Video detection system for science popularization recommendation

Abstract

The invention discloses a video detection system for science popularization recommendation, which relates to the technical field of video detection and comprises a node generation unit, a preference distribution output unit, a passive node division unit, a prediction result generation unit, a combined feature analysis unit and a content detection unit. According to the invention, the active nodes and the passive nodes are constructed through the node generation unit according to the behavior data of each user and the brief introduction data of the popular science videos, so that the system can more obviously reflect the operation behaviors of the users for different videos subsequently, and meanwhile, a corresponding associated node is set for each node; in this way, the similarity between the users and the similarity between the science popularization videos can be determined, the preference distribution output unit is utilized by the system to refer to the preferences and habits of the similar users when the preferences of each user are analyzed, the preferences of the similar users are fused with the preferences of the current user, and the user experience is improved. And the recommendation interface of the user is enriched.

Description

Technical Field

[0001] The present invention relates to the technical field of video detection, in particular to a video detection system for popular science recommendation. Background Art

[0002] The video detection system for popular science recommendation mainly detects the content in popular science videos to ensure the accuracy of the video. At the same time, it pushes relevant popular science videos according to user interests through a personalized recommendation algorithm to improve the efficiency of knowledge acquisition. The invention patent with application number 202210214422.3 discloses "a video sensitive content detection method and system, the method comprising: executing some steps of the decoding method on the video to be detected, extracting the compression domain information of the video to be detected; judging whether the quality of the video to be detected is qualified according to the compression domain information; if the quality of the video to be detected is unqualified, determining that there is no sensitive content in the video to be detected; if the quality of the video to be detected is qualified, performing sensitive content detection on the video to be detected according to the compression domain information to determine whether there is sensitive content in the video to be detected. The present invention reduces the resource consumption of sensitive content detection and improves detection efficiency and detection accuracy."

[0003] The above-mentioned existing technology solves the problems of difficulty in processing massive amounts of video data and poor generalization ability for low-quality videos. However, during operation, since the attribute representations of the extracted text information and image information cannot be combined, the final video attribute representation cannot retain the original detailed features of the video to the greatest extent. At the same time, this method cannot recommend appropriate video information to users based on their preferences and those of similar users. Summary of the invention

[0004] The purpose of the present invention is to provide a video detection system for popular science recommendation to solve the problems raised in the above background technology.

[0005] To achieve the above-mentioned object, the present invention provides the following technical solution: a video detection system for popular science recommendation, comprising a prediction result generation unit, a combined feature analysis unit and a content detection unit; A node generation unit, wherein the node generation unit extracts user behavior data and science video introduction data of online video websites, sets active nodes and passive nodes, initializes attribute representations of nodes using the behavior data and introduction data, counts the associated nodes of each node, constructs a connection graph of active nodes and passive nodes through the behavior data, inputs the graph into the GNN model, evaluates the influence of associated nodes on passive nodes, and generates multiple low-dimensional representations in combination with the initial attribute representations; A preference distribution output unit, wherein the preference distribution output unit extracts dependency vectors of active nodes and passive nodes, randomly selects active nodes, and counts low-dimensional representations of directly connected passive nodes, thereby deriving an initial preference distribution of the active nodes. After extracting associated nodes of each active node, the low-dimensional representations and dependency vectors of passive nodes directly connected to the associated nodes are determined, and all preference distributions of the active nodes are calculated using the low-dimensional representations and dependency vectors. The fused attribute representations of the passive nodes and the fused preference distribution of the active nodes are analyzed through the preference distribution, the attribute representations of the passive nodes, and the low-dimensional representations. A passive node division unit, after determining the affinity values ​​between active nodes and passive nodes, classifies the passive nodes according to the behavior data to obtain recorded nodes and non-recorded nodes, uses the video profile data to analyze the score approximation between passive nodes, calculates the score value of the recorded node through the behavior data and the fusion preference distribution, and after determining the consistency between active nodes, calculates the matching degree of the active node in combination with the affinity value, consistency and score approximation, and inputs it into the matching analysis model to determine the matching degree of the recorded node.

[0006] Preferably, the node generation unit includes a node setting module, a similarity calculation module, a connection graph generation module and a low-dimensional representation determination module. After the node setting module extracts the behavior data of all users in the online video publishing website and the introduction data of the popular science video, two different types of nodes are set, namely active nodes and passive nodes. After initializing the attribute representation of each active node using the behavior data of different users, the attribute representation initialization operation of each passive node is completed through the introduction data of each video. The similarity calculation module calculates the similarity between active nodes through the behavior data, and analyzes the similarity between passive nodes based on the introduction data. When the similarity between a node and other nodes is greater than a preset value, Other nodes are used as associated nodes of the node. When the similarity between the node and other nodes is less than or equal to a preset value, other nodes are not used as associated nodes of the node. After the connection graph generation module determines the connection graph of the active node and the passive node through the operation information in the behavior data, the conversion matrix corresponding to the initial attribute representation of each passive node is counted, and the new initial attribute representation of the passive node is calculated according to the conversion matrix and the initial attribute representation. The low-dimensional representation determination module inputs the path between the nodes in the connection graph into the GNN model for analysis, analyzes the passive node and its associated nodes through the model, evaluates the influence of each associated node on the passive node, and generates corresponding multiple low-dimensional representations in combination with the initial attribute representation of the passive node.

[0007] Preferably, the preference distribution output unit includes a dependency vector acquisition module, a preference distribution calculation module and an attribute representation optimization module. After the dependency vector acquisition module extracts the dependency vector between the active node and the passive node, it randomly selects an active node, counts the low-dimensional representations of all passive nodes directly connected to it, and derives the initial preference distribution of the current active node according to the dependency vector and the low-dimensional representation of the passive node. After the preference distribution calculation module counts all associated nodes of the current active node, it sorts them according to similarity, calculates the low-dimensional representations of the passive nodes directly connected to the associated nodes and their dependency vectors in turn, and derives the initial preference distribution of the current active node using the low-dimensional representation of the passive node and the dependency vector between the associated node and the passive node. After the preference distribution of the points is obtained, the active nodes are reselected to perform the same operation. After the attribute representation optimization module determines the preference distribution of all active nodes, the associated nodes corresponding to each preference distribution in the active nodes are counted, and the attribute representations of each passive node directly connected to the associated nodes are extracted. The attribute representations of the passive nodes are combined with the preference distribution of the active nodes to obtain the current attribute representation of the passive nodes. All current attribute representations of the same passive node are accumulated to obtain the fused attribute representation of the passive node. The preference distribution is combined with the low-dimensional representation of each passive node directly connected to the associated nodes to obtain the current preference distribution of the active node. All current preference distributions of the same active node are accumulated to obtain the fused preference distribution of the active node.

[0008] Preferably, the passive node division unit includes an independent set construction module, an affinity analysis module, a recording node determination module and a proximity calculation module. After the independent set construction module determines the associated nodes of each node, the associated nodes are stored in the corresponding independent set. The affinity analysis module counts the intersection and union results of each active node independent set with other active node independent sets, calculates the affinity value between the active node and other active nodes using the intersection and union results, extracts the intersection and union results of each passive node independent set with other passive node independent sets, and calculates the affinity value between the passive node and other passive nodes according to the intersection and union results. After the recording node determination module obtains the behavior data of each active node, it classifies each passive node according to the scoring information in the behavior data. If the video name corresponding to the current passive node exists in the scoring information, it is marked as a recorded node, otherwise it is marked as a non-recorded node. The proximity calculation module counts the total number of people who have scored each video and the number of people who have scored under each score value through the user's behavior data and the introduction data of the popular science video, then calculates the score ratio of each passive node, and uses the score ratio to analyze the score similarity between the passive nodes.

[0009] Preferably, the passive node division unit further comprises a consistency analysis module, a matching calculation module and a node analysis module. The consistency analysis module calculates the score value of the recorded node through the behavior data corresponding to the active node and the fusion preference distribution. After arbitrarily selecting two active nodes and the corresponding recorded nodes, the consistency between the active nodes is analyzed using the comprehensive score value in the recorded node profile data and the score values ​​of the current two active nodes. The calculation is cyclically performed until the consistency between each active node and other active nodes is obtained. The matching calculation module calculates the matching between the active nodes according to the affinity value, consistency and the score approximation of the recorded nodes between each active node and other active nodes. After obtaining the user's behavior data, the node analysis module calculates the ratio of the number of videos of the active node under different score values ​​to the total number of videos, calculates the score approximation between the active nodes using the ratio, inputs the consistency, affinity value, the score value of the recorded node corresponding to the preference distribution and the score approximation between the active nodes into the matching analysis model for deduction, thereby outputting the consistency between the recorded nodes. The matching between the recorded nodes is deduced according to the consistency, affinity value, the matching between the active nodes and the score approximation between the recorded nodes.

[0010] Preferably, the prediction result generating unit comprises a mean determination module, a candidate node setting module and a prediction score calculation module. After the mean determination module determines the matching degree between the recorded nodes of the active node and the recorded nodes of other active nodes, it selects the non-recorded nodes corresponding to the active node, analyzes the similarity between the non-recorded nodes and the recorded nodes of the current active node, selects the associated nodes of the non-recorded nodes according to the similarity, and calculates the corresponding mean using the score values ​​of all associated nodes. After the candidate node setting module traverses the recorded nodes corresponding to all active nodes, it extracts all nodes that use the current non-recorded node as the recorded node. The active node of the point is determined, and the affinity value between it and the current active node is determined, and the active node with the largest affinity value is used as the candidate node of the current active node. The matching degree between the non-recorded node and its associated node is calculated according to the candidate node and the current active node. The predicted score calculation module analyzes the matching degree and mean of the non-recorded node and its associated node through the predicted score analysis algorithm, so as to output the non-recorded node score value corresponding to the active node. After determining the recommended popular science video according to the size of the non-recorded node score value of each active node, the video and its introduction data are transmitted to the visualization interface. The predicted score analysis algorithm is specifically as follows:

[0011] in, Indicates the specified active node Corresponding non-record node Rating value, Represents a non-record node and associated nodes The matching degree, Indicates the specified active node Corresponding associated nodes The rating value of Represents an associated node The comprehensive rating value of Represents a non-record node The corresponding mean, Indicates the designated active node. Represents a non-record node, Represents an associated node, and represents the coefficient, Represents a non-record node The number of associated nodes.

[0012] Preferably, the combined feature analysis unit includes a label classification module, a sample generation module, a representation extraction module and an attribute combination module. The label classification module sets three different types of identification information, namely positive identification information, mildly negative identification information and severely negative identification information. After the sample generation module extracts all popular science videos from the online video publishing website, it determines the text identification information and image identification information in each video, constructs corresponding sample data using the video data and the identification information, and divides the sample data into training data and verification data according to a preset ratio. After the representation extraction module obtains the attribute representation of all text contents in the training data and the multidimensional attribute representation of the image, the multidimensional attribute representation of each image is integrated into a single-dimensional attribute representation, the target area contained in each image is determined, and the attribute representation of the target area is analyzed according to the attribute representation of the image and the position of the pixel point in the target area. After the attribute combination module counts the attribute representation of the text content, the single-dimensional attribute representation of the image, the attribute representation of the target area, the attribute representation of the text identifier and the attribute representation of the image identifier in each training data, all the attribute representations are combined to generate a joint attribute representation of the training data.

[0013] Preferably, the content detection unit includes a representation processing module, a model training module, an adjustment coefficient determination module and a video review module. After the representation processing module counts the joint attribute representations of all training data and determines the relative position vector of each object in the joint attribute representation, the relative position vector is combined with the joint attribute representation to obtain an input matrix, and three weight matrices are set. The relationship scores between objects are calculated through the input matrix and the weight matrix. The relationship vector corresponding to the input matrix is ​​analyzed according to the relationship scores between each object and other objects. After the relationship vector and the input matrix are fused using a vector analysis algorithm, they are used as a new input matrix for iterative analysis until the loop ends and a new joint attribute representation is output. The model training module transfers the joint attribute representation of each training data to the SVM model, and uses the loss function to optimize the model parameters according to the offset value between the predicted result and the actual identification information, thereby obtaining the optimal model parameters. After the coefficient determination module obtains the joint attribute representation of all verification data, it transfers it to the SVM model, calculates the accuracy of each type of identification information according to the offset value between the predicted result and the actual identification information, sets the adjustment coefficient according to the accuracy, and analyzes the new prediction result using the probability values ​​and adjustment coefficients of different types in the model. The accuracy of the type is calculated according to the offset value between the new prediction result and the actual identification information. When the accuracy of all types is greater than the threshold, the model training is completed, otherwise the training is resumed. After the video review module counts the joint attribute representation of the latest uploaded popular science video, it transfers it to the SVM model for analysis. If the obtained identification information type is positive identification information, no operation is performed. If the type is mildly negative identification information or severely negative identification information, the video data is directly transmitted to the review interface. If the review fails, the popular science video is removed from the shelves. If the review passes, the identification information of the video is modified to positive identification information.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention constructs active nodes and passive nodes according to the behavior data of each user and the profile data of the popular science video through the node generation unit, so that the system can more significantly reflect the user's operation behavior on different videos in the future, and set corresponding associated nodes for each node. In this way, the similarities between users and popular science videos can be clearly defined, and the system uses the preference distribution output unit to refer to the preferences and habits of similar users when analyzing the preferences of each user, and merges the preferences of these similar users with the preferences of the current user. At the same time, the attribute representations of similar popular science videos are also merged, so as to enrich the user's recommendation interface, and the prediction result generation unit combines the user's preferences with the rating values ​​of similar users when pushing popular science videos that the user has not rated, so as to infer the user's satisfaction with these unrecorded popular science videos, so that the video style in the recommendation interface is more in line with the user's preferences; 2. The present invention analyzes the text content and image information contained in the popular science video through a combined feature analysis unit, determines the attribute representations corresponding to these text content and image information, and then combines them with the identification information to obtain the joint attribute representation corresponding to each video. This design can completely retain the original feature information of the video, facilitate the distinction between different videos, and ensure that the classification effect of the SVM model during the training process can be more significant. After further optimizing the joint attribute representation of each video using the content detection unit, these data are used to determine the parameters of the model. At the same time, in order to prevent large differences in the number of different types in the training data, an adjustment coefficient is set in the model to improve the analysis performance of the model. The trained SVM model is used to determine the identification information of the newly-listed popular science video, which effectively reduces the workload of the reviewers. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Provides a schematic diagram of the overall system flow for an embodiment of the present invention; Figure 2 An internal module block diagram of a node generation unit provided in an embodiment of the present invention; Figure 3 A block diagram of the internal modules of the preference distribution output unit provided in an embodiment of the present invention; Figure 4 An internal module block diagram of a passive node partitioning unit provided in an embodiment of the present invention; Figure 5 An internal module block diagram of a prediction result generation unit provided in an embodiment of the present invention; Figure 6 This is a block diagram of the internal modules of the content detection unit provided by an embodiment of the present invention.

[0016] In the figure: 1. Node generation unit; 101. Node setting module; 102. Similarity calculation module; 103. Connection graph generation module; 104. Low-dimensional representation determination module; 2. Preference distribution output unit; 201. Dependency vector acquisition module; 202. Preference distribution calculation module; 203. Attribute representation optimization module; 3. Passive node division unit; 301. Independent set construction module; 302. Affinity analysis module; 303. Record node determination module; 304. Approximation calculation module; 305. Consistency analysis module ; 306, matching degree calculation module; 307, node analysis module; 4, prediction result generation unit; 401, mean determination module; 402, candidate node setting module; 403, prediction score calculation module; 5, combined feature analysis unit; 501, label classification module; 502, sample generation module; 503, representation extraction module; 504, attribute combination module; 6, content detection unit; 601, representation processing module; 602, model training module; 603, adjustment coefficient determination module; 604, video review module. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] See also Figure 1-Figure 6 , the present invention provides a technical solution: a video detection system for popular science recommendation, comprising a prediction result generation unit 4, a combined feature analysis unit 5 and a content detection unit 6; Node generation unit 1, after extracting user behavior data and science video introduction data from online video websites, sets active nodes and passive nodes, initializes attribute representations of nodes using behavior data and introduction data, counts associated nodes of each node, constructs a connection graph of active nodes and passive nodes through behavior data, inputs it into the GNN model, evaluates the influence of associated nodes on passive nodes, and generates multiple low-dimensional representations in combination with initial attribute representations; Preference distribution output unit 2, the preference distribution output unit 2 extracts the dependency vectors of active nodes and passive nodes, randomly selects active nodes, and counts the low-dimensional representations of the passive nodes directly connected to them, thereby deriving the initial preference distribution of the active nodes. After extracting the associated nodes of each active node, the low-dimensional representations and dependency vectors of the passive nodes directly connected to the associated nodes are determined, and all preference distributions of the active nodes are calculated using the low-dimensional representations and dependency vectors. The fused attribute representations of the passive nodes and the fused preference distribution of the active nodes are analyzed through the preference distribution, the attribute representations of the passive nodes, and the low-dimensional representations. Passive node division unit 3, after the passive node division unit 3 determines the affinity values ​​between active nodes and passive nodes, it classifies the passive nodes according to the behavior data to obtain recorded nodes and non-recorded nodes, uses the video profile data to analyze the score approximation between passive nodes, calculates the score value of the recorded node through the behavior data and the fusion preference distribution, and after determining the consistency between active nodes, calculates the matching degree of the active node in combination with the affinity value, consistency and score approximation, and inputs it into the matching analysis model to determine the matching degree of the recorded node.

[0019] The node generation unit 1 includes a node setting module 101, a similarity calculation module 102, a connection graph generation module 103 and a low-dimensional representation determination module 104. After the node setting module 101 extracts the behavior data of all users in the online video publishing website and the introduction data of the popular science video, two different types of nodes are set, namely active nodes and passive nodes. After the attribute representation of each active node is initialized using the behavior data of different users, the attribute representation initialization operation of each passive node is completed through the introduction data of each video. The similarity calculation module 102 calculates the similarity between active nodes through the behavior data, and analyzes the similarity between passive nodes based on the introduction data. When the similarity between a node and other nodes is greater than a preset value, the other nodes are used as associated nodes of the node. When the similarity between a node and other nodes is less than a preset value, the other nodes are used as associated nodes of the node. When it is equal to the preset value, other nodes are not used as associated nodes of the node. After the connection diagram generation module 103 determines the connection diagram of the active node and the passive node through the operation information in the behavior data, the conversion matrix corresponding to the initial attribute representation of each passive node is counted, and the new initial attribute representation of the passive node is calculated according to the conversion matrix and the initial attribute representation. The low-dimensional representation determination module 104 inputs the path between the nodes in the connection diagram into the GNN model for analysis, analyzes the passive node and its associated nodes through the model, evaluates the influence of each associated node on the passive node, and generates corresponding multiple low-dimensional representations in combination with the initial attribute representation of the passive node. The behavior data includes the user number, the popular science video number, the popular science video name, the operation information and the rating information, and the introduction data includes the popular science video number, the popular science video name, the identification information and the comprehensive rating value; The preference distribution output unit 2 includes a dependency vector acquisition module 201, a preference distribution calculation module 202 and an attribute representation optimization module 203. After the dependency vector acquisition module 201 extracts the dependency vector between the active node and the passive node, it randomly selects an active node, counts the low-dimensional representations of all passive nodes directly connected to it, and derives the initial preference distribution of the current active node based on the dependency vector and the low-dimensional representation of the passive node. After the preference distribution calculation module 202 counts all the associated nodes of the current active node, it sorts them by similarity, calculates the low-dimensional representations of the passive nodes directly connected to the associated nodes and their dependency vectors in turn, and derives the current active node based on the low-dimensional representation of the passive node and the dependency vector between the associated node and the passive node. After the preference distribution of the active node is obtained, the active node is reselected to perform the same operation. After the attribute representation optimization module 203 determines the preference distribution of all active nodes, the associated nodes corresponding to each preference distribution in the active node are counted, and the attribute representation of each passive node directly connected to the associated node is extracted. The attribute representation of the passive node is combined with the preference distribution of the active node to obtain the current attribute representation of the passive node, and all current attribute representations of the same passive node are accumulated to obtain the fused attribute representation of the passive node. The preference distribution is combined with the low-dimensional representation of each passive node directly connected to the associated node to obtain the current preference distribution of the active node, and all current preference distributions of the same active node are accumulated to obtain the fused preference distribution of the active node. The passive node division unit 3 includes an independent set construction module 301, an affinity analysis module 302, a record node determination module 303 and a proximity calculation module 304. After the independent set construction module 301 determines the associated nodes of each node, the associated nodes are stored in the corresponding independent set. The affinity analysis module 302 counts the intersection and union results of each active node independent set with other active node independent sets, calculates the affinity value between the active node and other active nodes using the intersection and union results, extracts the intersection and union results of each passive node independent set with other passive node independent sets, and calculates the affinity value between the active node and other active nodes according to the intersection and union results. The affinity value between the active node and other passive nodes, after the recording node determination module 303 obtains the behavior data of each active node, it classifies each passive node according to the scoring information in the behavior data. If the video name corresponding to the current passive node exists in the scoring information, it is marked as a recorded node, otherwise it is marked as a non-recorded node. The similarity calculation module 304 uses the user's behavior data and the introduction data of the popular science video to count the total number of people who have scored each video and the number of people who have scored under each score, and then calculates the score ratio of each passive node, and uses the score ratio to analyze the score similarity between the passive nodes; The passive node division unit 3 also includes a consistency analysis module 305, a matching calculation module 306 and a node analysis module 307. The consistency analysis module 305 calculates the score value of the recorded node through the behavior data corresponding to the active node and the fusion preference distribution. After arbitrarily selecting two active nodes and the corresponding recorded nodes, the consistency between the active nodes is analyzed using the comprehensive score value in the recorded node profile data and the score value of the current two active nodes. The calculation is cyclic until the consistency between each active node and other active nodes is obtained. The matching calculation module 306 calculates the consistency between each active node and other active nodes according to the affinity value, consistency and recorded The matching degree between active nodes is calculated based on the scoring approximation of recorded nodes. After the node analysis module 307 obtains the user's behavior data, it calculates the ratio of the number of videos of active nodes under different scoring values ​​to the total number of videos, and uses the ratio to calculate the scoring approximation between active nodes. The consistency, affinity value, and recorded node scoring values ​​and scoring approximation corresponding to the preference distribution between active nodes are input into the matching analysis model for deduction, thereby outputting the consistency between recorded nodes. The matching degree between recorded nodes is deduced based on the consistency, affinity value, matching degree between active nodes, and scoring approximation between recorded nodes. The matching analysis model is specifically as follows:

[0020]

[0021] in, Indicates active node and Corresponding recorded nodes and The matching degree between Indicates that a node has been recorded and Active Node and The joint value of Indicates that a node has been recorded and The affinity value between Indicates active node and The matching degree between Indicates active node and The similarity of the ratings between Indicates active node and Corresponding recorded nodes and The consistency between and Indicates that the node has been recorded. and represents the active node, and Indicates the parameters, represents the passive node set, represents the set of active nodes, Indicates active node The corresponding recorded nodes under the preference distribution Rating value, Indicates active node The corresponding recorded nodes under the preference distribution Rating value; The prediction result generation unit 4 includes a mean determination module 401, a candidate node setting module 402 and a prediction score calculation module 403. After the mean determination module 401 determines the matching degree between the recorded nodes of the active node and the recorded nodes of other active nodes, it screens out the non-recorded nodes corresponding to the active node, analyzes the similarity between the non-recorded nodes and the recorded nodes of the current active node, selects the associated nodes of the non-recorded nodes according to the similarity, and calculates the corresponding mean using the score values ​​of all associated nodes. After the candidate node setting module 402 traverses the recorded nodes corresponding to all active nodes, it extracts all nodes that use the current non-recorded nodes as the recorded nodes. The active node of the record node determines the affinity value between it and the current active node, and the active node with the largest affinity value is used as the candidate node of the current active node. The matching degree between the non-record node and its associated node is calculated based on the candidate node and the current active node. The prediction score calculation module 403 analyzes the matching degree and mean of the non-record node and its associated node through the prediction score analysis algorithm, thereby outputting the non-record node score value corresponding to the active node. After determining the recommended popular science video according to the size of the non-record node score value of each active node, the video and its introduction data are transmitted to the visualization interface. The prediction score analysis algorithm is specifically as follows:

[0022] in, Indicates the specified active node Corresponding non-record node Rating value, Represents a non-record node and associated nodes The matching degree, Indicates the specified active node Corresponding associated nodes The rating value of Represents an associated node The comprehensive rating value of Represents a non-record node The corresponding mean, Indicates the designated active node. Represents a non-record node, Represents an associated node, and represents the coefficient, Represents a non-record node The number of associated nodes; The combined feature analysis unit 5 includes a label classification module 501, a sample generation module 502, a representation extraction module 503 and an attribute combination module 504. The label classification module 501 sets three different types of identification information, namely positive identification information, mildly negative identification information and severely negative identification information. After the sample generation module 502 extracts all popular science videos from the online video publishing website, it determines the text identification information and image identification information in each video, constructs corresponding sample data using the video data and the identification information, and divides the sample data into training data and verification data according to a preset ratio. After the representation extraction module 503 obtains the attribute representation of all text contents in the training data and the multidimensional attribute representation of the image, it integrates the multidimensional attribute representation of each image into a single-dimensional attribute representation, determines the target area contained in each image, and analyzes the attribute representation of the target area according to the attribute representation of the image and the position of the pixel point in the target area. After the attribute combination module 504 counts the attribute representation of the text content in each training data, the single-dimensional attribute representation of the image, the attribute representation of the target area, the attribute representation of the text identification and the attribute representation of the image identification, all the attribute representations are combined to generate a joint attribute representation of the training data. The content detection unit 6 includes a representation processing module 601, a model training module 602, an adjustment coefficient determination module 603 and a video review module 604. After the representation processing module 601 counts the joint attribute representations of all training data, after determining the relative position vector of each object in the joint attribute representation, the relative position vector is combined with the joint attribute representation to obtain an input matrix, and three weight matrices are set. The relationship scores between the objects are calculated through the input matrix and the weight matrix. The relationship vector corresponding to the input matrix is ​​analyzed according to the relationship scores between each object and other objects. After the relationship vector and the input matrix are fused by the vector analysis algorithm, it is used as a new input matrix for iterative analysis until the loop ends and a new joint attribute representation is output. The model training module 602 transmits the joint attribute representation of each training data to the SVM model, and uses the loss function to optimize the model parameters according to the offset value between the predicted result and the actual identification information, so as to obtain the optimal model parameters. The adjustment coefficient is determined. After the module 603 obtains the joint attribute representation of all verification data, it transmits it to the SVM model, calculates the accuracy of each type of identification information according to the offset value between the predicted result and the actual identification information, sets the adjustment coefficient according to the accuracy, and uses the probability values ​​and adjustment coefficients of different types in the model to analyze the new prediction results. The accuracy of the type is calculated according to the offset value between the new prediction result and the actual identification information. When the accuracy of all types is greater than the threshold, the model training is completed, otherwise the training is resumed. After the video review module 604 counts the joint attribute representation of the latest uploaded popular science video, it transmits it to the SVM model for analysis. If the obtained identification information type is positive identification information, no operation is performed. If the type is mildly negative identification information and severely negative identification information, the video data is directly transmitted to the review interface. If the review fails, the popular science video is removed from the shelves. If the review passes, the identification information of the video is modified to positive identification information. The vector analysis algorithm is specifically as follows:

[0023]

[0024] in, represents the intermediate matrix, represents the relationship vector, represents the projection matrix, represents the original input matrix, represents the connection function, represents the feedforward network function, represents a combinatorial function, represents the new input matrix, and Indicates an ordinal number.

[0025] Working principle: After the node setting module 101 in the node generation unit 1 initializes the attribute representation of the active node and the passive node, the similarity calculation module 102 is used to count the associated nodes corresponding to each node, and the connection graph generation module 103 calculates the new initial attribute representation of the passive node. After the low-dimensional representation determination module 104 analyzes multiple low-dimensional representations of each passive node, the dependency vector acquisition module 201 in the preference distribution output unit 2 is used to deduce the initial preference distribution of the active node. After the preference distribution calculation module 202 determines all the preference distributions of the active node, the attribute representation preference is used to obtain the initial attribute representation of the passive node. The fusion attribute representation of the passive nodes and the fusion preference distribution of the active nodes are analyzed by the preference distribution and the low-dimensional representation of the passive nodes. The associated nodes of each node are determined by the independent set construction module 301 in the passive node division unit 3. The affinity analysis module 302 calculates the affinity values ​​between the active node and other active nodes and between the passive node and other passive nodes. The recorded node determination module 303 is used to filter out the recorded nodes and non-recorded nodes corresponding to each active node. The similarity calculation module 304 uses the score ratio to analyze the score similarity between the passive nodes. The consistency analysis module 302 calculates the affinity values ​​between the active node and other active nodes and between the passive node and other passive nodes. The recorded node determination module 303 is used to filter out the recorded nodes and non-recorded nodes corresponding to each active node. The similarity calculation module 304 uses the score ratio to analyze the score similarity between the passive nodes. The analysis module 305 obtains the consistency between each active node and other active nodes, and uses the matching degree calculation module 306 to determine the matching degree between the active nodes. The node analysis module 307 calculates the matching degree between the recorded nodes corresponding to the active nodes. The mean determination module 401 in the prediction result generation unit 4 uses the score values ​​of all associated nodes to calculate the corresponding mean. The candidate node setting module 402 analyzes the matching degree between the non-recorded node and its associated nodes. The prediction score calculation module 403 determines the recommended popular science video. The label classification module 501 in the combined feature analysis unit 5 sets three different The sample generation module 502 is used to divide the sample data into training data and verification data. The representation extraction module 503 obtains the attribute representation of all text content in the training data, the one-dimensional attribute representation of the image and the attribute representation of the target area. The joint attribute representation of the training data is generated by the attribute combination module 504. The representation processing module 601 in the content detection unit 6 is used to output a new joint attribute representation. The model training module 602 obtains the optimal model parameters. The adjustment coefficient determination module 603 sets the adjustment coefficient according to the accuracy. The video review module 604 is used to remove the video that fails the review.

[0026] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0027] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A video detection system for popular science recommendation, comprising a prediction result generation unit (4), a combined feature analysis unit (5) and a content detection unit (6), characterized in that: A node generation unit (1), wherein the node generation unit (1) extracts user behavior data and science video introduction data from an online video website, sets active nodes and passive nodes, initializes attribute representations of the nodes using the behavior data and introduction data, counts the associated nodes of each node, constructs a connection graph of active nodes and passive nodes using the behavior data, inputs the graph into a GNN model, evaluates the influence of the associated nodes on the passive nodes, and generates multiple low-dimensional representations in combination with the initial attribute representations; A preference distribution output unit (2), wherein the preference distribution output unit (2) extracts dependency vectors of active nodes and passive nodes, randomly selects active nodes, and counts low-dimensional representations of directly connected passive nodes, thereby deriving an initial preference distribution of the active nodes. After extracting associated nodes of each active node, the low-dimensional representations and dependency vectors of passive nodes directly connected to the associated nodes are determined, and all preference distributions of the active nodes are calculated using the low-dimensional representations and dependency vectors. The fused attribute representations of the passive nodes and the fused preference distribution of the active nodes are analyzed through the preference distribution, the attribute representations of the passive nodes, and the low-dimensional representations. A passive node division unit (3) is provided. After the passive node division unit (3) determines the affinity values ​​between active nodes and passive nodes, it classifies the passive nodes according to the behavior data, thereby obtaining recorded nodes and non-recorded nodes. The score approximation between passive nodes is analyzed using the video profile data. The score values ​​of recorded nodes are calculated using the behavior data and the fusion preference distribution. After the consistency between active nodes is determined, the matching degree of active nodes is calculated by combining the affinity value, consistency and scoring approximation, and the matching degree is input into the matching analysis model to determine the matching degree of recorded nodes.

2. The video detection system for popular science recommendation according to claim 1, characterized in that: The node generation unit (1) comprises a node setting module (101), a similarity calculation module (102), a connection graph generation module (103) and a low-dimensional representation determination module (104). The node setting module (101) extracts the behavior data of all users in the online video publishing website and the introduction data of the popular science video, and sets two different types of nodes, namely active nodes and passive nodes. After initializing the attribute representation of each active node using the behavior data of different users, the attribute representation initialization operation of each passive node is completed using the introduction data of each video. The similarity calculation module (102) calculates the similarity between active nodes based on the behavior data, and analyzes the similarity between passive nodes based on the introduction data. When the node is similar to other nodes, the similarity between active nodes and passive nodes is calculated. When the similarity is greater than a preset value, other nodes are used as associated nodes of the node. When the similarity between the node and other nodes is less than or equal to the preset value, other nodes are not used as associated nodes of the node. The connection graph generation module (103) determines the connection graph of the active node and the passive node through the operation information in the behavior data, and then counts the conversion matrix corresponding to the initial attribute representation of each passive node, and calculates the new initial attribute representation of the passive node based on the conversion matrix and the initial attribute representation. The low-dimensional representation determination module (104) inputs the path between the nodes in the connection graph into the GNN model for analysis, analyzes the passive node and its associated nodes through the model, evaluates the influence of each associated node on the passive node, and generates corresponding multiple low-dimensional representations in combination with the initial attribute representation of the passive node.

3. The video detection system for popular science recommendation according to claim 1, characterized in that: The preference distribution output unit (2) comprises a dependency vector acquisition module (201), a preference distribution calculation module (202) and an attribute representation optimization module (203). After extracting the dependency vector between the active node and the passive node, the dependency vector acquisition module (201) randomly selects an active node, counts the low-dimensional representations of all passive nodes directly connected to the active node, and derives the initial preference distribution of the current active node based on the dependency vector and the low-dimensional representation of the passive node. After counting all associated nodes of the current active node, the preference distribution calculation module (202) sorts them according to similarity, and sequentially calculates the low-dimensional representations of the passive nodes directly connected to the associated nodes and their dependency vectors, and uses the low-dimensional representation of the passive node and the dependency between the associated node and the passive node to derive the initial preference distribution of the current active node. After the preference distribution of the current active node is derived by the vector, the active node is reselected to perform the same operation. After the attribute representation optimization module (203) determines the preference distribution of all active nodes, the associated nodes corresponding to each preference distribution in the active node are counted, the attribute representation of each passive node directly connected to the associated node is extracted, the attribute representation of the passive node is combined with the preference distribution of the active node to obtain the current attribute representation of the passive node, all current attribute representations of the same passive node are accumulated to obtain the fused attribute representation of the passive node, the preference distribution is combined with the low-dimensional representation of each passive node directly connected to the associated node to obtain the current preference distribution of the active node, and all current preference distributions of the same active node are accumulated to obtain the fused preference distribution of the active node.

4. The video detection system for popular science recommendation according to claim 1, characterized in that: The passive node division unit (3) comprises an independent set construction module (301), an affinity analysis module (302), a record node determination module (303) and a proximity calculation module (304). After the independent set construction module (301) determines the associated nodes of each node, the associated nodes are stored in the corresponding independent set. The affinity analysis module (302) counts the intersection and union results of each active node independent set with other active node independent sets, calculates the affinity value between the active node and other active nodes using the intersection and union results, extracts the intersection and union results of each passive node independent set with other passive node independent sets, and calculates the affinity value between the active node and other active nodes according to the intersection and union results. The result is to calculate the affinity value between the passive node and other passive nodes. After the recording node determination module (303) obtains the behavior data of each active node, it classifies each passive node according to the scoring information in the behavior data. If the video name corresponding to the current passive node exists in the scoring information, it is marked as a recorded node, otherwise it is marked as a non-recorded node. The similarity calculation module (304) uses the user's behavior data and the introduction data of the popular science video to count the total number of people who have scored each video and the number of people who have scored under each score value, and then calculates the score ratio of each passive node, and uses the score ratio to analyze the score similarity between the passive nodes.

5. The video detection system for popular science recommendation according to claim 4, characterized in that: The passive node division unit (3) further comprises a consistency analysis module (305), a matching calculation module (306) and a node analysis module (307). The consistency analysis module (305) calculates the score value of the recorded node through the behavior data corresponding to the active node and the fusion preference distribution. After arbitrarily selecting two active nodes and the corresponding recorded nodes, the consistency between the active nodes is analyzed using the comprehensive score value in the recorded node profile data and the score values ​​of the current two active nodes. The calculation is repeated until the consistency between each active node and the other active nodes is obtained. The matching calculation module (306) calculates the consistency between each active node and the other active nodes according to the consistency between each active node and the other active nodes. The matching degree between active nodes is calculated based on the affinity value, consistency and the score approximation of recorded nodes. After obtaining the user's behavior data, the node analysis module (307) calculates the ratio of the number of videos of active nodes under different score values ​​to the total number of videos, and uses the ratio to calculate the score approximation between active nodes. The consistency, affinity value, and recorded node score values ​​corresponding to the preference distribution and the score approximation between active nodes are input into the matching analysis model for deduction, thereby outputting the consistency between recorded nodes. The matching degree between recorded nodes is deduced based on the consistency, affinity value, matching degree between active nodes and the score approximation between recorded nodes.

6. The video detection system for popular science recommendation according to claim 1, characterized in that: The prediction result generation unit (4) comprises a mean determination module (401), a candidate node setting module (402) and a prediction score calculation module (403). The mean determination module (401) determines the matching degree between the recorded nodes of the active node and the recorded nodes of other active nodes, and then selects the non-recorded nodes corresponding to the active node, and analyzes the similarity between the non-recorded nodes and the recorded nodes of the current active node, selects the associated nodes of the non-recorded nodes according to the similarity, and calculates the corresponding mean value using the score values ​​of all the associated nodes. The candidate node setting module (402) traverses all the recorded nodes corresponding to the active nodes, and then extracts the non-recorded nodes. All active nodes that use the current non-recorded node as a recorded node determine the affinity value between the current active node and the current active node, and use the active node with the largest affinity value as a candidate node of the current active node. The matching degree between the non-recorded node and its associated node is calculated based on the candidate node and the current active node. The predicted score calculation module (403) analyzes the matching degree and mean of the non-recorded node and its associated node through a predicted score analysis algorithm, thereby outputting the non-recorded node score value corresponding to the active node. After determining the recommended popular science video based on the size of the non-recorded node score value of each active node, the video and its introduction data are transmitted to the visualization interface.

7. The video detection system for popular science recommendation according to claim 1, characterized in that: The combined feature analysis unit (5) comprises a label classification module (501), a sample generation module (502), a representation extraction module (503) and an attribute combination module (504). The label classification module (501) sets three different types of identification information, namely positive identification information, mildly negative identification information and severely negative identification information. The sample generation module (502) extracts all popular science videos from an online video publishing website, determines text identification information and image identification information in each video, constructs corresponding sample data using video data and identification information, and divides the sample data into training data and validation data according to a preset ratio. The representation extraction module (503) obtains the attribute representations of all text contents in the training data and the multi-dimensional attribute representations of the images, integrates the multi-dimensional attribute representations of each image into a single-dimensional attribute representation, determines the target area contained in each image, and analyzes the attribute representation of the target area based on the attribute representation of the image and the position of the pixel points in the target area. The attribute combination module (504) calculates the attribute representations of the text contents in each training data, the single-dimensional attribute representations of the images, the attribute representations of the target area, the attribute representations of the text identifiers and the attribute representations of the image identifiers, and then combines all the attribute representations to generate a joint attribute representation of the training data.

8. The video detection system for popular science recommendation according to claim 1, characterized in that: The content detection unit (6) includes a representation processing module (601), a model training module (602), an adjustment coefficient determination module (603) and a video review module (604). The representation processing module (601) counts the joint attribute representations of all training data, determines the relative position vector of each object in the joint attribute representation, combines the relative position vector with the joint attribute representation to obtain an input matrix, sets three weight matrices, calculates the relationship scores between the objects through the input matrix and the weight matrix, analyzes the relationship vector corresponding to the input matrix according to the relationship scores between each object and other objects, fuses the relationship vector with the input matrix using a vector analysis algorithm, and uses it as a new input matrix for iterative analysis until the loop ends and a new joint attribute representation is output. The model training module (602) transmits the joint attribute representation of each training data to the SVM model, optimizes the model parameters according to the offset value between the predicted result and the actual identification information using a loss function, thereby obtaining the optimal The optimal model parameters, the adjustment coefficient determination module (603) obtains the joint attribute representation of all verification data and transmits it to the SVM model, calculates the accuracy of each type of identification information according to the offset value between the predicted result and the actual identification information, sets the adjustment coefficient according to the accuracy, analyzes the new prediction result using the probability values ​​and adjustment coefficients of different types in the model, calculates the accuracy of the type according to the offset value between the new prediction result and the actual identification information, and completes the model training when the accuracy of all types is greater than the threshold, otherwise the training is resumed. The video review module (604) counts the joint attribute representation of the latest uploaded popular science video and transmits it to the SVM model for analysis. If the obtained identification information type is positive identification information, no operation is performed. If the type is mildly negative identification information or severely negative identification information, the video data is directly transmitted to the review interface. If the review fails, the popular science video is removed from the shelves. If the review passes, the identification information of the video is modified to positive identification information.

Images