, the present application introduces interactive heat change indicators, topic linkage overlap indicators and content publication time distribution indicators for multi-dimensional behavior feature fusion, constructs a residual learning network to complete risk level identification, and dynamically adjusts the content publication subject matching priority, participation state and recommendation weight. By introducing a risk change trend monitoring mechanism, the content publication subject state is continuously evaluated and scheduled for adaptive adjustment, avoiding the problem of excessive exposure or potential illegal delivery caused by static configuration of the recommendation path. Overall, this method not only significantly improves the recommendation accuracy and brand marketing resource allocation efficiency, but also enhances the self-correction and dynamic feedback ability of the system to data anomalies and content operation behaviors. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments or prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0053] Figure 1 The flowchart of the AI intelligent marketing content publication subject matching recommendation method of the present application. DETAILED DESCRIPTION

[0054] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the inventive aspects of example implementations to those skilled in the art.

[0055] The present application providesFigure 1 The AI intelligent marketing content publishing subject matching recommendation method shown specifically includes the following steps:

[0056] S1, through content behavior data fusion processing of the text content, image content, voice content, publishing time information published by the content publishing subject and the attention user group distribution data, a behavior content fusion graph for detecting the cooperation information missing state is generated;

[0057] In this embodiment, S1 is specifically:

[0058] The text content, image content, voice content, publishing time information published by the content publishing subject and the attention user group distribution data are respectively executed deep feature extraction, the text content obtains a text feature vector through a context semantic encoding network, the image content obtains an image feature vector through a multi-layer visual perception network, the voice content obtains a voice feature vector through a time-frequency convolution encoder, the publishing time information obtains a time feature vector through a time series encoding structure, and the attention user group distribution data obtains an attention user group feature vector through a statistical normalization structure. All feature vectors are uniformly scaled and standardized to form a comparable feature set;

[0059] When the text content, image content, voice content, publishing time information published by the content publishing subject and the attention user group distribution data are deep feature extracted, the text content can be realized through a context semantic encoding network, which uses a bidirectional Transformer structure to model the long and short sentence context relationship, capturing features such as potential brand references, marketing jargon and language style in the text. The image content can be processed by a multi-layer visual perception network, which extracts key visual elements such as scene composition, brand logo and composition style from the content publishing subject image based on the combination of convolution and attention mechanism. The voice content uses a time-frequency convolution encoder to convert the voice signal into a frequency spectrum and extract voice emotion features, speech rate features and tone style, which are used to reflect the propagation characteristics in the voice-over or live broadcast scene. The publishing time information is processed by a time series encoding structure, which converts the time point into a periodic vector representation, enabling the model to understand the distribution density and rhythm of the content publishing time. The attention user group distribution data is modeled by a statistical normalization structure, taking population geographic attributes, activity changes, historical interaction frequencies, etc. as input to generate a standardized vector. All the above vectors are standardized by a unified scale, so that the features between different modalities are comparable, thereby providing a consistent semantic space basis for subsequent fusion.

[0060] The text feature vector generated by the text content is a numerical expression for representing the natural language structure in the content of the content publishing subject, which can reveal the brand inclination, content style consistency and language tone in the semantics. The image feature vector is constructed by the convolution-attention mechanism, which contains the content style features of the image of the content publishing subject, such as whether there are brand elements in the image and whether the scene template is repeated. The speech feature vector is derived from the modeling of the audio expression behavior of the content publishing subject, which has the important ability to reveal whether the content performance is similar to the previous commercial cooperation. The time feature vector reflects the periodicity and timeliness trend of the content publishing of the content publishing subject, which is an important reference for judging the behavior rule and checking the concentrated distribution of cooperation. The user group feature vector is based on the distribution structure, interaction ability and historical change of the user group, which is used to measure the influence range and change state of the content publishing subject. The core purpose of the unified scale standardization processing is to convert the feature vectors of different sources, different structures and different dimensions into a unified feature space, so that the feature weight imbalance caused by inconsistent dimensions can be avoided during the subsequent modal fusion, thereby ensuring the expression stability and robustness of the graph construction stage.

[0061] Based on the comparable feature set, the time sequence of the content publishing of the content publishing subject is executed for multi-modal sequence alignment, and the text feature vector, image feature vector, speech feature vector, time feature vector and attention user group feature vector are mapped into a multi-dimensional node sequence. The cross-modal fusion unit calculates the semantic similarity and time affinity between nodes and assigns a weighted edge value, forming an initial multi-modal association structure containing nodes and edges.

[0062] In order to realize the multi-modal sequence alignment processing based on the comparable feature set, the content published by the content publishing subject at different time points needs to be timestamped and sorted, and an original content time sequence arranged according to the publishing time is constructed. Each content unit has been mapped to a five-dimensional feature set containing text feature vector, image feature vector, speech feature vector, time feature vector and attention user group feature vector through the previous process. Through the multi-modal sequence alignment module, multiple modalities in the same time slice are processed synchronously to ensure that each node has five unified scale representation dimensions. Then, the cross-modal fusion unit is used to calculate the semantic similarity between different content nodes. The calculation method of semantic similarity is based on the attention weighting mechanism between modal pairs, which respectively evaluates the semantic coincidence degree of text and image, text and speech, image and speech, etc. At the same time, the time affinity factor is introduced to calculate the time proximity of nodes according to the time interval function between contents. Through the double-channel fusion structure, the semantic similarity and time affinity are nonlinearly combined to generate the weighted edge value between content nodes. Finally, through the combination of node features and edge weights, an initial multi-modal association structure with heterogeneous semantic expression ability is formed, which provides complete input for subsequent graph embedding modeling.

[0063] The text feature vector is a deep expression of the semantic of the text published by the content publishing subject, which captures the dimensions of content such as keyword use, syntax structure, sentiment orientation, and reflects the language style and content strategy of the content publishing subject. The image feature vector extracts key visual clues in the content publishing subject's pictures through a convolutional attention mechanism, such as product graphics, watermark identification, scene templates, etc., revealing the continuity or variability of visual style. The speech feature vector is used to model the audio propagation characteristics in the content publishing subject's content, such as speech rhythm, emotional tone, background sound characteristics, to help identify whether it is brand cooperation voice-over content. The time feature vector reflects the periodicity of the content publishing subject's content publishing time, and can identify whether the high-frequency dense publishing behavior coincides with the business task time. The focus user group feature vector represents the user base characteristics of the content publishing subject's audience in terms of geographical distribution, interaction depth, and growth trend, and is an important parameter for identifying the migration of the focus user group and the change in the activity of the focus user group. Multi-modal sequence alignment ensures that each content node has multiple semantic dimensions on the time axis, while the cross-modal fusion unit structures the complementarity between different modalities, so that the connection between content nodes is not only based on temporal proximity, but also based on deep semantic consistency. Finally, a multi-modal node association structure with logical coherence and semantic aggregation is generated, laying the foundation for subsequent judgment of content abnormal behavior and cooperation trace residues.

[0064] The initial multi-modal association structure is input into the heterogeneous graph embedding network, and the multi-head attention operation is used to perform feature interaction and spectral smoothing processing on the nodes and edges, outputting a behavior content fusion graph containing multi-dimensional behavior features of text content, image content, speech content, publishing time information, and focus user group distribution data, which is used to detect the cooperation information missing state.

[0065] To realize the conversion of the initial multi-modal association structure to the behavior content fusion graph, first, the node set composed of the text feature vector, the image feature vector, the speech feature vector, the time feature vector, and the user group feature vector and the edge weight connection relationship thereof need to be input into the heterogeneous graph embedding network. The embedding network contains multiple types of node and multiple types of edge input channels, and introduces a multi-head attention mechanism to enhance the interactive expression ability between different modalities. Each edge aggregates the modal feature vectors of its connected nodes and combines the edge weight representation to construct the context embedding of the edge. The multi-head attention mechanism captures the relevance between modal features in parallel during the node update process, thereby improving the stability and accuracy of multi-modal information fusion. In the feature update process, spectral smoothing processing is introduced, that is, the features of each node are locally aggregated by the graph convolution framework, so that the node expressions within the same subgraph are more consistent in the embedding space, avoiding interference of local outliers on global embedding. In the finally output graph structure, each node contains a fused high-dimensional behavior feature representation, and is interconnected with time, semantic, and audience features to form a behavior content fusion graph for detecting the cooperation information missing state of the content publishing subject.

[0066] The text content node in the multi-modal association structure carries the propagation features of the content publishing subject at the semantic level, and the stability and deviation trend of the language style can be effectively encoded through graph embedding. The image content node records the visual expression habit and can capture the reproduction of a specific image template or visual theme in multiple publishing cycles. The speech content node learns whether the voice broadcast style converges through the embedding network, thereby revealing potential advertising repetition behavior. The publishing time information node expresses the concentration and periodicity of the content in the time distribution dimension, which helps to infer the hidden time clustering of the cooperation task. The user group distribution data node provides user portrait support, and the embedded features reflect the consistency of audience migration, interaction fluctuation, and people portrait. Under the action of the multi-head attention mechanism, various nodes interact features through weighted aggregation, effectively eliminating the information island problem caused by a single modality. Spectral smoothing processing further ensures the structural stability of the embedding space, making the behavior features among semantics, time, and audience show strong continuity and interpretability. The finally constructed behavior content fusion graph not only retains the original multi-modal feature structure, but also strengthens the aggregation and expression of potential cooperation behavior features, providing a high credibility input basis for subsequent cooperation information missing state detection.

[0067] S2, performing keyword density mutation detection, brand semantic overlap detection, and content publishing cycle anomaly detection based on the behavior content fusion graph to determine whether the content publishing subject has a cooperation information missing but behavior feature showing a cooperation history;

[0068] In this embodiment, S2 specifically includes the following steps:

[0069] S201, extract text content nodes in the behavior content fusion graph, construct a time series word frequency matrix, use a sliding window algorithm to segment and statistically analyze the density change rate of keywords in continuous time periods, identify the density mutation point based on the mutation amplitude, mark the mutation node to establish a keyword density mutation detection sequence;

[0070] S202, based on the keyword density mutation detection sequence, extract the corresponding text content nodes and image content nodes in the graph, match the semantic similarity with the pre-defined brand semantic vector, extract the keyword semantic link by combining the connection relationship between nodes, construct the brand semantic overlap path graph, and realize brand semantic overlap detection;

[0071] S203, based on the brand semantic overlap path graph, extract the publishing time information node to construct the content publishing cycle sequence, use the cycle change recognition model to extract the variation point section of the publishing time, and convert the keyword density mutation detection sequence, brand semantic overlap path graph and content publishing cycle sequence into event node sets on the corresponding time axis respectively, use the event alignment algorithm to calculate the intersection number of the three types of event nodes in the sliding time window, when the intersection number exceeds the preset threshold and there is no corresponding cooperation record data in the content publishing subject portrait, it is determined that the content publishing subject has cooperation information missing but the behavior characteristics show that there is cooperation history.

[0072] In order to realize the judgment of whether the content publishing subject has cooperation information missing but the behavior characteristics show that there is cooperation history, first, the content publishing time information nodes related to keyword mutation are located based on the brand semantic overlap path graph, and the content publishing cycle sequence is constructed accordingly. This sequence records the intensity of content publishing of the content publishing subject in each time period, and extracts abnormal variation point sections through the cycle change recognition model, such as time intervals with significantly increased content publishing frequency or periodic abnormal fluctuations. Then, the mutation time points in the keyword density mutation detection sequence, the path appearance time nodes in the brand semantic overlap path graph, and the variation time nodes in the content publishing cycle sequence are respectively mapped into event node sets on the standard time axis. The event alignment algorithm is used to calculate the intersection number of the three types of node sets in the sliding time window. The algorithm aligns the three event streams in turn through sliding window scanning, and counts the intersection number of at least two or three events appearing at the same time in the window range. If the intersection number exceeds the preset threshold, and there is no explicit cooperation record data in the content publishing subject portrait, it can be judged that the content publishing subject's behavior presents cooperation signals but the cooperation information is missing. The preset threshold is the minimum intersection number threshold set by experience, which is used to avoid false judgment caused by accidental overlap, for example, set to at least twice in a three-day sliding window. This strategy can identify hidden cooperation behavior through structured cross-validation between multiple source behavior signals, effectively compensating for missing structured cooperation records.

[0073] In this embodiment, S201 is specifically:

[0074] In the behavior content fusion graph, the complete keyword set contained in the text content node is extracted, and a time series word frequency matrix segmented at equal time intervals is constructed according to the content publishing time information, and each matrix unit represents the word frequency count value of the corresponding keyword in the specified time period;

[0075] In the behavior content fusion graph, the complete keyword set contained in the text content node is extracted, that is, all keywords after natural language processing segmentation in the node are uniformly summarized to form a keyword full set for frequency analysis. In order to facilitate trend analysis in the time dimension, it is necessary to construct a time series structure segmented at equal time intervals in combination with the content publishing time information, and each segment represents a fixed time interval, for example, segmented by day or by hour. In each time period, the number of occurrences of each keyword in the keyword full set is counted to construct a time series word frequency matrix. Each row of the matrix corresponds to a keyword, each column corresponds to a time period, and the value of each cell represents the frequency of the keyword in the corresponding time period. This structure can be analyzed by a sliding window mechanism to capture the trend of keyword usage over time. The core purpose of this method is to establish a semantic intensity graph of content evolution over time, and to reveal possible abnormal semantic focusing behavior through time series trend, providing basic data support for subsequent identification of implicit signals of cooperative behavior. The time series word frequency matrix, as high-dimensional structure data, retains the dual semantic dimensions of semantic content and publishing time, laying a precise information foundation for keyword density mutation analysis and cycle synchronization detection.

[0076] On the basis of the time series word frequency matrix, the local interval of each keyword is processed based on the sliding window algorithm, the word frequency change rate of each keyword in adjacent time periods is calculated, and the change rate sequence is compared with the mutation amplitude threshold to identify the time period with a word frequency change rate greater than the mutation amplitude threshold as a mutation candidate segment;

[0077] On the basis of the time series word frequency matrix, a sliding window algorithm is performed on each keyword for local interval traversal processing, which means that fixed-length continuous time periods are taken in the time series in turn to form a sliding window, and the keyword frequency statistics value of adjacent time periods is extracted in each window. The word frequency change rate is calculated by the difference method. The word frequency change rate represents the relative growth or decline amplitude of the current time period compared with the previous time period, reflecting the use intensity fluctuation of the keyword in the short term. In order to identify the sudden content tendency, the word frequency change rate sequence of each keyword needs to be compared with the preset mutation amplitude threshold. The mutation amplitude threshold can be set by historical statistical analysis or dynamic adjustment strategy, which represents that the system considers that the growth rate of a certain keyword exceeds a certain proportion in a unit of time, which is abnormal change. For example, if the word frequency of a keyword increases from 10 times to 80 times in a day, the growth rate is 700%, which is much higher than the set mutation threshold of 200%, so the current time period is marked as a mutation candidate segment. This process can be accelerated by batch matrix operation to improve processing efficiency. The core significance of this detection method is to find the high-frequency short-period concentrated use behavior in the content. Such mutations are usually closely related to the concentrated exposure of commercial cooperation information, and are therefore one of the important clues to judge the behavior pattern change of the content publishing subject. The sliding window strategy ensures the fine-grained detection of word frequency change, and the mutation amplitude threshold provides a controllable sensitivity adjustment method, making the detection result more robust and adaptive.

[0078] The alignment analysis is performed on the keyword appearance time and content publishing time information involved in the mutation candidate segment, the keywords with word frequency mutation in the content intensive publishing section are screened out, the corresponding text content nodes are marked as mutation nodes, and the keyword density mutation detection sequence is formed in turn.

[0079] The alignment analysis is performed on the keyword appearance time and content publishing time information involved in the mutation candidate segment, specifically, the time point at which each keyword determined to have a keyword frequency mutation first or frequently appears in the text is time-stamped matched with the actual publishing time of the content corresponding to the content publishing subject, and a keyword-published time alignment sequence is constructed. In the sequence, further identify which keywords have mutation appearance time concentrated in the time period of continuous content publishing, that is, identify the "content intensive publishing section". The content intensive publishing section can be determined by counting the number of contents in the continuous time period exceeding the average publishing density in the sliding window, which represents the content publishing subject significantly increasing the frequency of publishing content in a specific period, usually corresponding to the behavior pattern of concentrated expression of a certain theme. Once the mutation time of a keyword coincides with such a high-density publishing section, it is determined to be a high-correlation behavior keyword, and the text content node corresponding to the keyword is marked as a mutation node. All marked mutation nodes are arranged in chronological order to form a keyword density mutation detection sequence. In this way, invalid mutations caused by accidental language fluctuations or non-cooperative backgrounds can be eliminated, and keywords with double aggregation in time and content structure are retained, so that the potential cooperative behavior characteristics can be more accurately identified. The process relies on the alignment mechanism of time semantics, local behavior density analysis and node time sequence modeling method, and can improve the detection ability of content manipulation or concentrated feature exposure.

[0080] In the embodiment, S202 is specifically:

[0081] Based on the keyword density mutation detection sequence, the corresponding text content nodes and image content nodes are extracted, and semantic vector representations are obtained through a text semantic encoding network and an image semantic decoding network, respectively, to uniformly map each node representation to a shared semantic space;

[0082] Based on the keyword density mutation detection sequence, the corresponding text content nodes and image content nodes are extracted, which refers to screening the nodes containing text information and the nodes containing image information from the content nodes associated with the identified mutation keywords as the target objects of semantic analysis. The text content nodes are processed by a text semantic encoding network, which usually adopts a bidirectional language model or a context understanding structure based on an attention mechanism to realize deep semantic encoding of the context of the sentence where the keyword is located, and output text semantic vectors that can be used for similarity calculation. The image content nodes are processed by an image semantic decoding network, which is usually composed of a visual perception network and a semantic projection layer, extracts multi-level visual features of the image, maps these visual features with label information in the semantic concept space, and finally obtains image semantic vectors. After the two categories of semantic vectors are processed by unified feature normalization, they are mapped into a shared semantic space, thereby realizing alignment and semantic consistency modeling between different modalities. The core of this operation is cross-modal representation alignment, so that text and image information can be uniformly measured in the same semantic dimension, thereby providing a unified calculation basis for subsequent similarity matching with brand semantic vectors. This way can significantly improve the collaborative expression ability of text and image content in brand relevance determination, and enhance the ability to identify potential semantic association features in complex content.

[0083] In the shared semantic space, the brand semantic vectors contained in the brand semantic vector library are called to perform cosine similarity calculation on each text content node and image content node respectively, and the nodes with similarity values exceeding the set semantic similarity threshold are marked as brand semantic candidate nodes.

[0084] In the shared semantic space, calling the brand semantic vector contained in the brand semantic vector library refers to using a pre-constructed brand semantic representation set, which is generated by a deep language model and a visual language alignment network based on brand historical content, advertising language, visual features, and other multi-dimensional inputs, forming a brand semantic vector representation that can be used for comparison. Each text content node and image content node has been embedded in the shared semantic space and is represented as a semantic vector with uniform dimensions and scales. By performing cosine similarity calculations, the directional consistency of these content nodes and each brand semantic vector in the semantic space can be measured, and their semantic relevance can be evaluated. The closer the cosine similarity value is to 1, the higher the semantic similarity between the two vectors. To ensure that the matching results have discriminative ability, a semantic similarity threshold needs to be set, which is obtained by cross-validation strategy on the training data and represents the minimum acceptable standard for semantic relevance. For example, if the threshold is set to 0.85, only when the cosine similarity between the text or image semantic vector and a certain brand semantic vector exceeds 0.85, can the node be marked as a brand semantic candidate node. This judgment standard helps to eliminate non-target brand content interference and retain content clues with potential brand association, thereby improving the accuracy of subsequent brand semantic overlap detection. This method ensures that the extracted candidate nodes have high credibility and business relevance through high-dimensional semantic alignment and strict similarity threshold screening.

[0085] Starting from the brand semantic candidate node, key word semantic links are constructed based on the connection relationships between nodes in the behavior content fusion graph. Semantic paths with similarity higher than a set value and appearance frequency meeting statistical conditions between consecutive nodes are screened, and brand semantic overlap path graphs are aggregated to complete brand semantic overlap detection.

[0086] With the brand semantic candidate node as the starting point, the keyword semantic link is constructed, which means that in the behavior content fusion graph, starting from the text content node or image content node that has been marked as a brand semantic candidate node, the existing connection relationship between nodes in the graph is used to traverse adjacent nodes layer by layer, and a path with semantic continuity is found. In the path screening process, the cosine similarity of the semantic vectors between consecutive nodes is calculated, and the node pairs with similarity higher than the set value are selected for path splicing, where the set value is usually determined according to experimental verification, and the minimum similarity threshold that can guarantee semantic coherence is usually selected, for example, 0.8 or higher. In order to avoid the interference of low-frequency accidental paths on the results, it is also necessary to set statistical conditions, that is, within a certain time window, the number of times the path appears reaches a certain frequency threshold, for example, more than 3 times, and it is considered as a stable and representative path. On this basis, the paths that meet the conditions are aggregated to generate a brand semantic overlap path graph, which is used to represent the potential semantic penetration trajectory of a brand in the content publishing subject content. This path graph can present multiple semantic similar content propagation chains, which facilitates the identification of whether brand-related content forms a high-frequency coverage phenomenon in the content publishing subject historical content. By combining the semantic consistency between nodes and the path appearance frequency for double constraints, the accuracy and stability of the detection results are effectively improved, providing structured basis for subsequent judgment of whether there is a cooperation information missing.

[0087] S3, based on the determination result, performing semantic co-occurrence path construction, propagation style similarity analysis and behavior residual discrimination operation, and obtaining cooperation frequency implicit signal in the case that the content publishing subject has cooperation information missing but the behavior characteristics show cooperation history;

[0088] In this embodiment, S3 specifically includes the following steps:

[0089] S301, by calling the node set that has been determined to have cooperation information missing but the behavior characteristics show cooperation history, a semantic connection graph structure is established based on the keyword density mutation detection sequence and the brand semantic overlap path graph, continuous node trajectories are generated in the graph structure according to semantic association strength and time proximity, and semantic co-occurrence paths are formed by connecting in turn;

[0090] The construction of the semantic connection graph structure first takes the node set with confirmed missing cooperation information but behavior characteristics showing cooperation history as the core input. The node types include text content nodes and image content nodes. These nodes are derived from the keyword density mutation detection sequence and the brand semantic coincidence path graph, and have high time sequence density and brand semantic coincidence degree. In order to construct the semantic connection graph, the semantic association strength between any two nodes in the node set needs to be calculated, which is usually measured by the cosine similarity between semantic vectors, and the content homology probability is added to correct the semantic similarity. After completing the semantic strength calculation, the time span threshold is set according to the time proximity, and the node pairs that meet the requirements of time continuity and semantic strength are selected as connectable edge pairs. Finally, the nodes are taken as the vertices in the graph structure, and the edge pairs that meet the conditions are connected to form a semantic connection graph with complete structure, smooth semantics and time sequence continuity. Each continuous connection path in the graph is a potential semantic co-occurrence path, which is used for subsequent propagation style analysis and frequency evaluation.

[0091] The keyword density mutation detection sequence is used to identify nodes with significant changes in keyword usage density within a short period of time, which represent key points of rapid content direction changes. The brand semantic coincidence path graph represents content flow paths with high semantic coincidence with existing brand semantic vectors, and its node set has high semantic concentration. The "semantic association strength" in the semantic connection graph structure is a quantitative indicator reflecting the intrinsic connection between nodes, usually calculated by the cosine similarity of the vector output by the semantic encoding model, and the consistency or style similarity of the content source is weighted to generate the final association score. "Time proximity" reflects the distance between nodes on the time axis, which is used to avoid connecting node pairs with large time span in the graph structure to ensure the reasonable content publishing continuity of the path. The path generation logic uses a sliding window strategy to traverse the continuous node sequence that meets the semantic strength and time proximity in the semantic connection graph, and records them as semantic co-occurrence paths in sequence. These paths are a deep abstraction of the historical behavior patterns of content publishing subjects, and reflect the hidden association structure in their potential cooperation trajectories.

[0092] S302, based on the semantic co-occurrence path, extracts the text content, image content, publishing time information and interaction behavior characteristics of each node in the path, generates content style encoding, publishing time interval encoding and interaction rhythm encoding using the propagation style feature extraction network, and calculates the propagation style similarity score of each adjacent node pair to complete the propagation style similarity analysis;

[0093] S303, input the propagation style similarity score and the semantic connection strength score of each semantic co-occurrence path into the residual difference discrimination structure, perform residual discrimination operation to calculate the path credibility residual, and select the path with a residual value lower than the preset discrimination threshold as the high credibility path set, count the appearance frequency of each node in the global semantic co-occurrence path graph in the high credibility path set, and output the cooperation frequency implicit signal.

[0094] The core purpose of the processing process is to rely on the fusion modeling of multi-modal behavior characteristics and semantic structure characteristics to mine high potential cooperation paths with stable propagation behavior characteristics in the case of missing content publishing subject cooperation information, so as to identify the cooperation frequency implicit signal. Because part of the content publishing subjects may consciously delete historical cooperation traces, making it difficult for the system to judge their cooperation frequency through explicit data, it is necessary to rely on the content style, release time regularity, interaction behavior and other propagation styles, and the internal continuity between text and image semantics for indirect modeling. By performing residual discrimination on the propagation style similarity score and the semantic connection strength score, it can be quantified whether the semantic cooperation between the contents is consistent with the actual propagation mode. If the two are highly matched, it means that the path has strong endogenous cooperation behavior. Further, by screening the paths with low residuals, accidental similar contents are avoided, and the nodes with structural repetition characteristics are extracted from the global graph, and the frequency thereof can be used as an important indicator of the implicit cooperation strength of the content publishing subject. This way not only avoids over-reliance on a single content, but also restores the real cooperation tendency of the content publishing subject from the structural repetition behavior, realizes high robustness identification of the cooperation behavior under the content cover, and provides more in-depth basic data for risk assessment and recommendation strategy.

[0095] In this embodiment, S302 is specifically:

[0096] Based on the node sequence in the semantic co-occurrence path, the text content, image content, release time information and interaction behavior characteristics corresponding to each node are extracted in turn, and the semantic encoding operation is performed on the content data, the interval discretization processing is performed on the time data, and the sequence normalization operation is performed on the interaction data, to generate standard input formats respectively;

[0097] In the node sequence processing process in the semantic co-occurrence path, the associated text content, image content, publication time information and interactive behavior features of each node need to be extracted and multi-dimensional data preprocessing is performed to form a unified format data input. The text content is processed by a pre-trained semantic encoding network, which often uses a context-aware language model to extract syntactic structure and semantic relationship information and outputs a semantic vector representation. The image content uses a convolutional neural network combined with a visual attention mechanism to extract theme elements and style elements in the image, ensuring that the text and image content have comparability in the semantic space. For publication time information, the time interval between adjacent nodes is calculated according to the chronological relationship of the nodes on the time axis, and a time discretization encoding method is used to convert the interval into a fixed-dimensional time sequence vector to reflect the periodicity and abruptness of the publication time. The interactive behavior features include comment quantity, like quantity and forwarding frequency, etc. First, time normalization and interval classification processing are performed to compress the numerical values into a relatively stable interval range, and then a normalized sequence vector is generated according to the interactive frequency trend. All features are standardized to a fixed dimension after generation to ensure that the subsequent propagation pattern encoding network can stably receive and integrate multi-modal information, and also facilitate the comparison and calculation of nodes within the path in different propagation feature dimensions. This processing method realizes the structural unification and alignment of cross-modal data, providing a structured input basis for path similarity and residual calculation.

[0098] The generated standard input format is input into the propagation pattern feature extraction network, which uses a multi-scale convolutional structure to extract the context association in the text content and image content to form content style encoding, uses a time recurrent unit to model the publication time information to generate publication time interval encoding, and uses a frequency-aware network to model the interactive behavior changes to generate interactive rhythm encoding.

[0099] The generated standard input format is input into the propagation style feature extraction network, and the text content, image content, publication time information and interactive behavior features need to be modeled and deeply coded in different channels. For text content and image content, a multi-scale convolution structure is constructed to extract context semantics and visual features under different receptive fields. The text path combines residual connections through a multi-core convolution network to obtain local word combination patterns and long-distance semantic dependencies. The image path extracts local texture, edge contour and image composition style through a multi-level convolution network. Finally, the extracted results are fused into content style encoding to represent the style coherence and visual language features of the content. For publication time information, a time recursion unit is designed to model the time interval sequence. The gating cycle mechanism is used to learn the progressive relationship and periodic pattern between time intervals, thereby outputting a publication time interval encoding that can reflect the rhythm of the time distribution. For interactive behavior features, a frequency perception network is used for modeling. This network combines spectrum decomposition and local trend detection structure to capture the rhythm changes of interactive indicators in continuous content, and then generates an interactive rhythm encoding. The above three types of encoding are unified into vector format and used as a multi-dimensional representation of path propagation style features, providing stable and structured feature support for subsequent propagation style similarity calculation. This multi-network collaborative modeling approach helps to restore the style consistency and publication behavior patterns behind content propagation, playing a key role in identifying implicit signals of content publishing subject cooperation frequency.

[0100] The content style encoding, publication time interval encoding and interactive rhythm encoding of any two adjacent nodes in the same semantic co-occurrence path are fused according to a preset weighting ratio to construct a composite propagation style vector. The cosine similarity between the composite propagation style vectors is calculated to generate the corresponding propagation style similarity score. The similarity scores between consecutive adjacent nodes are combined into a propagation style similarity score sequence, which is used in the subsequent behavior residual discrimination operation structure.

[0101] The propagation style similarity calculation of any two adjacent nodes in the same semantic co-occurrence path needs to first extract the content style code, the publishing time interval code and the interactive rhythm code corresponding to each node, and perform feature fusion operation according to the preset weighting proportion to construct a unified dimension composite propagation style vector. The weighting proportion can be optimized and determined based on the minimum prediction error of the propagation similarity of the verification set in the training stage to ensure that each type of feature reflects its actual influence in the similarity calculation. Subsequently, the cosine similarity operation is performed on the composite propagation style vector of any pair of adjacent nodes in the path to measure the similarity degree of the propagation style. The similarity value ranges from-1 to 1, and the higher the value, the more consistent the propagation mode. For example, if the content style of two nodes is concise text and image collocation, the publishing time interval is similar, and the interactive heat change trend is consistent, the corresponding propagation style similarity score is close to 1. The similarity scores of all adjacent node pairs are sequentially composed into a propagation style similarity score sequence, which is used as the continuity description of the path in the propagation dimension for further behavior residual discriminant analysis. In this process, ensuring the consistency of feature fusion, the uniformity of vector dimension and the differentiability of calculation method can guarantee the stability and generalization ability of the similarity sequence, and lay a foundation for subsequent identification of propagation chain abnormalities or implicit repeated behaviors.

[0102] In the embodiment, S303 is specifically:

[0103] The calculated propagation style similarity score in each semantic co-occurrence path is fused with the semantic connection strength score between the adjacent nodes in the path. The semantic connection strength score is obtained by calculating the cosine similarity of the semantic vectors of adjacent nodes and combining the content homology probability to form a comprehensive index reflecting the semantic continuity of the path. After fusion, a multi-dimensional feature input sequence is constructed, and the node order is kept consistent;

[0104] When performing multi-dimensional feature fusion processing on each semantic co-occurrence path, the sequence of the calculated propagation pattern similarity scores in the path is first extracted and paired with the semantic connection strength scores between each pair of adjacent nodes in the path. The calculation method of the semantic connection strength score includes performing cosine similarity calculation on the semantic vectors of adjacent nodes to measure the consistency of their semantic expressions, and combining the content homology probability, i.e., judging whether the contents of two nodes come from similar topics or backgrounds, such as brand keyword overlap, image subject consistency, text rhetoric repetition, etc. The content homology probability score is output by a pre-trained model, and is weighted with the cosine similarity according to a set proportion to obtain a comprehensive semantic connection strength score. Taking a text node as an example, if the semantic vector similarity of two nodes is 0.85 and the content homology probability is 0.9, and the weighting coefficients are 0.6 and 0.4 respectively, then the connection strength score will reflect that it has significant continuity in semantics. Then, the propagation pattern similarity score and the semantic connection strength score are aligned and fused to form a joint feature representation containing the propagation dimension and the semantic dimension. Each group of scores corresponds to a node connection segment in the path, and the joint features of all connection segments are arranged in sequence according to the nodes to construct a multi-dimensional feature input sequence, which provides continuity and consistency information input for the subsequent residual discriminant structure, thereby ensuring fine modeling of the path credibility. This method can accurately capture special paths with similar propagation but broken semantics or consistent semantics but abnormal propagation behavior, thereby improving the recognition accuracy of implicit high-frequency cooperation signals.

[0105] The fused feature input sequence is input into a residual difference discriminant structure composed of stacked difference normalization units and attention adjustment units to perform path residual discriminant operation, extract the deviation degree between the propagation features and the semantic structure, and output the credibility residual score of each semantic co-occurrence path;

[0106] To calculate the degree of deviation between the propagated features and the semantic structure in each semantic co-occurrence path, the fused feature input sequence needs to be input into a residual difference discrimination structure composed of stacked difference normalization units and attention adjustment units. The implementation of this structure is as follows: first, the stacked difference normalization unit is used to perform a difference operation on the continuous node connection segment in the multi-dimensional feature sequence, extract the local change rate information inside the feature sequence, such as the numerical jump area between the propagation pattern similarity and the semantic connection strength, and thus obtain the fluctuation trend feature inside the path. Subsequently, the normalized processing is performed on the difference feature sequence, and all values are mapped to a unified interval range, which is convenient for subsequent network unified perception and weight update, and prevents the scale inconsistency from affecting the stability of the model. Then, the attention adjustment unit performs weighted analysis on the normalized feature sequence, automatically identifies the local segment that has a key influence on the path credibility judgment, and improves the attention intensity of the model to important node relationships. For example, in a path, if two nodes show high consistency in semantics, but the propagation pattern difference is great, the attention mechanism will allocate a higher weight to amplify this contradiction, thereby significantly affecting the output of the residual score. Finally, the residual difference discrimination structure outputs the credibility residual score of each path through the end-to-end feedforward calculation process, reflecting the coupling degree of the path between semantic continuity and propagation consistency. The lower the credibility residual score, the higher the consistency of the path in the two dimensions of propagation and semantics, and the more likely it is to constitute a real implicit high-frequency cooperation path, thereby providing a quantitative basis for subsequent high-credibility path screening. This structure integrates difference trend modeling and attention selection mechanism, not only enhancing the perception ability of feature anomalies, but also effectively improving the recognition accuracy of behavior disguising and information masking.

[0107] The semantic co-occurrence paths with credibility residual scores lower than the preset discrimination threshold are screened as the high-credibility path set, the cumulative appearance frequency of each node in the global semantic co-occurrence path graph in the high-credibility path set is counted, and the cooperation frequency implicit signal in the case where the content publishing subject cooperation information is missing but the behavior feature shows a cooperation history is output.

[0108] To identify the cooperation frequency implicit signal from the semantic co-occurrence path graph, the credibility residual scores of all semantic co-occurrence paths need to be screened, and a high-credibility path set is extracted according to a preset discrimination threshold. First, the credibility residual score is quantified by the residual difference discrimination structure to the deviation degree between the semantic continuity and the propagation pattern consistency of each path. The lower the score, the stronger the coupling relationship between the two feature dimensions. The preset discrimination threshold can be set based on the residual distribution of the training set, for example, selecting the lower quantile threshold of the statistical distribution as the screening boundary of the credible path, so as to eliminate low-quality paths with high propagation structure disturbance or semantic instability. Then, all paths with scores lower than the discrimination threshold are collected as a high-credibility path set, representing potential real content coordination propagation trajectories. Next, in the global semantic co-occurrence path graph, cumulative statistical operations are performed on all nodes in the high-credibility path set to calculate the total frequency of each node appearing in the set. The higher the frequency of a node, the more it appears in multiple credible paths, indicating a higher potential active cooperation. By quantitatively modeling these frequency results, the cooperation frequency implicit signal can be output in the case where the content publishing subject's cooperation information is missing but the behavior characteristics show a cooperation history, reflecting the potential cooperation strength of the content publishing subject under hidden cooperation records, providing key support for subsequent risk identification and recommendation priority decision-making. This process realizes the modeling path of returning structural signals to the global behavior graph from the residual discrimination results, ensuring the semantic closure and propagation credibility of the frequency signal.

[0109] S4, fuse the cooperation frequency implicit signal, the interaction heat change index, the topic linkage coincidence index and the content publishing time distribution index, input the residual learning network for feature training, and output a risk level result for identifying whether the content publishing subject has exposure risk;

[0110] In this embodiment, S4 is specifically:

[0111] The cooperation frequency implicit signal, the interaction heat change index, the topic linkage coincidence index and the content publishing time distribution index are obtained, and after being obtained, they are respectively subjected to normalization processing and time sequence synchronization, a feature alignment sequence is constructed through a sliding window mechanism, a fusion weight is calculated based on the correlation matrix between feature groups, a fusion feature vector is generated by weight proportion superposition, and training input data containing multi-source behavior mode features are formed;

[0112] To construct the training input data containing multi-source behavior pattern features, first, the cooperation frequency implicit signal, the interaction heat change index, the topic linkage coincidence index and the content publishing time distribution index are obtained respectively through the data pipeline. These data are expressed in the form of time series to represent the content publishing subject behavior dynamics. After obtaining, the maximum and minimum value normalization algorithm is applied to each type of index to make the data values uniform in the standard numerical interval, avoiding the deviation of different dimension data to subsequent calculation. At the same time, time step alignment operation is performed on all time series to ensure the correspondence between each type of index at the same time point. Then, the sliding window mechanism is adopted to set a fixed time span window to slide and intercept the index sequence, thereby generating multiple groups of time segment samples. In each group of samples, based on the time dimension features between different indexes, the Pearson correlation coefficient matrix between feature groups is calculated to represent the correlation strength between each type of behavior signal. The fusion weight is obtained by normalizing the correlation matrix, and each dimension feature signal is weighted and combined according to the weight to generate a fusion feature vector reflecting the comprehensive features of the content publishing subject behavior in the time segment, which is used as the input data for subsequent model training.

[0113] The cooperation frequency implicit signal represents the cumulative appearance frequency of the content node of the content publishing subject in the high trust path graph under the premise of lacking explicit cooperation marks, reflecting the potential commercial cooperation evidence between contents; the interaction heat change index represents the fluctuation trend of interaction behaviors such as the number of user comments, likes and forwards in the time dimension, capturing the content heat change rule; the topic linkage coincidence index is based on the semantic coincidence of the topic labels appearing in the content of the content publishing subject and the brand-related topics in the same period, reflecting the potential topic co-occurrence relationship; the content publishing time distribution index quantifies the intensity of content publishing by the content publishing subject in each period. Normalization makes each type of data comparable, time series synchronization ensures consistent sample data dimensions, sliding window construction enhances the model's ability to perceive local behavior patterns, and the correlation matrix provides an important weighting basis for data fusion. The final fusion feature vector serves as a comprehensive behavior expression and can be used as training input for subsequent risk identification tasks. Each technical point in this process is supported by specific data structures and algorithms to ensure the feasibility and effectiveness of the processing logic.

[0114] The training input data is input into a residual learning network composed of a residual connection structure, a feature channel compression unit and a nonlinear activation structure. The residual connection structure retains cross-layer information flow to improve convergence ability, the feature channel compression unit maintains the core feature dimension, and the nonlinear activation structure performs pattern enhancement processing on high-order features. The output is a stable and separable feature expression vector for risk level discrimination;

[0115] After processing the constructed training input data, it is input into a residual learning network composed of a residual connection structure, a feature channel compression unit and a nonlinear activation structure in series, aiming to extract discriminative feature expressions from multi-source behavior patterns. The residual connection structure introduces a direct cross-layer data path between the layers of the backbone network, allowing the original input features to be weighted and fused with the output of the intermediate layers, thereby alleviating the gradient vanishing problem in deep networks and accelerating the convergence speed of the model. After the output of the residual path, the feature channel compression unit uses a one-dimensional convolution kernel or a learnable linear transformation matrix to reduce the dimension of the input feature channel, ensuring that the core feature dimensions with the most discriminative ability are retained while reducing the feature dimension. The compressed feature vector is sent to the nonlinear activation structure, which is commonly used structures such as ReLU or Swish activation functions, which serve to map different feature dimensions non-linearly, enhancing the network's ability to model complex patterns. Finally, the network outputs a stable and separable feature expression vector after multiple layers of abstraction, which can be used as the core basis for the subsequent risk level discrimination module.

[0116] The residual connection structure is a typical deep neural network optimization method that directly transmits the input signal of the previous layer to the subsequent layer for accumulation, avoiding the training degradation problem caused by the increase in network depth, and is particularly suitable for modeling deep dependencies between long sequences in behavior data. The design of the feature channel compression unit is based on the dual consideration of parameter efficiency and model generalization ability, which improves the convergence efficiency of the model during training and reduces the risk of overfitting by compressing redundant channels into compact vector representations. The introduction of the nonlinear activation structure breaks the bottleneck of linear model expression ability, allowing the feature vector to express nonlinear pattern relationships, especially for hidden nonlinear propagation characteristics in content publishing subject behavior data. The three modules cooperatively form the core computational graph of the residual learning network, enabling it to quickly focus on high-order patterns between behavior anomaly signals and cooperation frequencies during training, improving the stability and accuracy of subsequent risk level discrimination tasks.

[0117] The feature expression vector output by the residual learning network is input into the classification discrimination structure, which generates a risk prediction score through fully connected transformation and probability mapping. The risk prediction score is a continuous value in the interval [0, 1], representing the fitting strength between the behavior characteristics of the content publishing subject and the abnormal risk. Then, the risk prediction score is matched with the preset risk division interval, which is determined by the model training phase based on the distribution statistics of abnormal samples. Risk prediction scores below the first threshold are mapped to low risk, risk prediction scores between the first threshold and the second threshold are mapped to medium risk, and risk prediction scores above the second threshold are mapped to high risk. Finally, the risk level result of whether the content publishing subject exists exposure risk is output.

[0118] After inputting the feature expression vector output by the residual learning network into the classification discrimination structure, a linear transformation is performed on the high-dimensional behavior feature vector through one or more fully connected networks to obtain a scalar form of risk prediction score. To make the score interpretable, a Sigmoid function or a Softmax function is often used for probability mapping to normalize the output value to the [0, 1] interval, so that the value can be regarded as the fitting probability of the existence of potential exposure risk of the content publishing subject. For example, when the score is 0.8, it means that the model considers that the content publishing subject has strong behavior risk characteristics. This score is not directly used to output the risk level, but is used as the input basis for subsequent risk level division. Next, by matching the risk prediction score with the risk division interval set in the training stage, a discrete risk level label is correspondingly mapped. The interval matching process completes the classification output of the risk degree by comparing the relationship between the risk score and the first threshold and the second threshold defined in the training.

[0119] The risk division interval is composed of three parts: the low-risk interval is [0, the first threshold], the medium-risk interval is (the first threshold, the second threshold], and the high-risk interval is (the second threshold, 1]. The core basis for the division of the three is the density estimation and discrimination accuracy analysis of the abnormal sample distribution in the model training process. The first threshold is usually set to the quantile that can minimize the false positive rate, such as 0.35; the second threshold can be selected as the value corresponding to the cumulative probability of abnormal samples close to 70%, such as 0.65, so that a large number of content publishing subjects with medium behavior characteristics are classified as medium risk, and only those with highly fitted abnormal patterns are classified as high risk. Through this way, automatic stratification of risk levels can be achieved, which not only enhances the recognition robustness of the system, but also facilitates the platform to take differentiated intervention strategies in subsequent operation decisions, such as suspending cooperation, enhancing review, or increasing attention, etc. This method of mapping continuous probability output to discrete levels has good practicality and stability, and is especially suitable for application in abnormal behavior discrimination and risk control system construction scenarios.

[0120] S5, based on the risk level result, performs content publishing subject matching priority update, participation state classification, and recommendation weight adjustment processing, and executes dynamic control operations based on the risk change trend result output by the content update tracking to realize continuous judgment of exposure risk and optimization of recommendation accuracy.

[0121] In this embodiment, S5 is specifically:

[0122] According to the risk level result, the matching priority of the content publishing subject in the candidate set is adjusted, the content publishing subject with a low risk level result is assigned a first priority, the content publishing subject with a medium risk level result is assigned a second priority, and the content publishing subject with a high risk level result is assigned a third priority, and the updated priority is written into the content publishing subject recommendation scheduling cache table for controlling the recommendation triggering order in task allocation.

[0123] After the risk level result is generated, the risk level and the corresponding matching priority can be associated and set through a priority classification mapping table to establish a mapping rule between the risk level and the scheduling priority, for example, setting a first priority for a low risk, a second priority for a medium risk, and a third priority for a high risk. After the system receives the risk level of each content publishing subject, the priority is assigned according to the mapping rule, and the result is written into a structured content publishing subject recommendation scheduling cache table. The cache table takes the unique identifier of the content publishing subject as the primary key field, contains the priority field, the update timestamp field, and the recommendation task state field, and is used for quick query and sorting by the triggering engine in the actual recommendation scheduling process. When scheduling a recommendation task, the candidate content publishing subjects are arranged in ascending order according to the priority field, and the content publishing subjects with higher priority are preferentially selected to enter the recommendation execution queue. For example, when a content publishing subject has stable recent behavior and no risk characteristics, it is rated as low risk, and its priority is first, so it will be triggered first for matching and pushing among the same batch of candidates.

[0124] The "risk level result" in the priority allocation mechanism is a discrete level identifier output by the residual learning network and converted by the discrimination structure, which is used to reflect the potential abnormal behavior risk degree of the content publishing subject in the current period. The first, second, and third priorities are scheduling level division standards designed according to the risk control strategy, respectively representing three scheduling levels of priority recommendation, medium recommendation, and limited recommendation. The content publishing subject recommendation scheduling cache table is an intermediate structure existing in the memory dataset, which saves the scheduling state information of each content publishing subject through the Key-Value mapping structure, facilitating the scheduler to complete the selection of recommended candidates in milliseconds. The priority updating process is dynamic, and the system will reevaluate the risk level of the content publishing subject after a new round of content update and behavior analysis, and refresh the priority field in the scheduling cache table in real time to ensure the timeliness and risk perception ability of the matching mechanism. This mechanism ensures that the recommendation strategy has controllability, scalability, and security in actual operation.

[0125] The current risk level result of the content publishing subject and the historical recommended participation record are called to establish a participation state label mapping in the portrait data, the content publishing subject corresponding to a low risk is marked as a recommended participation in state, the content publishing subject corresponding to a medium risk is marked as a to be confirmed state, and the content publishing subject corresponding to a high risk is marked as a suspended participation state, and the participation state is used for controlling the filtering logic of the recommended trigger when being dispatched.

[0126] To realize the linkage control of the risk level and the recommended task scheduling mechanism, the participation state label field can be constructed based on the portrait data of the content publishing subject, and the label mapping relationship is dynamically generated by calling the current risk level result of the content publishing subject and the historical recommended participation record thereof. The label mapping relationship is used for marking the recommended participation state of each content publishing subject in the current period. In specific implementation, the system first analyzes the latest risk level of each content publishing subject, if it is a low risk, it is directly marked as a recommended participation in state, indicating that the content publishing subject can normally participate in the current recommended task, if it is a medium risk, it is queried whether there is a high-frequency participation or controversial content record in the recent period, if there is, it is marked as a to be confirmed state, and whether it can participate in the recommendation needs to be further judged by the content review mechanism or manual review, if it is a high risk, it does not need to be further analyzed, and is directly marked as a suspended participation state, which is automatically shielded by the recommendation system in the current task, so as to avoid the content being continuously pushed. After the marking is completed, the state label is written into the structured field in the portrait data of the content publishing subject, and is bound with the data filtering logic of the recommended trigger, so as to realize the state checking and filtering processing before the scheduling execution.

[0127] The risk level result in the mechanism is a discrete classification result of the multi-source behavior characteristics after being predicted and risk classified by the residual learning network, and is used to drive the downstream logic branch. The recommended participation record contains the calling times, content pass rate, controversial feedback amount and other fields of the content publishing subject in the recommended task in a certain time range in the past, which is used as a reference auxiliary feature for state judgment. The portrait data is a basic data set for structuring and archiving the multi-dimensional information such as behavior, interaction and content of the content publishing subject on the platform, and the participation state label field is a core control bit thereof, which is used for real-time retrieval by the scheduled module. The three states of recommended participation in, to be confirmed and suspended participation constitute an enumeration value set of the participation state, which corresponds to different scheduling strategies and filtering logics, so as to ensure the fine control of the content publishing subject state by the system and improve the stability and risk control ability of the task distribution. The mechanism realizes the dynamic management of the participation permission based on the risk level driving, and is a key technical point for connecting the risk perception and the task scheduling in the recommendation system.

[0128] The time sequence sample is extracted based on the content update log of the content publishing subject, and the risk level result after each update is associated in time sequence to construct a risk change trend sequence. The short-term trend change direction and fluctuation amplitude are calculated by using the moving average method. When the trend sequence presents a continuous downward trend and the risk level is continuously maintained in the low risk interval, the content publishing subject matching priority and the recommended weight are dynamically recovered. If the trend sequence presents an upward direction and the fluctuation amplitude exceeds the risk threshold, the current recommended weight is reduced and the participation state is updated to suspended observation to complete the dynamic regulation operation of the risk change trend result.

[0129] To realize dynamic monitoring of the risk level of the content publishing subject and real-time adjustment of the recommendation strategy, the content publishing samples at continuous time points can be extracted based on the content update log of the content publishing subject, and the risk level result generated after each content update is combined to construct a risk change trend sequence arranged in time sequence. In the implementation process, the content update time axis can be first discretized according to time granularity such as hour, day or publishing period, and the risk level label corresponding to each time node is associated to form a one-to-one mapping trend data structure. Then the trend sequence is smoothed by using the moving average method to eliminate the influence of accidental fluctuations, and the change direction of the short-term trend is judged by calculating the mean slope in the sliding window. If the trend sequence presents a continuous decrease in the risk level value in multiple time windows, and the risk level is stably in the low risk interval in the continuous period, the gradual recovery operation of the priority and the recommended weight can be performed. The specific method is to improve the priority field value of the content publishing subject in the recommended scheduling cache table, and call back the recommended weight factor to the system default benchmark value, so that it reenters the normal recommendation channel. If the trend sequence slope reverses and the fluctuation amplitude exceeds the system set risk response threshold, it represents that the risk is accumulating. To avoid content out of control, the recommended weight should be immediately reduced and the participation state should be modified to suspended observation. The system scheduling logic will automatically skip the content publishing subject in the subsequent task to realize timely response to the risk growth trend.

[0130] The risk change trend sequence is a time series data structure constructed by combining a time axis and a risk level label, and is used to monitor the risk dynamics of the content publishing subject in the content updating process. The content update log records the timestamp and content identifier of each content publishing behavior of the content publishing subject, which is the basic index information for constructing the trend sequence. The risk level result is a discrete label output by the residual learning network after prediction, which can be converted into a continuous variable of the trend sequence by numerical mapping. The moving average method is a commonly used time series smoothing method, which takes the average of continuous values by setting a fixed length sliding window, reduces the interference of sudden points, and extracts the main trend characteristics. The trend change direction can be calculated by the mean value difference between adjacent windows, and the fluctuation amplitude reflects the short-term fluctuation intensity by the difference between the maximum and minimum values in the window. The risk threshold is a controllable boundary set by the system in the training stage by statistical risk level change distribution, which is used to distinguish normal fluctuations from abnormal surges. When the trend indicators meet the determination conditions, the system will trigger the regulation rules to dynamically modify the priority field, recommendation weight value and participation state label of the content publishing subject in the scheduling parameters, so as to realize the positive feedback regulation of the content updating behavior on the recommendation strategy. The mechanism enhances the predictive response ability of the system to potential risks through the linkage of risk trend and recommendation mechanism.

[0131] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions according to the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired or wireless (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0132] It should be understood that in various embodiments of the present application, the size of the sequence number of each process does not mean the order of execution, and the execution order of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0133] Those skilled in the art can understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0134] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other ways. For example, the above-described embodiments are merely illustrative, for example, the division of units is only a logical function division, and actual implementation can have another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or components shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0135] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0136] In addition, the functional units in each embodiment of the present application can be integrated into a processing unit, or each unit can be physically present alone, or two or more units can be integrated into one unit.

[0137] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1.An AI intelligent marketing content publishing subject matching recommendation method, characterized in that, Specifically comprising the following steps: S1, by fusing the content behavior data of the text content, image content, voice content, publishing time information and the distribution data of the attention user group published by the content publishing subject, a behavior content fusion graph for detecting the cooperation information missing state is generated; S2, based on the behavior content fusion graph, keyword density mutation detection, brand semantic coincidence detection and content publishing cycle anomaly detection are performed to determine whether the content publishing subject has cooperation information missing but behavior characteristics showing cooperation history; S2 specifically comprises the following steps: S201, extract the text content node in the behavior content fusion graph, construct a time series word frequency matrix, use a sliding window algorithm to segment and statistically analyze the density change rate of the keyword in the continuous time period, identify the density mutation point based on the mutation amplitude, mark the mutation node to establish a keyword density mutation detection sequence; S202, based on the keyword density mutation detection sequence, extract the corresponding text content node and image content node in the graph, match the semantic similarity with the pre-defined brand semantic vector, extract the keyword semantic link through the connection relationship between the nodes, construct a brand semantic coincidence path graph, and realize brand semantic coincidence detection; S203, based on the brand semantic coincidence path graph, extract the publishing time information node to construct a content publishing cycle sequence, use a cycle change recognition model to extract the variation point section of the publishing time, and convert the keyword density mutation detection sequence, brand semantic coincidence path graph and content publishing cycle sequence into event node sets on the corresponding time axis respectively, use an event alignment algorithm to calculate the intersection number of the three types of event nodes in a sliding time window, when the intersection number exceeds a preset threshold and there is no corresponding cooperation record data in the content publishing subject portrait, it is determined that the content publishing subject has cooperation information missing but behavior characteristics showing cooperation history; S3, based on the determination result, perform semantic co-occurrence path construction, propagation style similarity analysis and behavior residual discrimination operation to obtain the cooperation frequency implicit signal of the content publishing subject with cooperation information missing but behavior characteristics showing cooperation history; S4, fuse the cooperation frequency implicit signal with the interactive heat change index, topic linkage coincidence index and content publishing time distribution index, input the residual learning network for feature training, and output the risk level result for identifying whether the content publishing subject has exposure risk; S5, based on the risk level result, perform content publishing subject matching priority update, participation state classification, recommendation weight adjustment processing, and combine the risk change trend result output by the content update tracking to perform dynamic control operation to realize continuous judgment of exposure risk and optimization of recommendation accuracy. 2.The AI intelligent marketing content publisher matching recommendation method of claim 1, characterized in that, S1 specifically is: The text content, image content, voice content, publishing time information and attention user group distribution data published by the content publishing subject are respectively subjected to deep feature extraction, the text content obtains a text feature vector through a context semantic encoding network, the image content obtains an image feature vector through a multi-layer visual perception network, the voice content obtains a voice feature vector through a time-frequency convolution encoder, the publishing time information obtains a time feature vector through a time series encoding structure, the attention user group distribution data obtains an attention user group feature vector through a statistical normalization structure, and all the feature vectors are subjected to unified scale normalization processing to form a comparable feature set; Based on the comparable feature set, a multi-modal sequence alignment is performed according to the time sequence of the content publishing of the content publishing subject, the text feature vector, the image feature vector, the voice feature vector, the time feature vector and the attention user group feature vector are mapped into a multi-dimensional node sequence, the semantic similarity and the time affinity between nodes are calculated through a cross-modal fusion unit, and a weighted edge value is assigned, to form an initial multi-modal association structure containing nodes and edges; The initial multi-modal association structure is input into a heterogeneous graph embedding network, the nodes and edges are subjected to feature interaction and spectral smoothing processing through multi-head attention operation, and a behavior content fusion graph containing multi-dimensional behavior features of the text content, the image content, the voice content, the publishing time information and the attention user group distribution data is output, which is used to detect the cooperative information missing state. 3.The AI intelligent marketing content publisher matching recommendation method of claim 1, characterized in that, S201 specifically is: In the behavior content fusion graph, a complete keyword set contained in the text content node is extracted, a time series word frequency matrix segmented at equal time intervals is constructed according to the content publishing time information, and each matrix unit represents the word frequency count value of the corresponding keyword in the specified time period; On the basis of the time series word frequency matrix, each keyword is subjected to local interval traversal processing based on a sliding window algorithm, the word frequency change rate of each keyword in adjacent time periods is calculated, and the change rate sequence is compared with a mutation amplitude threshold value, and the time period with a word frequency change rate greater than the mutation amplitude threshold value is identified as a mutation candidate segment; The keyword appearance time and the content publishing time information involved in the mutation candidate segment are subjected to alignment analysis, the keywords with word frequency mutation occurring in the content intensive publishing section are screened out, the corresponding text content node is marked as a mutation node, and a keyword density mutation detection sequence is formed in turn. 4.The AI intelligent marketing content publisher matching recommendation method of claim 1, characterized in that, S202 specifically is: Based on the keyword density mutation detection sequence, the corresponding text content node and image content node are extracted, and semantic vector representations are obtained through a text semantic encoding network and an image semantic decoding network respectively, and each node representation is uniformly mapped into a shared semantic space; In the shared semantic space, the brand semantic vector contained in the brand semantic vector library is called, the cosine similarity of each text content node and image content node is calculated, and the nodes with a similarity value exceeding a set semantic similarity threshold value are marked as brand semantic candidate nodes; The brand semantic candidate node is taken as a starting point, a keyword semantic link is constructed based on the connection relationship between nodes in the behavior content fusion graph, a semantic path with a similarity higher than a set value and a frequency meeting statistical conditions between continuous nodes is screened, a brand semantic overlap path graph is aggregated, and brand semantic overlap detection is completed. 5.The AI intelligent marketing content publisher matching recommendation method of claim 1, characterized in that, S3 specifically comprises the following steps: S301, by calling a node set in which cooperation information is determined to be missing but behavior characteristics show a cooperation history, a semantic connection graph structure is established based on a keyword density mutation detection sequence and a brand semantic overlap path graph, a continuous node track is generated in the graph structure according to semantic correlation strength and time proximity, and a semantic co-occurrence path is formed by sequentially connecting; S302, based on the semantic co-occurrence path, the text content, image content, publication time information and interactive behavior characteristics of each node in each path are extracted, a content style code, a publication time interval code and an interactive rhythm code are generated using a propagation pattern feature extraction network, a propagation pattern similarity score of each pair of adjacent nodes is calculated, and propagation pattern similarity analysis is completed; S303, the propagation pattern similarity score of each semantic co-occurrence path and the semantic connection strength score are input into a residual difference discrimination structure, residual discrimination operation is performed to calculate path credibility residuals, and paths with residuals lower than a preset discrimination threshold are selected as a high credibility path set, the frequency of each node in the global semantic co-occurrence path graph in the high credibility path set is counted, and a cooperation frequency implicit signal is output. 6.The AI intelligent marketing content publisher matching recommendation method of claim 5, characterized in that, S302 specifically comprises: Based on the node sequence in the semantic co-occurrence path, the text content, image content, publication time information and interactive behavior characteristics corresponding to each node are extracted in turn, and semantic encoding operation is performed on the content data, interval discretization processing is performed on the time data, and sequence normalization operation is performed on the interactive data, to generate standard input formats respectively; The generated standard input formats are input into the propagation pattern feature extraction network respectively, the context association in the text content and the image content is extracted using a multi-scale convolution structure to form a content style code, the publication time information is modeled using a time recurrent unit to generate a publication time interval code, and the interactive behavior change is modeled using a frequency perception network to generate an interactive rhythm code; The content style code, the publication time interval code and the interactive rhythm code of any two adjacent nodes in the same semantic co-occurrence path are fused according to a preset weighting ratio to construct a composite propagation pattern vector, and the cosine similarity between the composite propagation pattern vectors is calculated to generate a corresponding propagation pattern similarity score, the similarity scores between consecutive adjacent nodes are grouped into a propagation pattern similarity score sequence, and are used in subsequent behavior residual discrimination operation structures. 7.The AI intelligent marketing content publisher matching recommendation method of claim 5, characterized in that, S303 specifically comprises: The calculated propagation pattern similarity scores in each semantic co-occurrence path and the semantic connection strength scores between adjacent nodes in the path are fused, the semantic connection strength scores are calculated by calculating the cosine similarity of the semantic vectors of adjacent nodes and weighted by the content homology probability to form a comprehensive index reflecting the semantic continuity of the path; after fusion, a multi-dimensional feature input sequence is constructed, and the node order is kept consistent; The fused feature input sequence is input into a residual difference discrimination structure composed of stacked difference normalization units and attention adjustment units, a path residual discrimination operation is performed, a deviation degree between propagation features and semantic structures is extracted, and a credibility residual score of each semantic co-occurrence path is output. The semantic co-occurrence paths with a credibility residual score lower than a preset discrimination threshold are screened as a high-credibility path set, the cumulative appearance frequency of each node in the global semantic co-occurrence path graph in the high-credibility path set is counted, and an implicit cooperation frequency signal in a case where the content publishing subject cooperation information is missing but the behavior characteristics show a cooperation history is output. 8.The AI intelligent marketing content publisher matching recommendation method of claim 1, wherein, S4 specifically is: The cooperation frequency implicit signal, the interaction heat change index, the topic linkage overlap index and the content publishing time distribution index are obtained, and after being obtained, normalization processing and time sequence synchronization are respectively performed, a feature alignment sequence is constructed through a sliding window mechanism, a fusion weight is calculated based on a correlation matrix between feature groups, a fusion feature vector is generated by weight proportion superposition, and training input data containing multi-source behavior mode features are formed; The training input data is input into a residual learning network composed of a residual connection structure, a feature channel compression unit and a nonlinear activation structure, the residual connection structure retains cross-layer information flow to improve convergence ability, the feature channel compression unit maintains the core feature dimension, the nonlinear activation structure performs mode enhancement processing on high-order features, and a stable and separable feature expression vector is output for risk level discrimination; The feature expression vector output by the residual learning network is input into a classification discrimination structure, a risk prediction score is generated through full connection transformation and probability mapping, the risk prediction score takes continuous values in the [0, 1] interval, representing the fitting strength between the content publishing subject behavior characteristics and the abnormal risk; then the risk prediction score is matched with the preset risk division interval, the interval range is determined according to the abnormal sample distribution statistics in the model training stage, the risk prediction score lower than the first threshold is mapped to low risk, the risk prediction score between the first threshold and the second threshold is mapped to medium risk, and the risk prediction score higher than the second threshold is mapped to high risk. Finally, the risk level result of whether the content publishing subject exists exposure risk is output. 9.The AI intelligent marketing content publisher matching recommendation method of claim 1, wherein, S5 specifically is: According to the risk level result, the matching priority of the content publishing subject in the candidate set is adjusted, the content publishing subject with a low risk level result is assigned to a first priority, the content publishing subject with a medium risk level result is assigned to a second priority, and the content publishing subject with a high risk level result is assigned to a third priority. The updated priority is written into the content publishing subject recommendation scheduling cache table in synchronization, which is used to control the recommendation trigger order in task allocation; The current risk level result of the content publishing subject and the historical recommendation participation record are called, a participation state label mapping is established in the portrait data, the content publishing subject corresponding to the low risk is marked as a recommended participation state, the content publishing subject corresponding to the medium risk is marked as a to-be-confirmed state, and the content publishing subject corresponding to the high risk is marked as a temporary participation state. The participation state is used to control the filtering logic of the recommendation trigger in scheduling; Based on the content publishing subject content update log extraction time sequence sample, and according to the time sequence, the risk level result after each update is associated, the risk change trend sequence is constructed, the sliding average method is used to calculate the short-term trend change direction and fluctuation amplitude, when the trend sequence presents continuous downward trend and the risk level is continuously maintained in the low risk interval, the content publishing subject matching priority and the recommended weight are dynamically recovered; If the trend sequence presents an upward direction and the fluctuation amplitude exceeds the risk threshold, the current recommended weight is reduced and the participation state is updated to pause observation, so as to complete the dynamic regulation operation of the risk change trend result.

Images