College student ai capability literacy comprehensive evaluation method and device based on fusion perception improved graph neural network and storage medium

Abstract

This invention relates to the field of artificial intelligence technology, and discloses a comprehensive evaluation method for college students' AI competence and literacy that integrates perceptual improved graph neural networks. The method includes: processing structured and unstructured data from students across various dimensions of AI courses to obtain dimensional feature vectors for each dimension; calculating a first data score based on the semantic similarity between the sub-dimensional feature vectors of each dimension and a baseline vector; calculating a second data score for each dimension based on the first data score; constructing a bicentric heterogeneous bipartite graph as a student node graph based on multiple dimensional feature vectors and corresponding second data scores; generating a third data score for each dimension based on the student node graph using a graph neural network model and matching it with corresponding competence labels; and generating a student's AI competence analysis report based on the competence labels and third data scores. This invention achieves a comprehensive and quantifiable multi-dimensional analysis and evaluation of students' AI competence and literacy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to a method, device, and storage medium for comprehensively evaluating college students' AI competence and literacy by integrating a perception-improved graph neural network. Background Technology

[0002] In an era where artificial intelligence (AI) technology is sweeping the globe and becoming a key element of national core competitiveness, analyzing user behavior data generated by AI-related applications can effectively construct user profiles. Based on these profiles, suggestions can be made to improve users' abilities to utilize AI applications. For example, analyzing student data from online intelligent education applications can yield reports analyzing students' learning abilities in online courses (such as AI courses). These reports can identify students' strengths and weaknesses in these courses and provide improvement suggestions. Therefore, comprehensive analysis of user data generated by AI-related applications can guide users to more efficiently improve their ability to use AI applications based on the analysis results of large amounts of user data. However, because applications typically include multiple functional modules, the user behavior data they generate is multidimensional, making accurate and effective analysis a highly challenging technical problem. Summary of the Invention

[0003] In light of the above, it is necessary to provide a method for multi-dimensional analysis and processing of data generated by students or users in artificial intelligence courses through application apps. This method calculates the scores of the data in each dimension based on node graphs and graph neural networks and matches them with corresponding labels, thereby generating a data analysis report that is both comprehensive and has quantifiable benchmarks.

[0004] To achieve the above objectives, this invention provides a comprehensive evaluation method for college students' AI competence and literacy that integrates a perceptual improved graph neural network, comprising: Based on the multiple dimensions preset in the artificial intelligence course, the dimensional data corresponding to each dimension of the student in the artificial intelligence course is obtained. The dimensional data includes structured data and unstructured data. Each dimension includes multiple sub-dimensions. The data of each sub-dimension is set with multiple rating levels. Each rating level corresponds to a benchmark vector. The structured data contained in the dimensional data of each dimension is processed to obtain the first numerical feature vector with standardized numerical representation. The semantics of the unstructured data contained in the dimensional data of each dimension are extracted and encoded into the second numerical feature vector. The first numerical feature vector and the second numerical feature vector corresponding to each dimension are concatenated to obtain the dimensional feature vector corresponding to each dimension. The dimensional feature vectors corresponding to all dimensions are concatenated to obtain the comprehensive feature vector. Extract the sub-dimensional feature vectors corresponding to each sub-dimension from the dimensional feature vectors corresponding to each dimension, and calculate the semantic similarity between each sub-dimensional feature vector and the benchmark vector corresponding to each rating level. Calculate the first data score for each sub-dimension based on multiple semantic similarities for each sub-dimension, and then sum the weighted first data scores for all sub-dimensions included in each dimension to obtain the second data score for each dimension. Extract the baseline vector corresponding to the highest rating level of each sub-dimension, summarize the baseline vectors corresponding to the highest rating levels of all sub-dimensions included in each dimension to obtain the dimension baseline vector of each dimension, and calculate the average dimension baseline vector by averaging the dimension baseline vectors of all dimensions. A bicentric heterogeneous bipartite graph is constructed as the student node graph, wherein the first central node is embedded with the comprehensive feature vector, the second central node is embedded with the average dimension reference vector, each first child node of the first central node is embedded with the second numerical feature vector of each dimension, each second child node of the second central node is embedded with the dimension reference vector of each dimension, and the edges include the connection edges between the first central node and the first child node, the connection edges between the second central node and the second child node, and the cross-side matching edges between the first child node and the second child node in the same dimension; The second numerical feature vector corresponding to each first child node of the student node graph is aggregated and optimized by a pre-trained fusion-sensory graph neural network model. The fusion-sensory graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the dual-center heterogeneous bipartite graph at different receptive field scales. After fusing the features output by the above modules, the third data score of each first child node corresponding to each dimension is obtained by mapping through a fully connected layer. A student ability feature vector is constructed based on all third-party data scores. The corresponding ability labels are matched based on the student ability feature vector. An AI ability analysis report of the student is generated based on the ability labels and the third-party data scores.

[0005] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to execute the comprehensive evaluation method for college students' AI ability and literacy based on the fusion perception improved graph neural network as described above.

[0006] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the comprehensive evaluation method for college students' AI ability and literacy based on the fusion perception improved graph neural network as described above.

[0007] Furthermore, to achieve the above objectives, the present invention provides a user behavior analysis method based on graph neural networks, the method comprising: Based on multiple dimensions pre-set for the application APP, dimension data corresponding to each dimension is extracted from user behavior data generated by the user in the application APP. The dimension data includes structured data and unstructured data. Each dimension includes multiple sub-dimensions. The user behavior data of each sub-dimension is set with multiple rating levels. Each rating level corresponds to a baseline vector. The structured data contained in the dimensional data of each dimension is processed to obtain the first numerical feature vector with standardized numerical representation. The semantics of the unstructured data contained in the dimensional data of each dimension are extracted and encoded into the second numerical feature vector. The first numerical feature vector and the second numerical feature vector corresponding to each dimension are concatenated to obtain the dimensional feature vector corresponding to each dimension. The dimensional feature vectors corresponding to all dimensions are concatenated to obtain the comprehensive feature vector. Extract the sub-dimensional feature vectors corresponding to each sub-dimension from the dimensional feature vectors corresponding to each dimension, and calculate the semantic similarity between each sub-dimensional feature vector and the benchmark vector corresponding to each rating level. Calculate the first user behavior score for each sub-dimension based on multiple semantic similarities for each sub-dimension, and then sum the weighted scores of the first user behavior scores for all sub-dimensions included in each dimension to obtain the second user behavior score for each dimension. A dual-center heterogeneous bipartite graph is constructed as the user node graph, wherein the first center node embeds the comprehensive feature vector, the second center node embeds the average dimension reference vector, each first child node of the first center node embeds the second numerical feature vector of each dimension, each second child node of the second center node embeds the dimension reference vector of each dimension, and the edges include the connection edges between the first center node and the first child node, the connection edges between the second center node and the second child node, and the cross-side matching edges between the first child node and the second child node in the same dimension; The second numerical feature vector corresponding to each first child node of the user node graph is subjected to feature aggregation and feature optimization by a pre-trained fusion-aware graph neural network model. The fusion-aware graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the user node graph at different receptive field scales. After fusing the features output by the above modules, the third user behavior score corresponding to each first child node in each dimension is obtained by mapping through a fully connected layer. A user behavior feature vector is constructed based on all third-party user behavior scores. The corresponding user tags are matched based on the user behavior feature vectors. An analysis report for the user is generated based on the user tags and the third-party user behavior scores.

[0008] Optionally, the step of processing the structured data contained in the dimensional data of each dimension to obtain a first numerical feature vector with standardized numerical representation includes: The structured data is mapped to a preset numerical range to obtain standard structured data; Identify missing data in the standard structured data, calculate the padding value for the missing data, and pad the missing data in the standard structured data according to the padding value to obtain the padded data; The filled data is subjected to outlier identification, the identified outliers are replaced according to preset rules to obtain standardized values, and the standardized values ​​are converted into a one-dimensional first numerical feature vector.

[0009] Optionally, the step of extracting semantics from the unstructured data contained in the dimensional data of each dimension and encoding it into a second numerical feature vector includes: The unstructured data is input into a preset large language model. Redundant information and meaningless content in the unstructured data are removed by the prompts of the large language model to obtain the core semantic text. The core semantic text is feature-encoded using a pre-defined artificial intelligence text embedding model to obtain a second numerical feature vector.

[0010] Optionally, the step of calculating the first user behavior score corresponding to each sub-dimension based on multiple semantic similarities corresponding to each sub-dimension, and then performing a weighted summation of the first user behavior scores corresponding to all sub-dimensions included in each dimension to obtain the second user behavior score corresponding to each dimension, includes: Obtain the preset weight values ​​for each rating level of each sub-dimension as the first weight value of the semantic similarity corresponding to each rating level in each sub-dimension; Obtain the pre-set second weight value for the standardized numerical values ​​corresponding to each dimension; Based on the semantic similarity of each rating level in each sub-dimension and the first weight value, as well as the standardized value of the dimension to which each sub-dimension belongs and the second weight value, the first user behavior score corresponding to each sub-dimension is calculated using a weighted summation formula. Obtain the third weight value pre-set for each sub-dimension, and calculate the second user behavior score corresponding to each dimension using a weighted summation formula based on the first user behavior score of the multiple sub-dimensions contained in each dimension and the third weight value.

[0011] Optionally, the construction of a bicentric heterogeneous bipartite graph as a user node graph further includes: The second user behavior score for each dimension is used as the edge weight of the edge connecting the corresponding first child node and the first central node. Obtain the fourth weight value pre-set for each dimension as the edge weight of the edge connecting each second child node and the second center node; The cosine similarity between the second numerical feature vector of the first child node and the dimensional reference vector of the second child node in the same dimension is calculated as the edge weight of the edge connecting the first child node and the second child node in the same dimension.

[0012] Optionally, the second numerical feature vector corresponding to each first child node of the user node graph is subjected to feature aggregation and feature optimization by a pre-trained fusion-aware graph neural network model. The fusion-aware graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the user node graph at different receptive field scales. After fusing the features output by the above modules, a third user behavior score corresponding to each first child node in each dimension is obtained through mapping via a fully connected layer, including: The graph neural network model constructs an adjacency matrix based on the nodes and edges of the user node graph. The rows and columns of the adjacency matrix represent the nodes in the user node graph, and the elements in the adjacency matrix indicate whether there are weighted edges between the nodes. The adjacency matrix is ​​then normalized. The graph convolutional layer performs local neighborhood feature aggregation on the normalized adjacency matrix to obtain the node feature matrix corresponding to each first child node of the user node graph. The dependency features of the user node graph are obtained by capturing the global temporal and long-distance dependencies between nodes in the node feature matrix through the multi-head self-attention mechanism module. The fine-grained and coarse-grained features of the node feature matrix are extracted through the multi-scale feature extraction module. The node features of the node feature matrix are optimized based on the dependency features and the fine-grained and coarse-grained features. The resulting optimized node features are then input into the fully connected layer and output layer of the graph neural network model for normalization and mapping to obtain the third user behavior score corresponding to each first child node.

[0013] Optionally, generating the user's analysis report based on the user tags and the third user behavior score includes: Get the pre-set analysis report template; Obtain the corresponding tag interpretation text based on the user tags and write it into the preset tag interpretation area in the analysis report template; Obtain the third-party user behavior scores corresponding to all dimensions and write them into the preset dimension display area in the analysis report template; The user rating is generated based on the third user behavior score corresponding to all dimensions, and written into the preset rating analysis area in the analysis report template.

[0014] This invention processes multi-dimensional structured and unstructured data generated by users in an application (APP) to obtain dimensional feature vectors for each dimension. It calculates a second user behavior score for each dimension based on the similarity between the feature vectors and a baseline vector. A user node graph is constructed based on the dimensional feature vectors and corresponding second user behavior scores. A graph neural network model is used to generate a third user behavior score for each dimension based on the user node graph and match it with corresponding user tags. An analysis report is generated based on the user tags and the third user behavior score. This invention achieves standardized processing and deep semantic fusion of multi-source heterogeneous data. It explicitly models the semantic and quantitative relationships of dimensional data through the node graph, dynamically aggregates information from the graph using a graph neural network, optimizes and calibrates the second user behavior score, and outputs a highly interpretable user behavior analysis report. Attached Figure Description

[0015] Figure 1 This is an exemplary system architecture diagram in which the present invention can be applied; Figure 2 This is a flowchart of an embodiment of the comprehensive evaluation method for college students' AI ability and literacy that integrates perceptual improved graph neural networks according to the present invention; Figure 3 This is a flowchart of an embodiment of the user behavior analysis method based on graph neural networks of the present invention; Figure 4 This is a schematic diagram of an embodiment of the computer device of the present invention; Figure 5 This is a schematic diagram of a module of an embodiment of the user behavior analysis device based on graph neural networks of the present invention; Figure 6 This is a model architecture diagram of an embodiment of the fusion perceptual graph neural network model of the present invention.

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

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0018] It should be noted that the descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0019] like Figure 1 As shown, system architecture 100 may include terminals 101, 102, and 103, network 104, and server 105. Network 104 is used as a medium to provide a communication link between terminals 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0020] Users can use terminals 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminals 101, 102, and 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0021] Terminals 101, 102, and 103 can be various electronic devices with displays and web browsing capabilities, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), laptops, and desktop computers, etc.

[0022] Server 105 can be a server that provides various services, such as a backend server that supports the pages displayed on terminals 101, 102, and 103.

[0023] It should be noted that the comprehensive evaluation method for college students' AI ability and literacy based on fusion perception improved graph neural network and the user behavior analysis method based on graph neural network provided in the embodiments of this application are generally executed by the terminal, and correspondingly, the computer-readable storage medium is generally set in the terminal.

[0024] It should be understood that Figure 1 The number of terminals, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminals, networks, and servers can be included.

[0025] like Figure 2 The diagram shows a flowchart of an embodiment of the comprehensive evaluation method for college students' AI competence and literacy based on an improved perceptual graph neural network according to the present invention. The user behavior analysis method based on the graph neural network includes steps S11-S17: S11. Based on the multiple dimensions preset in the artificial intelligence course, obtain the dimension data corresponding to each dimension of the student in the artificial intelligence course. The dimension data includes structured data and unstructured data. Each dimension includes multiple sub-dimensions. The data of each sub-dimension is set with multiple rating levels. Each rating level corresponds to a benchmark vector. S12. Process the structured data contained in the dimension data of each dimension to obtain the first numerical feature vector of standardized numerical representation. Extract semantics from the unstructured data contained in the dimension data of each dimension and encode it into the second numerical feature vector. Concatenate the first numerical feature vector and the second numerical feature vector corresponding to each dimension to obtain the dimension feature vector corresponding to each dimension. Concatenate the dimension feature vectors corresponding to all dimensions to obtain the comprehensive feature vector. S13. Extract the sub-dimension feature vectors corresponding to each sub-dimension from the dimensional feature vectors corresponding to each dimension, calculate the semantic similarity between each sub-dimension feature vector and the benchmark vector corresponding to each rating level, calculate the first data score corresponding to each sub-dimension based on the multiple semantic similarities corresponding to each sub-dimension, and perform a weighted summation of the first data scores corresponding to all sub-dimensions included in each dimension to obtain the second data score corresponding to each dimension. S14. Extract the baseline vector corresponding to the highest rating level of each sub-dimension, summarize the baseline vectors corresponding to the highest rating levels of all sub-dimensions included in each dimension to obtain the dimension baseline vector of each dimension, and calculate the average dimension baseline vector of all dimensions. S15. Construct a bi-center heterogeneous bipartite graph as a student node graph, wherein the first center node embeds the comprehensive feature vector, the second center node embeds the average dimension reference vector, each first child node of the first center node embeds the second numerical feature vector of each dimension, each second child node of the second center node embeds the dimension reference vector of each dimension, and the edges include the connection edges between the first center node and the first child node, the connection edges between the second center node and the second child node, and the cross-side matching edges between the first child node and the second child node in the same dimension; S16. The second numerical feature vector corresponding to each first child node of the student node graph is aggregated and optimized by a pre-trained fusion-sensory graph neural network model. The fusion-sensory graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the dual-center heterogeneous bipartite graph at different receptive field scales. After fusing the features output by the above modules, the third data score of each first child node corresponding to each dimension is obtained by mapping through a fully connected layer. S17. Construct a student ability feature vector based on all third-party data scores, match the corresponding ability tags based on the student ability feature vector, and generate an AI ability analysis report for the student based on the ability tags and the third-party data scores.

[0026] like Figure 3 The flowchart shown is an embodiment of the user behavior analysis method based on graph neural networks of the present invention, which includes steps S21-S27.

[0027] S21. Based on multiple dimensions pre-set for the application APP, extract the dimension data corresponding to each dimension from the user behavior data generated by the user in the application APP. The dimension data includes structured data and unstructured data. Each dimension includes multiple sub-dimensions. The user behavior data of each sub-dimension is set with multiple rating levels. Each rating level corresponds to a baseline vector. In one embodiment, the application (APP) can be an online education APP for artificial intelligence. The users of the application APP are students who use the online APP to learn artificial intelligence courses. The user behavior data generated by the users in the application APP includes learning data such as students' learning grades or class reports generated through the online APP for artificial intelligence courses. For example, test scores generated by students taking online tests on the online APP, or class reports generated by students uploading relevant information from offline practice to the online APP.

[0028] In one embodiment, multiple dimensions are pre-defined for the application app based on the types of user behavior generated within the app. For example, when the application app is an online AI education app, five dimensions are generated based on the course content and course indicators contained in the app: a first dimension representing the understanding of the history of AI development, a second dimension representing the theoretical foundation of AI, a third dimension representing the understanding of how AI operates, a fourth dimension representing the practical application ability of AI, and a fifth dimension representing the ethical and moral understanding of AI.

[0029] In one embodiment, each dimension includes multiple sub-dimensions. For example, the sub-dimensions of the first dimension, representing the understanding of the history of artificial intelligence development, include: mastery of key time nodes in the history of artificial intelligence development, understanding of the evolution of core artificial intelligence technologies, and understanding of the contributions of important scientists in artificial intelligence; the sub-dimensions of the second dimension, representing the theoretical foundation of artificial intelligence, include: understanding of basic artificial intelligence concepts, understanding of the logic of core artificial intelligence algorithms, and the ability to judge the applicable scenarios of artificial intelligence technology; the sub-dimensions of the third dimension, representing the understanding of how artificial intelligence operates, include: the ability to deconstruct the principles of artificial intelligence technology, the ability to sort out the logic of artificial intelligence operation processes, and the ability to identify key links in the operation of artificial intelligence; the sub-dimensions of the fourth dimension, representing the practical application ability of artificial intelligence, include: proficiency in operating artificial intelligence tools, accuracy in adapting to artificial intelligence scenarios, and the degree of innovation in artificial intelligence achievements; and the sub-dimensions of the fifth dimension, representing the ethical and moral understanding of artificial intelligence, include: the ability to identify ethical risks of artificial intelligence, the level of awareness of compliant use of artificial intelligence, and the understanding of the social impact of artificial intelligence.

[0030] In one embodiment, three rating levels are pre-set for user behavior data in each sub-dimension: Basic, Good, and Excellent. Corresponding textual description standards are formulated for each standard level of each sub-dimension. The textual description standards corresponding to each rating level of each sub-dimension are converted into semantic vectors through a large language model. A set of standard vectors for that dimension is generated based on the semantic vectors corresponding to the rating levels belonging to the same sub-dimension.

[0031] For example, the sub-dimension of AI core algorithm logic cognition in the second dimension representing the theoretical foundation of artificial intelligence has the following textual description standards: Basic level: "Able to identify common AI algorithm names and understand the basic functions of algorithms," which is converted into the corresponding 768-dimensional semantic vector B1; Good level: "Able to explain the basic logic of two core algorithms and make a preliminary judgment on the applicable scenarios of the algorithms," which is converted into the corresponding 768-dimensional semantic vector B2; Excellent level: "Able to clearly explain the working principles of more than three core algorithms such as neural networks and decision trees, and accurately analyze the applicable scenarios and limitations of the algorithms," which is converted into the corresponding 768-dimensional semantic vector B3. Based on the above semantic vectors, the standard vector set B={B1, B2, B3} of the sub-dimension of AI core algorithm logic cognition in the second dimension is obtained.

[0032] In this embodiment, by pre-setting the data content classification of "dimension-subdimension-rating level" for the application APP, as well as the data structure classification of structured and unstructured data, user behavior data can be comprehensively and meticulously divided. This provides quantitative standards and standard values ​​for similarity calculation for subsequent user behavior data processing and analysis, making it easier to match the corresponding level of user behavior data to calculate refined analysis indicator values.

[0033] In one embodiment, based on multiple dimensions pre-set for the application (APP), dimension data corresponding to each dimension is extracted from user behavior data generated by the user in the application (APP), including: A data collection list for user behavior data is established based on multiple dimensions pre-set for the application APP. The data collection list includes structured data collection items and unstructured data collection items for each dimension. Based on the structured data collection items of each dimension, obtain multiple structured data generated by the user in the application APP, and import the multiple structured data and corresponding traceability information into the data table corresponding to each dimension. Based on the unstructured data collection items of each dimension, obtain multiple unstructured data generated by the user in the application APP, and convert the multiple unstructured data into plain text documents in a preset format; The structured data stored in the data table corresponding to each dimension and the unstructured data stored in plain text document format are used as the dimension data corresponding to each dimension.

[0034] In one embodiment, the structured data can be obtained through cross-validation by two people to control the data error rate; the unstructured data can be obtained through data transcription (such as audio to text), and after acquisition, it is checked and verified to avoid semantic deviations.

[0035] In one embodiment, after extracting the dimension data corresponding to each dimension from the user behavior data generated by the application APP, the method further includes: when there are missing data items in the structured and unstructured data collection items of the data collection list, determining the reason for the missing data in the collection item and marking it. For example, if the structured data collection item "Practical task accuracy rate" is missing, the reason is marked as "not participating in the practical task"; if the unstructured data collection item "Development history special analysis report" is missing, the reason is marked as "not submitting the special report".

[0036] For example, in an online AI education app, the structured data includes: students' developmental history unit test scores in AI courses as the first dimension; students' theoretical closed-book exam scores as the second dimension; students' principle breakdown test scores as the third dimension; students' group collaboration contribution scores, practical task accuracy rates, and process breakdown completion times as the fourth dimension; and students' ethical risk identification test scores and scenario simulation decision compliance scores as the fifth dimension. The source information for the first dimension's structured data is the developmental history unit test scores (data source). The source information for the practical task accuracy rate in the fourth dimension's structured data is: Practical task accuracy rate = Number of correct operation steps / Total number of steps (the calculation formula for obtaining structured data).

[0037] In an online AI education app, the following data are collected: first-dimensional unstructured data includes students' historical analysis reports on AI development in the course; second-dimensional unstructured data includes students' assignments on breaking down algorithm principles in the course; third-dimensional unstructured data includes students' text analyzing the technological evolution in the course; fourth-dimensional unstructured data includes subtitles of key segments of students' practical processes in the course, transcripts of course graduation defense recordings, and group project results documents; and fifth-dimensional unstructured data includes students' AI ethics scenario simulation decision-making explanations, tutor evaluation records, and classroom discussion speeches.

[0038] In this embodiment, structured data is recorded using a unified data table and associated with corresponding traceability information to ensure data consistency and traceability; unstructured data is stored using a unified plain text format, retaining complete semantic information while removing irrelevant format markers, which facilitates accurate and convenient semantic extraction in the future.

[0039] S22. Process the structured data contained in the dimensional data of each dimension to obtain the first numerical feature vector of standardized numerical representation. Extract semantics from the unstructured data contained in the dimensional data of each dimension and encode it into the second numerical feature vector. Concatenate the first numerical feature vector and the second numerical feature vector corresponding to each dimension to obtain the dimensional feature vector corresponding to each dimension. Concatenate the dimensional feature vectors corresponding to all dimensions to obtain the comprehensive feature vector. In one embodiment, the structured data contained in the dimensional data of each dimension is processed to obtain a first numerical feature vector with a standardized numerical representation, including: The structured data is mapped to a preset numerical range to obtain standard structured data; Identify missing data in the standard structured data, calculate the padding value for the missing data, and pad the missing data in the standard structured data according to the padding value to obtain the padded data; The filled data is subjected to outlier identification, the identified outliers are replaced according to preset rules to obtain standardized values, and the standardized values ​​are converted into a one-dimensional first numerical feature vector.

[0040] In one embodiment, the structured data is mapped to a preset numerical range to obtain standard structured data, including: Identify positive and negative indicators in the structured data; The positive index standardization formula using the extreme value standardization method maps the value of the positive index to a preset value range to obtain the positive standardized value; The negative index standardization formula using the extreme value standardization method maps the value of the negative index to a preset value range to obtain the negative standardized value; The structured data is scaled according to a preset ratio by combining the positive and negative standardized values.

[0041] In one embodiment, the preset numerical range can be the range [0.1, 0.9], and by mapping the data to the range [0.1, 0.9], the influence of extreme values ​​can be avoided.

[0042] In one embodiment, the positive indicator is an indicator whose larger value corresponds to a higher level (such as test score, accuracy rate), and the standardized formula for the positive indicator is: Where x is the original observation value of the positive structured index (such as students' "closed-book theoretical exam score" or "accuracy rate of practical tasks"). , These are the minimum and maximum values ​​of the positive structured index in the entire student sample, respectively. Let x be the positively standardized value.

[0043] In one embodiment, the negative indicator is an indicator whose smaller value corresponds to a higher level (such as completion time), and the standardized formula for the negative indicator is: Where x is the original observation value of the negative structured index, , These are the minimum and maximum values ​​of the negative structured index in the entire student sample, respectively. This is the negative standardized value corresponding to x.

[0044] In this embodiment, the maximum and minimum values ​​of each structured data indicator are calculated independently to preserve the original distribution characteristics of different indicators.

[0045] In one embodiment, missing data in the standard structured data is identified, imputation values ​​for the missing data are calculated, and the missing data in the standard structured data is imputed according to the imputation values ​​to obtain imputed data, including: Identify the missing data and the dimensions corresponding to the missing data in the standard structured data, calculate the correlation between the missing data and data in other dimensions, and obtain the associated data of other dimensions corresponding to the missing data based on the calculation results; The correlation coefficient between the missing data and data of the same dimension is calculated using a contextual adaptive imputation (AI) strategy, and the correlation coefficient between the missing data and related data of other dimensions is also calculated. Data with a correlation coefficient greater than or equal to 0.3 with the missing data from data in the same dimension and related data in other dimensions are selected as initial filling data. The filling value of the missing data is calculated by the initial filling data using a weighted average formula. The missing data is then filled with the filling value to obtain the filled data.

[0046] In one embodiment, outlier identification is performed on the filled data, and the identified outliers are replaced according to preset rules to obtain standardized values, including: The relative performance thresholding strategy is used to calculate the mean and standard deviation of each indicator in the populated data in the entire user behavior data sample, and to identify whether the value of each indicator is an outlier based on the preset threshold, mean and standard deviation. If the value of an indicator is an outlier, determine whether the outlier is higher than the preset standard value of the indicator; if yes, replace the outlier with the median of the indicator value corresponding to the sample with the indicator level of "excellent" in the entire user behavior data sample; if no, replace the outlier with the median of the indicator value corresponding to the sample with the indicator level of "basic" in the entire user behavior data sample.

[0047] In one embodiment, semantics are extracted from the unstructured data contained in the dimensional data of each dimension and encoded into a second numerical feature vector, including: The unstructured data is input into a preset large language model. Redundant information and meaningless content in the unstructured data are removed by the prompts of the large language model to obtain the core semantic text. The core semantic text is feature-encoded using a pre-defined artificial intelligence text embedding model to obtain a second numerical feature vector.

[0048] In one embodiment, the preset large language model is a natural language model. When using the natural language model to extract core semantics from the unstructured data, redundant information (such as repetitive statements and irrelevant background information) and meaningless content (such as formatting marks and typos) in the unstructured data are removed by setting prompts in the natural language model to obtain the core semantic text. This ensures that the extracted core semantic text focuses on the content related to the indicators of the course evaluation dimensions. For example, the prompts in the natural language model are: "Extract the core information related to {dimensional name} from the following text, including key knowledge points, ability performance, and viewpoints and attitudes, retain the original semantics, and remove redundant content: {text content}".

[0049] In one embodiment, the large text embedding model of the artificial intelligence can be the text-embedding-3-large model of Open AI. The core semantic text is feature-encoded by the large text embedding model of the artificial intelligence to obtain a 768-dimensional second numerical feature vector, and the value range of each dimension of the second numerical feature vector is [-1, 1].

[0050] In one embodiment, concatenating the first numerical feature vector and the second numerical feature vector corresponding to each dimension to obtain the dimension feature vector corresponding to each dimension includes: concatenating the first numerical feature vector corresponding to the dimension data of the same dimension to the end of the second numerical feature vector to obtain the dimension feature vector corresponding to the dimension data of that dimension.

[0051] For example, the structured data for the second dimension is the theoretical closed-book exam score, and the unstructured data is the algorithm principle decomposition assignment. The standardized values ​​of the theoretical closed-book exam score are used as an additional dimension vector and concatenated to the end of the 768-dimensional second numerical feature vector of the algorithm principle decomposition assignment, resulting in the 769-dimensional initial feature vector corresponding to the second dimension. This initial feature vector contains the quantized representation and semantic features of the second-dimensional data. Following the same operation for obtaining the initial feature vector of the second dimension, the initial feature vectors for the first, third, fourth, and fifth dimensions are obtained respectively.

[0052] In one embodiment, a comprehensive feature vector is obtained by concatenating the dimensional feature vectors corresponding to the dimensional data of all dimensions, including: concatenating the 769-dimensional feature vectors of the user's multiple dimensional data, and performing linear projection dimensionality reduction on the concatenated feature vector to obtain a 768-dimensional comprehensive feature vector.

[0053] This embodiment generates a unified-dimensional feature vector by standardizing, semantically extracting, and encoding multi-source data, laying the foundation for subsequent matching calculations and graph construction.

[0054] S23. Extract the sub-dimension feature vectors corresponding to each sub-dimension from the dimensional feature vectors corresponding to each dimension, and calculate the semantic similarity between each sub-dimension feature vector and the benchmark vector corresponding to each rating level. In one embodiment, before calculating the semantic similarity between the feature vectors of each sub-dimension and the baseline vector corresponding to each rating level, the method further includes: The first baseline vector of each sub-dimension is calculated based on the level-based standard indicator values ​​in the sample user behavior data of each sub-dimension contained in each dimension. The second baseline vector for each sub-dimension is calculated based on the standard indicator values ​​of good grades in the sample user behavior data of each sub-dimension included in each dimension. The third baseline vector for each sub-dimension is calculated based on the excellent standard indicator values ​​in the sample user behavior data of each sub-dimension included in each dimension.

[0055] In one embodiment, extracting the sub-dimension feature vectors corresponding to each sub-dimension from the dimensional feature vectors corresponding to each dimension includes: extracting vectors related to each sub-dimension contained in the dimension from the dimensional feature vectors corresponding to the dimensional data of each dimension, thereby obtaining the sub-dimension feature vectors of each sub-dimension.

[0056] In one embodiment, calculating the semantic similarity between the feature vector of each sub-dimension and the baseline vector corresponding to each rating level includes: Calculate the cosine similarity between each sub-dimension feature vector and the first reference vector of the same sub-dimension to obtain the first semantic similarity; Calculate the cosine similarity between each sub-dimension feature vector and the second reference vector of the same sub-dimension to obtain the second semantic similarity; The third semantic similarity is obtained by calculating the cosine similarity between the feature vectors of each sub-dimension and the third reference vector of the same sub-dimension.

[0057] For example: If the student's feature vector in sub-dimension a is v, the first reference vector is k1, the second reference vector is k2, and the third reference vector is k3, then the first semantic similarity sim1 is calculated as follows: sim1=(v·k1) / (‖v‖·‖k1‖); The second semantic similarity sim2 is calculated as follows: sim2=(v·k2) / (‖v‖·‖k2‖); Calculate the third semantic similarity sim3 as follows sim3=(v·k3) / (‖v‖·‖k3‖).

[0058] S24. Calculate the first user behavior score corresponding to each sub-dimension based on the multiple semantic similarities corresponding to each sub-dimension, and perform a weighted summation of the first user behavior scores corresponding to all sub-dimensions included in each dimension to obtain the second user behavior score corresponding to each dimension. In one embodiment, a first user behavior score corresponding to each sub-dimension is calculated based on multiple semantic similarities corresponding to each sub-dimension, and a second user behavior score corresponding to each dimension is obtained by weighted summation of the first user behavior scores corresponding to all sub-dimensions included in each dimension, including: Obtain the preset weight values ​​for each rating level of each sub-dimension as the first weight value of the semantic similarity corresponding to each rating level in each sub-dimension; Obtain the pre-set second weight value for the standardized numerical values ​​corresponding to each dimension; Based on the semantic similarity of each rating level in each sub-dimension and the first weight value, as well as the standardized value of the dimension to which each sub-dimension belongs and the second weight value, the first user behavior score corresponding to each sub-dimension is calculated using a weighted summation formula. Obtain the third weight value pre-set for each sub-dimension, and calculate the second user behavior score corresponding to each dimension using a weighted summation formula based on the first user behavior score of the multiple sub-dimensions contained in each dimension and the third weight value.

[0059] For example: the weight of the first semantic similarity sim1 corresponding to the basic level of sub-dimensional a is 0.2, the weight of the second semantic similarity sim2 corresponding to the good level is 0.3, the weight of the third semantic similarity sim3 corresponding to the excellent level is 0.5, and the weight of the standardized value n of the structured data corresponding to sub-dimensional a is 0.1. The first user behavior score S of sub-dimensional a is then calculated. a for S a =sim1·0.2+sim4·0.3+sim3·0.5+n·0.1 The second dimension comprises three sub-dimensions: b1, b2, and b3. The weight of b1 is 0.35, the weight of b2 is 0.45, and the weight of b3 is 0.2. The second user behavior score, score2, calculated from the data of this second dimension is... score2=b1·0.35+b2·0.45+b3·0.2 S25. Construct a dual-center heterogeneous bipartite graph as a user node graph, wherein the first center node embeds the comprehensive feature vector, the second center node embeds the average dimension reference vector, each first child node of the first center node embeds the second numerical feature vector of each dimension, each second child node of the second center node embeds the dimension reference vector of each dimension, and the edges include the connection edges between the first center node and the first child node, the connection edges between the second center node and the second child node, and the cross-side matching edges between the first child node and the second child node in the same dimension; In one embodiment, constructing a bicentric heterogeneous bipartite graph as a user node graph further includes: The second user behavior score for each dimension is used as the edge weight of the edge connecting the corresponding first child node and the first central node. Obtain the fourth weight value pre-set for each dimension as the edge weight of the edge connecting each second child node and the second center node; The cosine similarity between the second numerical feature vector of the first child node and the dimensional reference vector of the second child node in the same dimension is calculated as the edge weight of the edge connecting the first child node and the second child node in the same dimension.

[0060] This embodiment constructs a user node graph by calculating the matching degree between user features and the highest-level benchmark features, transforming user behavior sub-features into a graph structure, capturing the complex relationships between user features and between user features and the highest-level features, and ensuring the precision and interpretability of feature matching.

[0061] S26. The second numerical feature vector corresponding to each first child node of the user node graph is subjected to feature aggregation and feature optimization by a pre-trained fusion-aware graph neural network model. The fusion-aware graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the user node graph at different receptive field scales. After fusing the features output by the above modules, the third user behavior score corresponding to each first child node in each dimension is obtained by mapping through a fully connected layer. In this embodiment, as Figure 6 As shown, the fusion-aware graph neural network model includes a graph convolution module, a multi-head self-attention mechanism module, a multi-scale feature extraction module, and a fusion output module. The graph neural network module's model architecture includes an input layer, a first graph convolution layer, a second graph convolution layer, and a batch normalization layer. The input layer maps the feature vector dimensions of each node in the user node graph to a 256-dimensional vector through linear projection. The first graph convolution layer aggregates the features of the 256-dimensional vector using an adjacency matrix, resulting in a 128-dimensional vector. The second graph convolution layer aggregates the features of the 256-dimensional vector using an adjacency matrix, resulting in a 64-dimensional vector. The batch normalization layer is used as part of the model architecture to perform normalization on mini-batch samples during each training iteration.

[0062] In one embodiment, the training method for the graph neural network model includes: Obtain a sample dataset, which includes sample data and sample scores corresponding to the sample data, and divide the sample dataset into a training set and a validation set according to a preset ratio; The initial graph neural network model is repeatedly trained using the training set according to the preset training rounds. The model is validated using the validation set and the model loss is calculated using the weighted cross-entropy loss function. At the same time, the model parameters are optimized using the AdamW optimizer. Training is completed when the model loss meets the preset early stopping strategy, and the graph neural network model is obtained.

[0063] In this embodiment, the AdamW optimizer is pre-set with a weight decay coefficient λ = 1e-5 and a learning rate η = 1e-4. A dropout rate of 0.5 is preset in the fully connected layers of the initial graph neural network model to randomly zero out some neurons, preventing overfitting. The preset training epochs are 300, and the preset early stopping strategy has a patience value of 50, meaning that training stops when the model loss calculated from the validation set does not decrease for 50 consecutive epochs. The weighted cross-entropy loss function assigns a weight of 1.2 to samples with scores lower than or equal to a preset value (e.g., scores ≤ 60), and a weight of 1.0 to other samples, to balance the potentially imbalanced distribution of the evaluation data.

[0064] In one embodiment, a pre-trained fusion-aware graph neural network model is used to perform feature aggregation and feature optimization on the second numerical feature vector corresponding to each first child node of the user node graph. The fusion-aware graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the user node graph at different receptive field scales. After fusing the features output by the above modules, a third user behavior score corresponding to each first child node in each dimension is obtained through mapping via a fully connected layer, including: The graph neural network model constructs an adjacency matrix based on the nodes and edges of the user node graph. The rows and columns of the adjacency matrix represent the nodes in the user node graph, and the elements in the adjacency matrix indicate whether there are weighted edges between the nodes. The adjacency matrix is ​​then normalized. The graph convolutional layer performs local neighborhood feature aggregation on the normalized adjacency matrix to obtain the node feature matrix corresponding to each first child node of the user node graph. The dependency features of the user node graph are obtained by capturing the global temporal and long-distance dependencies between nodes in the node feature matrix through the multi-head self-attention mechanism module. The fine-grained and coarse-grained features of the node feature matrix are extracted through the multi-scale feature extraction module. The node features of the node feature matrix are optimized based on the dependency features and the fine-grained and coarse-grained features. The resulting optimized node features are then input into the fully connected layer and output layer of the graph neural network model for normalization and mapping to obtain the third user behavior score corresponding to each first child node.

[0065] This embodiment uses a graph neural network model to learn features from the user node graph, dynamically aggregates multi-dimensional information, and outputs optimized third-dimensional user behavior scores for each node.

[0066] S27. Construct a user behavior feature vector based on all third-party user behavior scores, match the corresponding user tags based on the user behavior feature vector, and generate an analysis report for the user based on the user tags and the third-party user behavior scores.

[0067] In one embodiment, a user behavior feature vector is constructed based on all third-party user behavior scores, and corresponding user tags are matched based on the user behavior feature vector, including: Construct an ordered sequence vector using third-party user behavior scores from multiple dimensions as components; The ordered sequence vector is mapped to the interval [0, 1] using a 0-1 normalization algorithm to obtain the user behavior feature vector; Calculate the Euclidean distance between the user behavior feature vector and the preset behavior feature cluster center, and take the user label corresponding to the cluster center with the smallest distance to the user behavior feature vector as the user label matched by the user behavior feature vector.

[0068] For example: the user node graph contains 5 dimension nodes, and the corresponding third user behavior scores are Q1, Q2, Q3, Q4, and Q5, respectively, constructing the corresponding ordered sequence vector Q. i ={Q1, Q2, Q3, Q4, Q5}; The user behavior feature vector Q is calculated using a 0-1 normalization algorithm. norm as follows Among them, Q min Q is the minimum value among the third user behavior scores corresponding to all dimension nodes. max This represents the maximum value among the third user behavior scores corresponding to all dimension nodes.

[0069] In one embodiment, the preset behavioral feature cluster centers are five cluster centers obtained by iteratively clustering the standardized behavioral feature vectors corresponding to all sample users using the K-means clustering algorithm based on the silhouette coefficient method. The silhouette coefficient method predetermines the optimal number of clusters K=5; the iterative clustering uses the K-means++ algorithm for initial clustering to avoid local optima, and sets the iteration termination condition as "the change in cluster centers is less than 1e-4" or "the maximum number of iterations reaches 1000". The distance metric of the clustering algorithm is calculated using Euclidean distance, with the formula: in , These are the standardized behavioral feature vectors for two users.

[0070] In one embodiment, the five cluster centers are assigned corresponding user tags and core feature descriptions based on the distribution characteristics of the indicator values. For example: The distribution characteristics of the index values ​​of the first cluster center are that the third user behavior score corresponding to the fourth dimension node is greater than or equal to 0.9, and the third user behavior score corresponding to the other dimension nodes is greater than or equal to 0.6. The corresponding user tag is "practical expertise type", and the core feature is described as "outstanding practical application ability and strong tool operation and scenario adaptation ability". The distribution characteristics of the index values ​​of the second cluster center are that the third user behavior score corresponding to the second dimension node is greater than or equal to 0.9, the third user behavior score corresponding to the first dimension node is greater than or equal to 0.85, and the third user behavior score corresponding to the other dimension nodes is greater than or equal to 0.6. The corresponding user tag is "theoretical solid type", and the core feature is described as "solid understanding of artificial intelligence theory and history, and in-depth understanding of algorithm logic". The distribution characteristics of the index values ​​of the third cluster center are that the third user behavior scores corresponding to all dimension nodes are in the range of [0.75, 0.85], the corresponding user label is "comprehensive and balanced", and the core feature is described as "balanced ability in all dimensions with no obvious weaknesses". The distribution characteristics of the index values ​​of the fourth cluster center are that the third user behavior score corresponding to the fifth dimension node is greater than or equal to 0.9, and the third user behavior score corresponding to the other dimension nodes is greater than or equal to 0.6. The corresponding user label is "ethics-oriented", and the core feature is described as "prominent identification of ethical risks and compliance awareness, and clear understanding of social impact". The distribution characteristics of the index values ​​of the fifth cluster center are that the third user behavior score corresponding to at least two dimension nodes is less than or equal to 0.6, and the third user behavior score corresponding to other dimension nodes is in the numerical range of [0.6, 0.75]. The corresponding user label is "potential improvement type", and the core feature is described as "there are obvious shortcomings in ability, which need to be improved in a targeted manner".

[0071] In one embodiment, generating an analysis report for the user based on the user tags and the third user behavior score includes: Get the pre-set analysis report template; Obtain the corresponding tag interpretation text based on the user tags and write it into the preset tag interpretation area in the analysis report template; Obtain the third-party user behavior scores corresponding to all dimensions and write them into the preset dimension display area in the analysis report template; The user rating is generated based on the third user behavior score corresponding to all dimensions, and written into the preset rating analysis area in the analysis report template.

[0072] In one embodiment, the analysis report also includes the user's unique identifier (ID), the distance between the user's behavior feature vector and the cluster center of the behavior features corresponding to the user's tag, and the matching degree, etc.

[0073] In one embodiment, a user rating is generated based on the third user behavior score corresponding to all dimensions, and written into a preset rating analysis area in the analysis report template, including: The dimension corresponding to the third user behavior score that is higher than the first score threshold is used as the dominant dimension for user rating; Calculate the similarity of the second numerical feature vector corresponding to the dominant dimension with the dominant dimension similarity of the highest-level benchmark vector of the same dimension; The dimension corresponding to the third user behavior score that is below the second score threshold is taken as the weak dimension of user rating; Calculate the similarity of the second numerical feature vector corresponding to the shortest dimension with the shortest-level benchmark vector of the same dimension. The strengths and their similarity, as well as the weaknesses and their similarity, are written into the preset rating analysis area in the analysis report template.

[0074] For example, in an online AI education app, a student's analysis report is as follows: "Student ID: 2024001, Ability Tag: Practical Expertise (Distance to the cluster center of ability feature is 0.12, matching degree is 92%)."

[0075] I. Tag Characteristics Interpretation: This tag corresponds to students with outstanding practical application abilities. Their core strengths are proficiency in operating artificial intelligence tools, accuracy in scenario adaptation, and innovativeness of results, which aligns with the core training objective of the general education course on artificial intelligence: "practice-oriented".

[0076] II. Dimensional Performance Details (GNN Output Final Score): First Dimension (78 points), Second Dimension (58 points), Third Dimension (82 points), Fourth Dimension (92 points), Fifth Dimension (86 points).

[0077] III. Analysis of Strengths and Weaknesses: The strengths are in the fourth dimension (practical application ability of artificial intelligence, 92 points) and the fifth dimension (ethical and moral cognition of artificial intelligence, 86 points), with similarities of 0.89 and 0.87 respectively to the excellent level benchmark vector, reaching the excellent level standard; the weakness is in the second dimension (theoretical foundation of artificial intelligence, 58 points), with a similarity of only 0.61 to the basic level benchmark vector, failing to meet the basic requirements. The core weakness is insufficient understanding of core algorithm logic.

[0078] IV. Personalized Improvement Suggestions: 1. Take a targeted elective course on "Introduction to Core Algorithms of Artificial Intelligence," focusing on mastering the fundamental principles of algorithms such as neural networks and decision trees; 2. Deepen theoretical learning through practical case studies, such as simultaneously reviewing the application logic of algorithms during practical tasks; 3. Participate in theoretical discussion groups to improve the ability to express and analyze algorithm principles. In one embodiment, after generating the user's analysis report based on the user tags and the third user behavior score, the method further includes: The system periodically acquires new behavioral data generated by the user in the application APP based on a preset time period. The graph neural network model is updated based on the new behavioral data of all users in the application APP; The user's analysis report is updated using the updated graph neural network model based on the user's new behavioral data.

[0079] For example, in an online AI education app, the system periodically acquires students' dimensional data for the new semester, updates the graph neural network model based on all students' dimensional data for the new semester, and then updates the students' analysis reports based on the updated graph neural network model.

[0080] This embodiment uses a clustering algorithm to classify users based on third-party user behavior scores across various dimensions and generates unique tags. It then combines the tags with dimensional features to generate a personalized analysis report, providing clear guidance for application usage feedback and user development.

[0081] As can be seen from the above embodiments, the user behavior analysis method based on graph neural networks proposed in this invention processes structured and unstructured data generated by users in an application (APP) across multiple dimensions to obtain dimensional feature vectors corresponding to each dimension. It calculates a second user behavior score for each dimension based on the similarity between the feature vectors and the baseline vector. A user node graph is constructed based on the dimensional feature vectors and the corresponding second user behavior scores. A third user behavior score corresponding to each dimension is generated based on the user node graph using a graph neural network model, and corresponding user tags are matched. An analysis report is generated based on the user tags and the third user behavior score. This achieves standardized processing and deep semantic fusion of multi-source heterogeneous data. The node graph explicitly models the semantic and quantitative associations of dimensional data, and the graph neural network dynamically aggregates information from the graph to optimize and calibrate the second user behavior score, outputting a highly interpretable user behavior analysis report.

[0082] like Figure 4The diagram shown is a schematic representation of an embodiment of the computer device of the present invention. The computer device 1 is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions. The computer device 1 can be a single network server, a server group composed of multiple network servers, or a cloud based on cloud computing consisting of a large number of hosts or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers.

[0083] In this embodiment, the computer device 1 includes, but is not limited to, a memory 11, a processor 12, and a network interface 13 that can be interconnected via a system bus. The memory 11 stores a user behavior analysis program 10 based on a graph neural network, which can be executed by the processor 12. Figure 4 Only computer device 1, which includes components 11-13 and a user behavior analysis program 10 based on a graph neural network, is shown. Those skilled in the art will understand that... Figure 4 The structure shown does not constitute a limitation on computer device 1 and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0084] The memory 11 includes RAM and at least one type of readable storage medium. The RAM provides a cache for the operation of the computer device 1; the readable storage medium can be volatile or non-volatile. Specifically, the readable storage medium can be a storage medium such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the readable storage medium can be an internal storage unit of the computer device 1, such as the hard disk of the computer device 1; in other embodiments, the storage medium can also be an external storage device of the computer device 1, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 1. In this embodiment, the readable storage medium of the memory 11 mainly includes a program storage area and a data storage area. The program storage area is typically used to store the operating system and various application software installed on the computer device 1, such as the code of the user behavior analysis program 10 based on graph neural network in one embodiment of the present invention. The data storage area can store data created based on the use of blockchain nodes, such as various types of data that have been output or will be output.

[0085] In some embodiments, processor 12 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. This processor 12 is typically used to control the overall operation of the computer device 1, such as performing control and processing related to data interaction or communication with other devices. In this embodiment, the processor 12 is used to run program code stored in the memory 11 or process data, for example, running a user behavior analysis program 10 based on a graph neural network.

[0086] The network interface 13 may include a wireless network interface or a wired network interface, which is used to establish a communication connection between the computer device 1 and a server or other terminal (not shown in the figure).

[0087] Optionally, the computer device 1 may further include a user interface, which may include a display, an input unit such as a keyboard, and optionally, a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an organic light-emitting diode (OLED) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the computer device 1 and to display a visual user interface.

[0088] In one embodiment of the present invention, when the user behavior analysis program 10 based on graph neural network is executed by the processor 12, the following steps S21-S27 are implemented.

[0089] S21. Based on multiple dimensions pre-set for the application APP, extract the dimension data corresponding to each dimension from the user behavior data generated by the user in the application APP. The dimension data includes structured data and unstructured data. Each dimension includes multiple sub-dimensions. The user behavior data of each sub-dimension is set with multiple rating levels. Each rating level corresponds to a baseline vector. S22. Process the structured data contained in the dimensional data of each dimension to obtain the first numerical feature vector of standardized numerical representation. Extract semantics from the unstructured data contained in the dimensional data of each dimension and encode it into the second numerical feature vector. Concatenate the first numerical feature vector and the second numerical feature vector corresponding to each dimension to obtain the dimensional feature vector corresponding to each dimension. Concatenate the dimensional feature vectors corresponding to all dimensions to obtain the comprehensive feature vector. S23. Extract the sub-dimension feature vectors corresponding to each sub-dimension from the dimensional feature vectors corresponding to each dimension, and calculate the semantic similarity between each sub-dimension feature vector and the benchmark vector corresponding to each rating level. S24. Calculate the first user behavior score corresponding to each sub-dimension based on the multiple semantic similarities corresponding to each sub-dimension, and perform a weighted summation of the first user behavior scores corresponding to all sub-dimensions included in each dimension to obtain the second user behavior score corresponding to each dimension. S25. Construct a dual-center heterogeneous bipartite graph as a user node graph, wherein the first center node embeds the comprehensive feature vector, the second center node embeds the average dimension reference vector, each first child node of the first center node embeds the second numerical feature vector of each dimension, each second child node of the second center node embeds the dimension reference vector of each dimension, and the edges include the connection edges between the first center node and the first child node, the connection edges between the second center node and the second child node, and the cross-side matching edges between the first child node and the second child node in the same dimension; S26. The second numerical feature vector corresponding to each first child node of the user node graph is subjected to feature aggregation and feature optimization by a pre-trained fusion-aware graph neural network model. The fusion-aware graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the user node graph at different receptive field scales. After fusing the features output by the above modules, the third user behavior score corresponding to each first child node in each dimension is obtained by mapping through a fully connected layer. S27. Construct a user behavior feature vector based on all third-party user behavior scores, match the corresponding user tags based on the user behavior feature vector, and generate an analysis report for the user based on the user tags and the third-party user behavior scores.

[0090] The specific operation steps implemented in steps S21-S27 above are largely the same as those in an embodiment of the user behavior analysis method based on graph neural networks of the present invention, and will not be repeated here.

[0091] In other embodiments, the graph neural network-based user behavior analysis program 10 may be divided into one or more modules, which are stored in the memory 11 and executed by one or more processors (processor 12 in this embodiment) to complete the present invention. The module referred to in the present invention refers to a series of computer program instruction segments that can perform specific functions and are used to describe the execution process of the graph neural network-based user behavior analysis program 10 in the terminal 1.

[0092] like Figure 5The diagram shown is a schematic representation of a module of a user behavior analysis device based on a graph neural network according to the present invention.

[0093] In a first embodiment of the present invention, a user behavior analysis device 100 based on a graph neural network includes a data acquisition module 110, a data processing module 120, a similarity calculation module 130, a first analysis module 140, a node graph generation module 150, a second analysis module 160, and an analysis report generation module 170, exemplarily: The data acquisition module 110 is used to extract dimension data corresponding to each dimension from the user behavior data generated by the user in the application APP according to multiple dimensions pre-set for the application APP. The dimension data includes structured data and unstructured data. Each dimension includes multiple sub-dimensions. The user behavior data of each sub-dimension is set with multiple rating levels. Each rating level corresponds to a baseline vector. The data processing module 120 is used to process the structured data contained in the dimensional data of each dimension to obtain a first numerical feature vector with standardized numerical representation, extract semantics from the unstructured data contained in the dimensional data of each dimension and encode it into a second numerical feature vector, concatenate the first numerical feature vector and the second numerical feature vector corresponding to each dimension to obtain the dimensional feature vector corresponding to each dimension, and concatenate the dimensional feature vectors corresponding to all dimensions to obtain a comprehensive feature vector. The similarity calculation module 130 is used to extract the sub-dimension feature vectors corresponding to each sub-dimension from the dimensional feature vectors corresponding to each dimension, and calculate the semantic similarity between each sub-dimension feature vector and the benchmark vector corresponding to each rating level. The first analysis module 140 is used to calculate the first user behavior score corresponding to each sub-dimension based on multiple semantic similarities corresponding to each sub-dimension, and to perform a weighted summation of the first user behavior scores corresponding to all sub-dimensions included in each dimension to obtain the second user behavior score corresponding to each dimension. The node graph generation module 150 is used to construct a dual-center heterogeneous bipartite graph as a user node graph, wherein the first center node embeds the comprehensive feature vector, the second center node embeds the average dimension reference vector, each first child node of the first center node embeds the second numerical feature vector of each dimension, each second child node of the second center node embeds the dimension reference vector of each dimension, and the edges include the connection edges between the first center node and the first child node, the connection edges between the second center node and the second child node, and the cross-side matching edges between the first child node and the second child node in the same dimension; The second analysis module 160 is used to perform feature aggregation and feature optimization on the second numerical feature vector corresponding to each first child node of the user node graph through a pre-trained fusion-aware graph neural network model. The fusion-aware graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the user node graph at different receptive field scales. After fusing the features output by the above modules, the third user behavior score corresponding to each first child node in each dimension is obtained by mapping through a fully connected layer. The analysis report generation module 170 is used to construct a user behavior feature vector based on all third user behavior scores, match the corresponding user tags based on the user behavior feature vector, and generate an analysis report for the user based on the user tags and the third user behavior scores.

[0094] The specific operation steps implemented when the above-mentioned data acquisition module 110, data processing module 120, similarity calculation module 130, first analysis module 140, node graph generation module 150, second analysis module 160 and analysis report generation module 170 are executed are generally the same as those in the above embodiments, and will not be repeated here.

[0095] Furthermore, this invention also proposes a computer-readable storage medium, which can be volatile or non-volatile. Specifically, the computer-readable storage medium can be any one or any combination of several of the following: hard disk, multimedia card, SD card, flash memory card, SMC, read-only memory (ROM), erasable programmable read-only memory (EPROM), portable compact disc read-only memory (CD-ROM), USB memory, etc. The computer-readable storage medium stores a user behavior analysis program 10 based on a graph neural network.

[0096] When the user behavior analysis program 10 based on graph neural network is executed by the processor, it performs the following operations: A1. Based on multiple dimensions pre-set for the application APP, extract the dimension data corresponding to each dimension from the user behavior data generated by the user in the application APP. The dimension data includes structured data and unstructured data. Each dimension includes multiple sub-dimensions. The user behavior data of each sub-dimension is set with multiple rating levels. Each rating level corresponds to a baseline vector. A2. Process the structured data contained in the dimensional data of each dimension to obtain the first numerical feature vector with standardized numerical representation. Extract semantics from the unstructured data contained in the dimensional data of each dimension and encode it into the second numerical feature vector. Concatenate the first numerical feature vector and the second numerical feature vector corresponding to each dimension to obtain the dimensional feature vector corresponding to each dimension. Concatenate the dimensional feature vectors corresponding to all dimensions to obtain the comprehensive feature vector. A3. Extract the sub-dimension feature vectors corresponding to each sub-dimension from the dimensional feature vectors corresponding to each dimension, and calculate the semantic similarity between each sub-dimension feature vector and the benchmark vector corresponding to each rating level. A4. Calculate the first user behavior score for each sub-dimension based on the multiple semantic similarities corresponding to each sub-dimension, and then perform a weighted summation of the first user behavior scores for all sub-dimensions included in each dimension to obtain the second user behavior score for each dimension. A5. Construct a dual-center heterogeneous bipartite graph as the user node graph, wherein the first center node embeds the comprehensive feature vector, the second center node embeds the average dimension reference vector, each first child node of the first center node embeds the second numerical feature vector of each dimension, each second child node of the second center node embeds the dimension reference vector of each dimension, and the edges include the connection edges between the first center node and the first child node, the connection edges between the second center node and the second child node, and the cross-side matching edges between the first child node and the second child node in the same dimension; A6. The second numerical feature vector corresponding to each first child node of the user node graph is aggregated and optimized by a pre-trained fusion-aware graph neural network model. The fusion-aware graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the user node graph at different receptive field scales. After fusing the features output by the above modules, the third user behavior score corresponding to each first child node in each dimension is obtained by mapping through a fully connected layer. A7. Construct a user behavior feature vector based on all third-party user behavior scores, match the corresponding user tags based on the user behavior feature vector, and generate an analysis report for the user based on the user tags and the third-party user behavior scores.

[0097] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0098] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0099] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0100] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A comprehensive evaluation method for college students' AI competence and literacy that integrates perceptual improved graph neural networks, characterized in that, The method includes: Based on the multiple dimensions preset in the artificial intelligence course, the dimensional data corresponding to each dimension of the student in the artificial intelligence course is obtained. The dimensional data includes structured data and unstructured data. Each dimension includes multiple sub-dimensions. The data of each sub-dimension is set with multiple rating levels. Each rating level corresponds to a benchmark vector. The structured data contained in the dimensional data of each dimension is processed to obtain the first numerical feature vector with standardized numerical representation. The semantics of the unstructured data contained in the dimensional data of each dimension are extracted and encoded into the second numerical feature vector. The first numerical feature vector and the second numerical feature vector corresponding to each dimension are concatenated to obtain the dimensional feature vector corresponding to each dimension. The dimensional feature vectors corresponding to all dimensions are concatenated to obtain the comprehensive feature vector. Extract the sub-dimensional feature vectors corresponding to each sub-dimension from the dimensional feature vectors corresponding to each dimension, and calculate the semantic similarity between each sub-dimensional feature vector and the benchmark vector corresponding to each rating level. Calculate the first data score for each sub-dimension based on multiple semantic similarities for each sub-dimension, and then sum the weighted first data scores for all sub-dimensions included in each dimension to obtain the second data score for each dimension. Extract the baseline vector corresponding to the highest rating level of each sub-dimension, summarize the baseline vectors corresponding to the highest rating levels of all sub-dimensions included in each dimension to obtain the dimension baseline vector of each dimension, and calculate the average dimension baseline vector by averaging the dimension baseline vectors of all dimensions. A bicentric heterogeneous bipartite graph is constructed as the student node graph, wherein the first central node is embedded with the comprehensive feature vector, the second central node is embedded with the average dimension reference vector, each first child node of the first central node is embedded with the second numerical feature vector of each dimension, each second child node of the second central node is embedded with the dimension reference vector of each dimension, and the edges include the connection edges between the first central node and the first child node, the connection edges between the second central node and the second child node, and the cross-side matching edges between the first child node and the second child node in the same dimension; The second numerical feature vector corresponding to each first child node of the student node graph is aggregated and optimized by a pre-trained fusion-sensory graph neural network model. The fusion-sensory graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the dual-center heterogeneous bipartite graph at different receptive field scales. After fusing the features output by the above modules, the third data score of each first child node corresponding to each dimension is obtained by mapping through a fully connected layer. A student ability feature vector is constructed based on all third-party data scores. The corresponding ability labels are matched based on the student ability feature vector. An AI ability analysis report of the student is generated based on the ability labels and the third-party data scores.

2. A computer device, the computer device comprising: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the comprehensive evaluation method for college students' AI competence and literacy based on the fusion perception improved graph neural network as described in claim 1.

3. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the comprehensive evaluation method for college students' AI ability and literacy based on the fusion perception improved graph neural network as described in claim 1.

4. A user behavior analysis method based on graph neural networks, characterized in that, The method includes: Based on multiple dimensions pre-set for the application APP, dimension data corresponding to each dimension is extracted from user behavior data generated by the user in the application APP. The dimension data includes structured data and unstructured data. Each dimension includes multiple sub-dimensions. The user behavior data of each sub-dimension is set with multiple rating levels. Each rating level corresponds to a baseline vector. The structured data contained in the dimensional data of each dimension is processed to obtain the first numerical feature vector with standardized numerical representation. The semantics of the unstructured data contained in the dimensional data of each dimension are extracted and encoded into the second numerical feature vector. The first numerical feature vector and the second numerical feature vector corresponding to each dimension are concatenated to obtain the dimensional feature vector corresponding to each dimension. The dimensional feature vectors corresponding to all dimensions are concatenated to obtain the comprehensive feature vector. Extract the sub-dimensional feature vectors corresponding to each sub-dimension from the dimensional feature vectors corresponding to each dimension, and calculate the semantic similarity between each sub-dimensional feature vector and the benchmark vector corresponding to each rating level. Calculate the first user behavior score for each sub-dimension based on multiple semantic similarities for each sub-dimension, and then sum the weighted scores of the first user behavior scores for all sub-dimensions included in each dimension to obtain the second user behavior score for each dimension. A dual-center heterogeneous bipartite graph is constructed as the user node graph, wherein the first center node embeds the comprehensive feature vector, the second center node embeds the average dimension reference vector, each first child node of the first center node embeds the second numerical feature vector of each dimension, each second child node of the second center node embeds the dimension reference vector of each dimension, and the edges include the connection edges between the first center node and the first child node, the connection edges between the second center node and the second child node, and the cross-side matching edges between the first child node and the second child node in the same dimension; The second numerical feature vector corresponding to each first child node of the user node graph is subjected to feature aggregation and feature optimization by a pre-trained fusion-aware graph neural network model. The fusion-aware graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the user node graph at different receptive field scales. After fusing the features output by the above modules, the third user behavior score corresponding to each first child node in each dimension is obtained by mapping through a fully connected layer. A user behavior feature vector is constructed based on all third-party user behavior scores. The corresponding user tags are matched based on the user behavior feature vectors. An analysis report for the user is generated based on the user tags and the third-party user behavior scores.

5. The user behavior analysis method based on graph neural networks as described in claim 4, characterized in that, The process of processing the structured data contained in the dimensional data of each dimension to obtain the first numerical feature vector with standardized numerical representation includes: The structured data is mapped to a preset numerical range to obtain standard structured data; Identify missing data in the standard structured data, calculate the padding value for the missing data, and pad the missing data in the standard structured data according to the padding value to obtain the padded data; The filled data is subjected to outlier identification, the identified outliers are replaced according to preset rules to obtain standardized values, and the standardized values ​​are converted into a one-dimensional first numerical feature vector.

6. The user behavior analysis method based on graph neural networks as described in claim 4, characterized in that, The process of extracting semantics from the unstructured data contained in the dimensional data of each dimension and encoding it into a second numerical feature vector includes: The unstructured data is input into a preset large language model. Redundant information and meaningless content in the unstructured data are removed by the prompts of the large language model to obtain the core semantic text. The core semantic text is feature-encoded using a pre-defined artificial intelligence text embedding model to obtain a second numerical feature vector.

7. The user behavior analysis method based on graph neural networks as described in claim 4, characterized in that, The step of calculating the first user behavior score corresponding to each sub-dimension based on multiple semantic similarities for each sub-dimension, and then weighting and summing the first user behavior scores corresponding to all sub-dimensions included in each dimension to obtain the second user behavior score corresponding to each dimension, includes: Obtain the preset weight values ​​for each rating level of each sub-dimension as the first weight value of the semantic similarity corresponding to each rating level in each sub-dimension; Obtain the pre-set second weight value for the standardized numerical values ​​corresponding to each dimension; Based on the semantic similarity of each rating level in each sub-dimension and the first weight value, as well as the standardized value of the dimension to which each sub-dimension belongs and the second weight value, the first user behavior score corresponding to each sub-dimension is calculated using a weighted summation formula. Obtain the third weight value pre-set for each sub-dimension, and calculate the second user behavior score corresponding to each dimension using a weighted summation formula based on the first user behavior score of the multiple sub-dimensions contained in each dimension and the third weight value.

8. The user behavior analysis method based on graph neural networks as described in claim 4, characterized in that, The construction of a bicentric heterogeneous bipartite graph as a user node graph also includes: The second user behavior score for each dimension is used as the edge weight of the edge connecting the corresponding first child node and the first central node. Obtain the fourth weight value pre-set for each dimension as the edge weight of the edge connecting each second child node and the second center node; The cosine similarity between the second numerical feature vector of the first child node and the dimensional reference vector of the second child node in the same dimension is calculated as the edge weight of the edge connecting the first child node and the second child node in the same dimension.

9. The user behavior analysis method based on graph neural networks as described in claim 4, characterized in that, The pre-trained fusion-aware graph neural network model performs feature aggregation and optimization on the second numerical feature vector corresponding to each first child node of the user node graph. The fusion-aware graph neural network model includes a graph convolution module for local neighborhood feature aggregation, a multi-head self-attention mechanism module for capturing global temporal and long-distance dependencies between nodes, and a multi-scale feature extraction module for simultaneously extracting fine-grained and coarse-grained features of the user node graph at different receptive field scales. After fusing the features output by the above modules, a third user behavior score corresponding to each first child node in each dimension is obtained through mapping via a fully connected layer, including: The graph neural network model constructs an adjacency matrix based on the nodes and edges of the user node graph. The rows and columns of the adjacency matrix represent the nodes in the user node graph, and the elements in the adjacency matrix indicate whether there are weighted edges between the nodes. The adjacency matrix is ​​then normalized. The graph convolutional layer performs local neighborhood feature aggregation on the normalized adjacency matrix to obtain the node feature matrix corresponding to each first child node of the user node graph. The dependency features of the user node graph are obtained by capturing the global temporal and long-distance dependencies between nodes in the node feature matrix through the multi-head self-attention mechanism module. The fine-grained and coarse-grained features of the node feature matrix are extracted through the multi-scale feature extraction module. The node features of the node feature matrix are optimized based on the dependency features and the fine-grained and coarse-grained features. The resulting optimized node features are then input into the fully connected layer and output layer of the graph neural network model for normalization and mapping to obtain the third user behavior score corresponding to each first child node.

10. The user behavior analysis method based on graph neural networks as described in claim 4, characterized in that, The process of generating an analysis report for the user based on the user tags and the third user behavior score includes: Get the pre-set analysis report template; Obtain the corresponding tag interpretation text based on the user tags and write it into the preset tag interpretation area in the analysis report template; Obtain the third-party user behavior scores corresponding to all dimensions and write them into the preset dimension display area in the analysis report template; The user rating is generated based on the third user behavior score corresponding to all dimensions, and written into the preset rating analysis area in the analysis report template.

Images