Intelligent Recommendation Methods and Systems for Computer Information Teaching Resources

Abstract

This application provides an intelligent recommendation method and system for computer information teaching resources. First, a recommendation graph is constructed, comprising user nodes, resource nodes, and at least one behavioral interaction relationship. Then, the nodes in the recommendation graph are vectorized to generate user preference vectors and resource feature vectors. Contextual information from the current learning task is integrated, and the contribution weights of different interaction paths in the recommendation graph to the recommendation decision are calculated using a graph attention network. The preference drift between the user's current cognitive state and their historical cognitive state is periodically monitored. When the preference drift exceeds a preset threshold, a weight reallocation mechanism in the graph attention network is triggered, and a dynamically adjusted list of recommended teaching resources is generated based on the matching degree between the updated user preference vector and the resource feature vector. Using the scheme of this application, dynamic reallocation of graph attention based on preference drift can be achieved, thereby improving the adaptive accuracy of teaching resource recommendations.

Description

Technical Field

[0001] This application relates to the field of resource recommendation technology, and more specifically, to a method and system for intelligent recommendation of computer information teaching resources. Background Technology

[0002] With the development of information technology, computer-based teaching platforms have been widely used in universities, vocational education, and online learning environments. The number and types of teaching resources are growing rapidly, including course videos, experimental guides, exercise sets, and online tests. Against this backdrop, how to recommend the most suitable teaching resources based on students' learning behaviors and knowledge needs has become a key technical issue for improving teaching efficiency and personalized learning experiences. Effective resource recommendations can not only save learners time in searching for resources, but also guide learners to efficiently advance the learning process according to knowledge logic, thereby improving learning outcomes and knowledge mastery. Thus, it has significant value in the application of educational informatization and intelligentization.

[0003] Existing computer information teaching resource recommendation technologies largely rely on user historical behavior records or static user profiles to generate recommendation lists. While these can reflect learning preferences to some extent, they struggle to effectively address the dynamic changes in learners' cognitive levels and interests at different stages. Specifically, when a user's short-term learning focus shifts from their long-term learning preferences, existing recommendation models cannot adjust their strategies in a timely manner. This leads to a decreased match between recommended resources and the user's current learning needs, resulting in delayed recommendations and even resource redundancy or omissions of key knowledge points. This problem is particularly pronounced when learning tasks are complex and resource types are diverse, severely hindering the improvement of personalized recommendations and the optimization of teaching quality. Therefore, how to achieve dynamic reallocation of graph attention based on preference drift to improve the adaptive accuracy of teaching resource recommendations has become a challenging issue for the industry. Summary of the Invention

[0004] This application provides a method and system for intelligent recommendation of computer information teaching resources, which can realize dynamic reallocation of graph attention based on preference drift, thereby improving the adaptive accuracy of teaching resource recommendation.

[0005] Firstly, this application provides an intelligent recommendation method for computer information teaching resources, including: Collect user learning behavior data and teaching resource attribute data in the computer information teaching platform, and construct a recommendation graph that includes user nodes, resource nodes and at least one behavioral interaction relationship; Using graph embedding learning methods, the nodes in the recommendation graph are vectorized to generate user preference representation vectors and resource feature representation vectors; By integrating the contextual information of the current learning task, the contribution weights of different interaction paths in the recommendation graph to the recommendation decision are calculated through a graph attention network, thereby dynamically adjusting the user preference representation vector. Periodically monitor the amount of preference drift between the user's current cognitive state and historical cognitive state; When the detected preference drift exceeds a preset threshold, the weight redistribution mechanism in the graph attention network is triggered, and a dynamically adjusted list of recommended teaching resources is generated based on the matching degree between the updated user preference representation vector and the resource feature representation vector.

[0006] Preferably, constructing a recommendation graph that includes user nodes, resource nodes, and at least one behavioral interaction relationship specifically includes: Each user in the computer information teaching platform is uniquely identified, and a corresponding user node is generated. Based on the attribute data of teaching resources, the knowledge point tags, difficulty levels and resource types of teaching resources in the computer information teaching platform are structured and encoded to generate resource nodes; Learning behavior data is divided according to the time dimension, and long-term interactive behavior features that represent users' long-term learning preferences and short-term interactive behavior features that represent recent changes in learning focus are extracted. Based on the long-term and short-term interaction behavior characteristics, the interaction strength between user nodes and resource nodes is calculated, and at least one behavioral interaction relationship edge is determined based on the interaction strength. A recommendation graph is constructed based on the user nodes, resource nodes, and behavioral interaction relationships.

[0007] Preferably, the nodes in the recommendation graph are vectorized using a graph embedding learning method to generate user preference representation vectors and resource feature representation vectors, specifically including: Load corresponding attribute features, including user history behavior, resource knowledge point tags, difficulty level and resource type, into user nodes and resource nodes in the recommendation graph as initial node features; Based on the behavioral interaction relationships between nodes, the initial node features of adjacent nodes are aggregated, and the node representation is updated by combining the node features of the node itself, thus obtaining the node vector of the corresponding node. Repeated aggregation and multiple rounds of updating the node's own features enable the node vector to contain both its own attribute information and neighborhood structure information. The updated node vector of the user node is output as the user preference representation vector, and the node vector of the resource node is output as the resource feature representation vector.

[0008] Preferably, the contribution weights of different interaction paths in the recommendation graph to the recommendation decision are calculated using a graph attention network by integrating contextual information from the current learning task. Specifically, this includes: The knowledge point tags, difficulty level, and learning stage information associated with the current learning task are vectorized and encoded to form a task context vector; Extract the meta-path features corresponding to various interaction relationship paths in the recommendation graph, and jointly represent the meta-path features with the task context vector; The input graph attention network will be jointly represented, and the relative importance weights of different interaction paths in the current learning task will be evaluated in parallel through a multi-head attention structure. The relative importance weights corresponding to each interaction path are normalized to obtain the contribution weights of different interaction paths in the recommendation graph to the recommendation decision.

[0009] Preferably, periodically monitoring the preference drift between a user's current cognitive state and their historical cognitive state specifically includes: Extract the latest user preference representation vector as an approximate representation of the current cognitive state; The user's historical cognitive state benchmark is determined based on the cognitive state obtained from historical periodic monitoring. The difference between the current cognitive state and the historical cognitive state benchmark is calculated to obtain the preference difference value, and the preference difference value is used as the preference drift amount.

[0010] Preferably, the dynamically adjusted list of recommended teaching resources, generated based on the matching degree between the updated user preference representation vector and the resource feature representation vector, specifically includes: Calculate the similarity measure between the updated user preference representation vector and the resource feature representation vector of the candidate teaching resources to obtain the initial matching result; The initial matching degree result is corrected by combining the knowledge point connection relationship between teaching resources and the contextual constraints of the current learning task; The candidate teaching resources are ranked based on the corrected matching results; A list of recommended teaching resources is generated based on the sorting results.

[0011] Preferably, the preset threshold is an adaptive threshold, which is dynamically updated based on the statistical distribution of the target user's preference drift over multiple historical monitoring periods, so that the preference drift triggering conditions match the fluctuation range of the user's individual learning behavior.

[0012] Preferably, the learning behavior data includes operation event data collected by the computer information teaching platform during the user's learning process. The operation event data includes at least the behavior type, occurrence time, and associated resource identifier corresponding to the user's access to, interaction with, and completion of teaching resources, which is used to characterize the user's behavioral features during the learning process.

[0013] Preferably, the teaching resource attribute data is descriptive information that provides a structured representation of the content characteristics and organizational structure characteristics of the teaching resource itself. The descriptive information is used at least to characterize the knowledge content attributes, difficulty level, and resource type of the teaching resource.

[0014] Secondly, this application provides an intelligent recommendation system for computer information teaching resources, comprising: The data collection module is used to collect user learning behavior data and teaching resource attribute data in the computer information teaching platform, and to construct a recommendation graph that includes user nodes, resource nodes and at least one behavioral interaction relationship. The processing module is used to vectorize the nodes in the recommendation graph using graph embedding learning methods, and generate user preference representation vectors and resource feature representation vectors. The processing module is also used to integrate the contextual information of the current learning task, calculate the contribution weight of different interaction relationship paths in the recommendation graph to the recommendation decision through a graph attention network, and thereby dynamically adjust the user preference representation vector. The processing module is also used to periodically monitor the amount of preference drift between the user's current cognitive state and historical cognitive state; The execution module is used to trigger the weight reallocation mechanism in the graph attention network when the detected preference drift amount is greater than a preset threshold, and generate a dynamically adjusted teaching resource recommendation list based on the matching degree between the updated user preference representation vector and the resource feature representation vector.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: In this embodiment, by collecting user learning behavior data and teaching resource attribute data from a computer information teaching platform, a recommendation graph containing user nodes, resource nodes, and behavioral interaction relationships is constructed. This enables structured modeling of user preferences and resource characteristics, providing a foundation for refined recommendations. Furthermore, the scheme employs a graph embedding learning method to vectorize the nodes of the recommendation graph, mapping user preferences and resource features to a unified vector space. This preserves the node's own attribute information while integrating neighborhood structural features, achieving a quantitative representation of user historical behavior and resource characteristics. Further, considering the current learning task context, a graph attention network is used to calculate the importance of different interaction paths, allowing the user preference vector to dynamically adjust under the task scenario, improving the targeting of recommendation decisions. The scheme periodically monitors user cognition. The preference drift between the current state and the historical cognitive state is analyzed. When the preference drift exceeds a preset threshold, a graph attention network weight reallocation mechanism is triggered. This mechanism reduces the attention weight of paths highly correlated with historical preferences and increases the weight of paths aligned with the current preference change direction, achieving dynamic adaptation of user preferences. This mechanism ensures that the recommendation model responds in real-time to cognitive changes and interest shifts, effectively solving the problems of lag and bias in traditional recommendations. Finally, a dynamically adjusted list of recommended teaching resources is generated based on the matching degree between the updated user preference vector and the resource feature vector. The matching degree ranking is then corrected by combining knowledge point connections and task context, ensuring that the recommendation results balance the accuracy of learning needs with the completeness of knowledge logic. In summary, this scheme can achieve dynamic reallocation of graph attention based on preference drift, thereby improving the adaptive accuracy of teaching resource recommendations. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating an application scenario of the intelligent recommendation method for computer information teaching resources according to some embodiments of this application; Figure 2 This is an exemplary flowchart of a computer information teaching resource intelligent recommendation method according to some embodiments of this application; Figure 3 This is a schematic flowchart illustrating the process of determining the amount of preference drift according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of an intelligent recommendation system for computer information teaching resources according to some embodiments of this application; Figure 5 This is a schematic diagram of the structure of a computer device that implements a method for intelligent recommendation of computer information teaching resources according to some embodiments of this application. Detailed Implementation

[0017] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] refer to Figure 1 This figure is a schematic diagram of an application scenario for an intelligent recommendation method for computer information teaching resources according to some embodiments of this application. The figure includes a data acquisition device, a server, a communication network, and a terminal. The data acquisition device is connected to the server via a network, and the terminal is connected to the server system via a communication network. The server acquires user learning behavior data and teaching resource attribute data provided by the data acquisition device. After organizing and associating the data, it constructs a recommendation graph containing user nodes, resource nodes, and at least one behavioral interaction relationship. Subsequently, using a graph embedding learning method, the nodes in the recommendation graph are vectorized to generate user preference representation vectors and resource feature representation vectors. Then, the context information of the current learning task is fused, and the graph attention network is used to calculate the... The recommendation graph assigns weights to different interaction paths in the recommendation decision-making process, thereby dynamically adjusting the user preference representation vector. Simultaneously, it periodically monitors the preference drift between the user's current and historical cognitive states. When the preference drift exceeds a preset threshold, a weight reallocation mechanism in the graph attention network is triggered. Based on the matching degree between the updated user preference representation vector and the resource feature representation vector, a dynamically adjusted list of recommended teaching resources is generated. When an adjustment request for a specific learning user group or teaching resource recommendation is received via a terminal, the server feeds back the dynamic recommendation list, resource matching degree indicators, and preference drift prompts to the terminal, allowing teaching administrators, course instructors, or student users to view and select suitable teaching resources.

[0019] The data collection devices may include a learning behavior log recording system, a teaching resource attribute annotation interface, a teaching platform interactive data collection tool, and a user cognitive status monitoring module; the terminal may be, but is not limited to, a teaching management backend, a course learning terminal, a teacher preparation platform, or a student learning client; the server may be a teaching resource computing cluster deployed locally on campus, or a distributed teaching resource intelligent recommendation service platform built on the education cloud.

[0020] refer to Figure 2 The figure is an exemplary flowchart of a computer information teaching resource intelligent recommendation method according to some embodiments of this application. The method mainly includes the following steps: In step 101, user learning behavior data and teaching resource attribute data in the computer information teaching platform are collected, and a recommendation graph containing user nodes, resource nodes and at least one behavioral interaction relationship is constructed.

[0021] It should be noted that the learning behavior data in this embodiment refers to the digital operation records generated by users during the learning process, which are recorded by the computer information teaching platform and used to analyze their learning preferences and status; the teaching resource attribute data refers to the structured descriptive information that describes the content and structural characteristics of the teaching resources themselves.

[0022] In some embodiments, the collection of user learning behavior data and teaching resource attribute data in the computer information teaching platform can be achieved in the following ways: It can be achieved through a data collection module deployed on the backend of the teaching platform server. This module first listens for and intercepts all legitimate application programming interface requests and responses between the platform's front-end application and the server. For learning behavior data, the module parses and extracts key fields from user requests, including the user identifier (User ID), the identifier of the operated teaching resource (Item ID), the behavior type (e.g., click, play, pause, favorite, rate, submit assignment, complete quiz), the timestamp of the behavior, and accompanying contextual parameters (e.g., video playback progress, quiz score). Simultaneously, the module periodically synchronizes batch historical behavior records from the platform's database operation log table. For teaching resource attribute data, the module directly connects to the platform's resource management database, periodically extracting or monitoring resource change events in real time through a predetermined data interface to obtain the structured attributes of each teaching resource, including its associated knowledge point, preset difficulty level tags, resource media type (e.g., video, text, code environment, exercise set), and metadata such as creator, duration, and file size.

[0023] In some embodiments, constructing a recommendation graph that includes user nodes, resource nodes, and at least one behavioral interaction relationship can be achieved through the following steps: Each user in the computer information teaching platform is uniquely identified, and a corresponding user node is generated. Based on the attribute data of teaching resources, the knowledge point tags, difficulty levels and resource types of teaching resources in the computer information teaching platform are structured and encoded to generate resource nodes; Learning behavior data is divided according to the time dimension, and long-term interactive behavior features that represent users' long-term learning preferences and short-term interactive behavior features that represent recent changes in learning focus are extracted. Based on the long-term and short-term interaction behavior characteristics, the interaction strength between user nodes and resource nodes is calculated, and at least one behavioral interaction relationship edge is determined based on the interaction strength. A recommendation graph is constructed based on the user nodes, resource nodes, and behavioral interaction relationships.

[0024] It should be noted that, in this embodiment, user nodes are node entities used to represent a single user learning subject and the objects carrying their behavior in the recommendation graph; resource nodes are node entities used to represent the attribute structure of teaching resources and the objects they interact with in the recommendation graph; long-term interaction behavior characteristics refer to behavioral characteristics that characterize the stable learning preference trends formed by users over a long period of time; short-term interaction behavior characteristics refer to behavioral characteristics that characterize the changes in the learning attention state of users within the most recent time window; interaction intensity is a numerical representation that quantifies the closeness of interaction between user nodes and resource nodes; behavioral interaction relationship edges are connection structures used to represent the learning behavior associations between user nodes and resource nodes; and the recommendation graph is a structured relationship network used to uniformly represent users, teaching resources, and their behavioral interaction relationships.

[0025] In practical applications, firstly, each user in the computer information teaching platform is uniquely identified, and the corresponding user node is generated in the following way: The unique identifier of each user, such as user ID or username, can be extracted from the user center database or registration information table of the computer information teaching platform. This identifier is then used as the core index of the user in the graph. At the same time, some static attributes of the user (such as registration time and class) can be selected as the initial attribute features of the user node. Furthermore, a graph vertex indexed by the user's unique identifier is created in the graph database, and this graph vertex is used as the corresponding user node.

[0026] Secondly, based on the attribute data of teaching resources, the knowledge point tags, difficulty levels, and resource types of teaching resources in the computer information teaching platform are structured and encoded to generate resource nodes. This can be achieved in the following way: For each teaching resource, extract its three key dimensions: knowledge point tags, difficulty levels, and resource types. For the knowledge point tags, map them to the pre-built platform knowledge graph to obtain their standardized knowledge point codes. For the difficulty levels and resource types, use one-hot encoding or preset enumeration values ​​to perform numerical conversion. Concatenate the codes of the above three dimensions into a unified multi-dimensional feature vector, create a graph vertex for each resource, use the above feature vector as its attribute, and use this graph vertex as the corresponding resource node.

[0027] Secondly, the learning behavior data can be divided according to the time dimension to extract long-term interaction behavior features that represent users' long-term learning preferences and short-term interaction behavior features that represent recent changes in learning focus. This can be achieved in the following way: First, two time windows can be set. For example, the learning behavior data within the past three months can be divided into a long-term window, and the learning behavior data within the past week can be divided into a short-term window. The behavior data within each window can be statistically analyzed. For all interaction records between the user and each resource, basic indicators such as the total number of interactions, the average duration of each interaction, and the last interaction time can be calculated. For the long-term window data, the focus is on calculating the smoothed mean and variance to reflect stability. For the short-term window data, the focus is on calculating the recent growth rate or the recency index based on time decay weighting. Finally, the statistical vector calculated under the long-term window is used as the long-term interaction behavior feature, and the statistical vector calculated under the short-term window is used as the short-term interaction behavior feature.

[0028] Then, based on the long-term and short-term interaction behavior characteristics, the interaction strength between user nodes and resource nodes is calculated, and at least one behavioral interaction relationship edge is determined according to the interaction strength. This can be achieved in the following way: After obtaining the long-term and short-term interaction behavior characteristics of the user, the learning behavior generated by the user with each teaching resource at different time scales is quantified. The long-term interaction behavior characteristics are used to statistically analyze the frequency of the user's access to the teaching resource, the cumulative learning time, and the number of repeated learning sessions over a longer period, reflecting the user's stable learning tendency towards that teaching resource. The short-term interaction behavior characteristics are used to statistically analyze the user's access to the teaching resource within the most recent time window. Click behavior, learning start and end times, and continuous access status of teaching resources are used to reflect changes in the user's learning focus at the current stage. Based on this, long-term and short-term interactive behavior features are weighted and fused according to a preset time weighting rule to obtain the interaction intensity between user nodes and corresponding resource nodes. The interaction intensity is then compared with a preset relationship establishment threshold. When the interaction intensity reaches or exceeds the threshold, at least one behavioral interaction relationship edge is established between the corresponding user node and resource node, and the interaction intensity is used as the weight attribute of the behavioral interaction relationship edge. This realizes the construction of behavioral interaction relationships driven by both the user's historical stable preferences and recent learning changes.

[0029] Finally, the recommendation graph can be constructed based on the user nodes, resource nodes, and behavioral interaction edges as follows: After determining the user nodes, resource nodes, and the behavioral interaction edges between user nodes and resource nodes, each user node and resource node is managed uniformly as a basic entity node in the graph. The behavioral interaction edges are used as the association structure connecting user nodes and resource nodes. During the construction process, each user node and resource node is first assigned a unique node identifier, and the corresponding attribute information is stored as node attributes. Then, the behavioral interaction edges determined based on the interaction strength are written into the connection relationship between nodes, and the interaction strength is used as the weight parameter of the relationship edge to characterize the degree of association between users and teaching resources. On this basis, all nodes and their association relationships are structured to form a graph structure data model with user nodes and resource nodes as the core and behavioral interaction edges as the connection, thereby constructing a recommendation graph that can simultaneously express user learning behavior characteristics and teaching resource attribute characteristics.

[0030] In step 102, the nodes in the recommendation graph are vectorized using a graph embedding learning method to generate user preference representation vectors and resource feature representation vectors.

[0031] In some embodiments, the following steps can be taken to vectorize the nodes in the recommendation graph using a graph embedding learning method to generate user preference representation vectors and resource feature representation vectors: Load corresponding attribute features, including user history behavior, resource knowledge point tags, difficulty level and resource type, into user nodes and resource nodes in the recommendation graph as initial node features; Based on the behavioral interaction relationships between nodes, the initial node features of adjacent nodes are aggregated, and the node representation is updated by combining the node features of the node itself, thus obtaining the node vector of the corresponding node. Repeated aggregation and multiple rounds of updating the node's own features enable the node vector to contain both its own attribute information and neighborhood structure information. The updated node vector of the user node is output as the user preference representation vector, and the node vector of the resource node is output as the resource feature representation vector.

[0032] It should be noted that the neighborhood structure information in this embodiment is structural feature information used to characterize the relationship pattern between a node and its multi-level adjacent nodes in the recommendation graph; the user preference representation vector is a vector representation used to characterize the user's comprehensive learning interests and behavioral tendencies in the teaching resource space; and the resource feature representation vector is a vector representation used to characterize the comprehensive attribute features of teaching resources and their performance in user interaction relationships.

[0033] In practical application, firstly, corresponding attribute features are loaded onto user nodes and resource nodes in the recommendation graph, including user historical behavior, resource knowledge point tags, difficulty level, and resource type. These can be implemented as initial node features in the following way: For user nodes, statistical features are extracted from user historical behavior logs, such as the total number of learning resources, the distribution ratio of each behavior type, and the average learning time, and these statistics are combined into a numerical vector; For resource nodes, the structured encoded knowledge point tag vector, difficulty level encoding, and resource type encoding are concatenated to form a unified attribute vector; This attribute vector prepared for each node is bound to the unique identifier of the node in the graph, and this bound attribute vector is used as the initial node feature of the node.

[0034] Secondly, based on the behavioral interaction relationships between nodes, the initial node features of adjacent nodes are aggregated, and the node representation is updated by combining the node's own node features to obtain the node vector of the corresponding node. This can be achieved in the following way: Select a graph neural network layer as the basic operator for aggregation and updating, such as a graph attention layer; For the currently processed center node, firstly, find all its first-order adjacent nodes through the behavioral interaction relationship edges in the recommendation graph; read the initial node features of these adjacent nodes; according to the type and weight of the behavioral interaction relationship edges connecting the center node and each adjacent node, use the attention mechanism to calculate the weighted sum of the features of each adjacent node, which is the aggregation process; concatenate the neighborhood feature vector obtained by aggregation with the initial node features of the center node itself, and input the concatenated vector into a fully connected neural network layer, and then process it through a non-linear activation function to obtain a new feature vector; use this newly generated feature vector as the node vector of the center node after this round of updates.

[0035] Then, the aggregation and update process is iterated multiple times. A fixed number of iterations is set, such as three rounds. In each iteration, each node performs new aggregation and updates based on the node vectors obtained by its neighboring nodes in the previous iteration (using the initial node features in the first round), thereby generating its own new node vector after the current iteration. In this way, after the second iteration, the node vector of each node already contains the attribute information of its one-hop neighbors. After the third iteration, the node vector of each node further integrates the attribute information and structural patterns of its neighbors within two hops. Finally, the final node vector obtained by each node after the last iteration contains both its own initial attribute information and the structural information passed from its multi-hop neighborhood.

[0036] Finally, after completing the iteration, all nodes in the recommendation graph have obtained their final node vectors. At this point, the final node vectors of all user nodes are extracted and formed into a set; similarly, the final node vectors of all resource nodes are extracted and formed into another set; the final node vector of each user node in the set is used as the user preference representation vector of that user; and the final node vector of each resource node in the set is used as the resource feature representation vector of that resource.

[0037] In step 103, the contextual information of the current learning task is fused, and the contribution weights of different interaction paths in the recommendation graph to the recommendation decision are calculated through a graph attention network, thereby dynamically adjusting the user preference representation vector.

[0038] In some embodiments, the contribution weights of different interaction paths in the recommendation graph to the recommendation decision can be calculated using a graph attention network by incorporating contextual information of the current learning task: The knowledge point tags, difficulty level, and learning stage information associated with the current learning task are vectorized and encoded to form a task context vector; Extract the meta-path features corresponding to various interaction relationship paths in the recommendation graph, and jointly represent the meta-path features with the task context vector; The input graph attention network will be jointly represented, and the relative importance weights of different interaction paths in the current learning task will be evaluated in parallel through a multi-head attention structure. The relative importance weights corresponding to each interaction path are normalized to obtain the contribution weights of different interaction paths in the recommendation graph to the recommendation decision.

[0039] It should be noted that, in this embodiment, the task context vector is a vector representation used to characterize the comprehensive state of the current learning task in terms of knowledge content, learning difficulty, and learning process; the joint representation refers to the comprehensive feature representation that integrates the structural features of the interaction relationship path and the context information of the current learning task; the relative importance weight is a weight parameter that measures the degree of influence of different interaction relationship paths on the recommendation decision under the current learning task conditions; and the contribution weight is a weight value used to quantify the actual degree of influence of different interaction relationship paths on the final teaching resource recommendation decision.

[0040] In practical application, firstly, the knowledge point tags, difficulty level, and learning stage information associated with the current learning task are vectorized and encoded to form a task context vector. This can be achieved in the following way: After receiving the current learning task, the knowledge point tags, difficulty level, and learning stage information corresponding to the learning task are obtained from the teaching platform. Among them, the knowledge point tags are used to describe the topic range of the learning content, the difficulty level is used to reflect the cognitive load level of the task, and the learning stage is used to indicate the position of the task in the overall learning path. Then, the above types of information are numerically represented in a unified dimension and combined in a preset order to obtain a feature vector that can simultaneously reflect the attributes of the learning content and the characteristics of the learning process. This feature vector is then used as the task context vector.

[0041] Secondly, the extraction of meta-path features corresponding to various interaction paths in the recommendation graph and the joint representation of the meta-path features and the task context vector can be achieved in the following way: Based on the constructed recommendation graph, enumerate the interaction paths formed by user nodes and resource nodes through different types of behavioral interaction relationships, and extract meta-path features that can reflect the path structure characteristics for each type of interaction path. The meta-path features are used to describe the overall pattern of node type combination and interaction relationship type in the path. On this basis, the meta-path features and the corresponding task context vector are concatenated or superimposed according to a unified feature dimension to form a joint feature representation that simultaneously contains path structure information and task context constraint information.

[0042] Then, the joint representation is input into a graph attention network. The relative importance weights of different interaction paths in the current learning task are evaluated in parallel using a multi-head attention structure. This can be achieved as follows: the joint representation is fed as input samples into a pre-constructed and trained graph attention network, which is trained on historical recommendation samples. During training, actual user learning feedback is used as a supervision signal, and the network's internal parameters are adjusted iteratively to enable the network to distinguish the importance of different interaction paths in different learning task scenarios. In the inference phase, multiple attention branches in the graph attention network process the input joint representation in parallel, calculating the attention score corresponding to each interaction path, thereby obtaining the relative importance weights reflecting the differences in the relative contributions of different interaction paths. It should be further noted that the graph attention network is used to evaluate the relative importance of different interaction paths in graph-structured data. This deep learning model adaptively assigns weights based on the importance of nodes or relationship paths. Its technical principle lies in introducing an attention mechanism to differentiate the contribution of different interaction paths in the graph under specific task conditions. In this scheme, the graph attention network uses the joint representation extracted from the recommendation graph as input. By calculating attention scores for each interaction path, it characterizes the strength of different paths' influence on user decisions under the current learning task context. During the training phase, the model iteratively updates its internal parameters using historical learning behavior and recommendation feedback, ensuring that the attention allocation results reflect real changes in user preferences. During the inference phase, the graph attention network automatically assigns different weights to different interaction paths based on the learned parameter structure, thereby effectively filtering and strengthening multi-source relationship information and providing a more suitable decision-making basis for subsequent teaching resource recommendations.

[0043] Finally, the relative importance weights corresponding to each interaction path are normalized to obtain the contribution weights of different interaction paths in the recommendation graph to the recommendation decision. This can be achieved in the following way: the relative importance weights corresponding to each interaction path are processed with a unified scale, and each weight value is mapped to the same numerical range through a preset normalization rule, so that the sum of the weights of all interaction paths satisfies the preset constraint condition. After normalization, the weight value corresponding to each interaction path can directly reflect its actual contribution ratio to the recommendation result under the current learning task condition. Thus, the normalized weight result is used as the contribution weight of different interaction paths in the recommendation graph to the recommendation decision.

[0044] In step 104, the amount of preference drift between the user's current cognitive state and historical cognitive state is periodically monitored.

[0045] In some embodiments, reference Figure 3As shown in the figure, this is a flowchart illustrating the determination of preference drift in some embodiments of this application. In this embodiment, the periodic monitoring of preference drift between the user's current cognitive state and historical cognitive state can be achieved through the following steps: In step 1041, the latest user preference representation vector is extracted as an approximate representation of the current cognitive state; In step 1042, the user's historical cognitive state benchmark is determined based on the cognitive state obtained from historical periodic monitoring; In step 1043, the difference between the current cognitive state and the historical cognitive state benchmark is calculated to obtain the preference difference value, and the preference difference value is used as the preference drift amount.

[0046] It should be noted that the preference drift in this embodiment is a state change indicator that reflects the extent to which a user's learning preferences change over time.

[0047] In practical application, firstly, after each preset monitoring period, the vector result corresponding to the target user is selected from the latest user preference representation vector obtained from graph embedding learning and graph attention network processing, as an approximate representation of the user's cognitive state at the current time point; secondly, the system continuously stores the approximate representation of the user's cognitive state over multiple historical monitoring periods, and when monitoring preference drift, it selects one or more historical states that meet the preset time span requirement from the historical cognitive state approximate representation as references. By statistically processing the selected historical states, a historical cognitive state benchmark for comparative analysis is obtained; then, the difference between the approximate representation of the current cognitive state and the historical cognitive state benchmark is calculated, and the degree of change of the two in each corresponding dimension is uniformly measured to obtain a preference difference value reflecting the degree of difference between the two cognitive states. The preference difference value is directly used as the user preference drift amount to indicate the magnitude of change of the user's current learning preference relative to the historical stable state, and the preference difference value is output as the preference drift amount.

[0048] In step 105, when the detected preference drift amount is greater than a preset threshold, the weight reallocation mechanism in the graph attention network is triggered, and a dynamically adjusted list of recommended teaching resources is generated based on the matching degree between the updated user preference representation vector and the resource feature representation vector.

[0049] It should be noted that the preset threshold for preference drift in this application is an adaptive threshold. The adaptive threshold is dynamically updated based on the statistical distribution of the preference drift of the target user in multiple historical monitoring periods, so as to match the preference drift triggering conditions with the fluctuation range of the user's individual learning behavior, thereby reducing the probability of false triggering and improving the effectiveness of weight redistribution.

[0050] In some embodiments, when the detected preference shift exceeds a preset threshold, a weight reallocation mechanism in the graph attention network is triggered. This weight reallocation mechanism in the graph attention network can be implemented using the following steps: The adjustment range of attention branches in the graph attention network is determined based on the magnitude and direction of preference drift; For interaction paths that are highly correlated with historical preference features, reduce their corresponding attention weights; Increase the attention weight of interaction paths that are consistent with the direction of current preference change; The adjusted attention weights are normalized, and the weight configuration of the graph attention network is updated.

[0051] It should be noted that the adjustment range of the attention branches in this embodiment is a control parameter used to limit the adjustment range of the weights of each attention branch in the graph attention network; the attention weight is a weight parameter that characterizes the degree of influence of the interaction relationship path in the recommendation decision.

[0052] In practical application, firstly, after detecting that the preference drift exceeds a preset threshold, the system first reads the magnitude of the preference drift to measure the extent of the change in user preference. Simultaneously, it combines the direction of the preference drift to determine whether the user preference is shifting towards a new learning topic or refining within the existing topic range. Based on the magnitude of the preference drift, the system maps it to a pre-set adjustment range, thereby determining the maximum allowable adjustment range for each attention branch in the graph attention network. Secondly, it performs correlation analysis between the dominant user preference features in the historical cognitive state baseline and the interaction paths in the recommendation graph, identifying interaction paths that highly match the user's historical stable learning preferences. Within the adjustment range of the attention branches, the attention weight of the corresponding interaction path is gradually reduced, weakening its influence in the current recommendation decision. The adjusted value is used as the attention weight of that interaction path. Then, based on the preference drift reflected by the preference... The system identifies interaction paths that align with current learning interest shifts, such as frequently interacted resource types or newly emerging knowledge point associations. Within a defined attention branch adjustment range, the attention weights corresponding to these interaction paths are incrementally adjusted to enhance their role in recommendation decisions. The adjusted values ​​are then used as the attention weights for the corresponding interaction paths. This approach allows the recommendation model to more sensitively respond to changes in user learning needs at the current stage. Finally, after adjusting the attention weights for each interaction path, all attention weights are uniformly normalized to maintain them within a preset range and ensure the stability of the overall weight distribution. The normalized attention weights are then written into the parameter storage module of the graph attention network, replacing the original weight configuration. This completes the weight redistribution of the graph attention network during the current monitoring period, and the updated parameter set is used as the new weight configuration.

[0053] It should also be noted that when the graph attention network in this application performs the weight redistribution mechanism, it can introduce a weight change smoothing constraint to limit the magnitude of attention weight changes on the same interaction path within adjacent monitoring periods, so as to suppress drastic fluctuations in recommendation results caused by short-term abnormal behavior and improve the stability and continuity of the recommendation process.

[0054] In some embodiments, generating a dynamically adjusted list of recommended teaching resources based on the matching degree between the updated user preference representation vector and the resource feature representation vector can be achieved through the following steps: Calculate the similarity measure between the updated user preference representation vector and the resource feature representation vector of the candidate teaching resources to obtain the initial matching result; The initial matching degree result is corrected by combining the knowledge point connection relationship between teaching resources and the contextual constraints of the current learning task; The candidate teaching resources are ranked based on the corrected matching results; A list of recommended teaching resources is generated based on the sorting results.

[0055] It should be noted that the teaching resource recommendation list in this embodiment is used to display a set of teaching resources that match the user's current learning needs and preferences; context constraints are used to limit the teaching resource recommendations to meet the requirements of the current learning task objectives and learning stage.

[0056] In practical application, after updating the user preference representation vector, the vector similarity calculation module is first invoked to compare the updated user preference representation vector with the resource feature representation vectors corresponding to the candidate teaching resources one by one. An initial matching degree result for each candidate teaching resource is obtained through conventional vector similarity calculation. Then, based on the established knowledge point connections between teaching resources, the initial matching degree result is checked for sequential consistency. Furthermore, considering the contextual constraints of knowledge point levels, difficulty progression, and learning stages in the current learning task, the initial matching degree results for teaching resources that do not meet the requirements of the current learning path are adjusted to obtain a corrected matching degree result. Based on this, all candidate teaching resources are sorted according to the corrected matching degree result to form an ordered candidate resource sequence. Finally, according to the preset recommendation list generation rules, teaching resources are selected sequentially from the ordered candidate resource sequence, outputting a dynamically adjusted teaching resource recommendation list.

[0057] On the other hand, in some embodiments, this application provides an intelligent recommendation system for computer information teaching resources, with reference to... Figure 4 The figure is a schematic diagram of the structure of a computer information teaching resource intelligent recommendation system 400 according to some embodiments of this application. The computer information teaching resource intelligent recommendation system 400 includes: a data acquisition module 401, a processing module 402, and an execution module 403, which are described below: The data collection module 401 is used to collect user learning behavior data and teaching resource attribute data in the computer information teaching platform, and to construct a recommendation graph that includes user nodes, resource nodes and at least one behavioral interaction relationship. Processing module 402 is used to vectorize the nodes in the recommendation graph using a graph embedding learning method to generate user preference representation vectors and resource feature representation vectors; The processing module 402 is also used to integrate the contextual information of the current learning task, calculate the contribution weight of different interaction relationship paths in the recommendation graph to the recommendation decision through the graph attention network, and dynamically adjust the user preference representation vector accordingly. The processing module 402 is also used to periodically monitor the amount of preference drift between the user's current cognitive state and historical cognitive state; The execution module 403 is used to trigger the weight reallocation mechanism in the graph attention network when the detected preference drift amount is greater than a preset threshold, and generate a dynamically adjusted teaching resource recommendation list based on the matching degree between the updated user preference representation vector and the resource feature representation vector.

[0058] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described intelligent recommendation method for computer information teaching resources.

[0059] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device implementing a method for intelligent recommendation of computer information teaching resources according to some embodiments of this application. The intelligent recommendation method for computer information teaching resources in the above embodiments can... Figure 5 The computer device shown is used to implement this, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0060] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0061] The communication bus 502 can be used to transmit information between the aforementioned components.

[0062] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0063] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. In the above embodiments, the intelligent recommendation method for computer information teaching resources can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0064] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0065] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0066] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0067] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent recommendation method for computer information teaching resources.

[0068] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0069] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for intelligent recommendation of computer information teaching resources, characterized in that, include: Collect user learning behavior data and teaching resource attribute data in the computer information teaching platform, and construct a recommendation graph that includes user nodes, resource nodes and at least one behavioral interaction relationship; Using graph embedding learning methods, the nodes in the recommendation graph are vectorized to generate user preference representation vectors and resource feature representation vectors; By integrating the contextual information of the current learning task, the contribution weights of different interaction paths in the recommendation graph to the recommendation decision are calculated through a graph attention network, thereby dynamically adjusting the user preference representation vector. Periodically monitor the amount of preference drift between the user's current cognitive state and historical cognitive state; When the detected preference drift exceeds a preset threshold, the weight redistribution mechanism in the graph attention network is triggered, and a dynamically adjusted list of recommended teaching resources is generated based on the matching degree between the updated user preference representation vector and the resource feature representation vector.

2. The method as described in claim 1, characterized in that, Constructing a recommendation graph that includes user nodes, resource nodes, and at least one behavioral interaction relationship specifically includes: Each user in the computer information teaching platform is uniquely identified, and a corresponding user node is generated. Based on the attribute data of teaching resources, the knowledge point tags, difficulty levels and resource types of teaching resources in the computer information teaching platform are structured and encoded to generate resource nodes; Learning behavior data is divided according to the time dimension, and long-term interactive behavior features that represent users' long-term learning preferences and short-term interactive behavior features that represent recent changes in learning focus are extracted. Based on the long-term and short-term interaction behavior characteristics, the interaction strength between user nodes and resource nodes is calculated, and at least one behavioral interaction relationship edge is determined based on the interaction strength. A recommendation graph is constructed based on the user nodes, resource nodes, and behavioral interaction relationships.

3. The method as described in claim 1, characterized in that, Using graph embedding learning methods, nodes in the recommendation graph are vectorized to generate user preference representation vectors and resource feature representation vectors. Specifically, this includes: Load corresponding attribute features, including user history behavior, resource knowledge point tags, difficulty level and resource type, into user nodes and resource nodes in the recommendation graph as initial node features; Based on the behavioral interaction relationships between nodes, the initial node features of adjacent nodes are aggregated, and the node representation is updated by combining the node features of the node itself, thus obtaining the node vector of the corresponding node. Repeated aggregation and multiple rounds of updating the node's own features enable the node vector to contain both its own attribute information and neighborhood structure information. The updated node vector of the user node is output as the user preference representation vector, and the node vector of the resource node is output as the resource feature representation vector.

4. The method as described in claim 1, characterized in that, By integrating contextual information from the current learning task, the contribution weights of different interaction paths in the recommendation graph to the recommendation decision are calculated using a graph attention network. Specifically, this includes: The knowledge point tags, difficulty level, and learning stage information associated with the current learning task are vectorized and encoded to form a task context vector; Extract the meta-path features corresponding to various interaction relationship paths in the recommendation graph, and jointly represent the meta-path features with the task context vector; The input graph attention network will be jointly represented, and the relative importance weights of different interaction paths in the current learning task will be evaluated in parallel through a multi-head attention structure. The relative importance weights corresponding to each interaction path are normalized to obtain the contribution weights of different interaction paths in the recommendation graph to the recommendation decision.

5. The method as described in claim 1, characterized in that, Periodically monitoring the preference drift between a user's current cognitive state and their historical cognitive state specifically includes: Extract the latest user preference representation vector as an approximate representation of the current cognitive state; The user's historical cognitive state benchmark is determined based on the cognitive state obtained from historical periodic monitoring. The difference between the current cognitive state and the historical cognitive state benchmark is calculated to obtain the preference difference value, and the preference difference value is used as the preference drift amount.

6. The method as described in claim 1, characterized in that, Based on the matching degree between the updated user preference representation vector and the resource feature representation vector, a dynamically adjusted list of recommended teaching resources is generated, specifically including: Calculate the similarity measure between the updated user preference representation vector and the resource feature representation vector of the candidate teaching resources to obtain the initial matching result; The initial matching degree result is corrected by combining the knowledge point connection relationship between teaching resources and the contextual constraints of the current learning task; The candidate teaching resources are ranked based on the corrected matching results; A list of recommended teaching resources is generated based on the sorting results.

7. The method as described in claim 1, characterized in that, The preset threshold is an adaptive threshold, which is dynamically updated based on the statistical distribution of the target user's preference drift over multiple historical monitoring periods, so that the preference drift triggering conditions match the fluctuation range of the user's individual learning behavior.

8. The method as described in claim 1, characterized in that, The learning behavior data includes operation event data collected by the computer information teaching platform during the user's learning process. The operation event data includes at least the behavior type, occurrence time, and associated resource identifier corresponding to the user's access to, interaction with, and completion of teaching resources, which is used to characterize the user's behavioral features during the learning process.

9. The method as described in claim 1, characterized in that, The teaching resource attribute data is descriptive information that provides a structured representation of the content characteristics and organizational structure characteristics of the teaching resources themselves. The descriptive information is used to characterize at least the knowledge content attributes, difficulty level, and resource type of the teaching resources.

10. A computer information teaching resource intelligent recommendation system, characterized in that, include: The data collection module is used to collect user learning behavior data and teaching resource attribute data in the computer information teaching platform, and to construct a recommendation graph that includes user nodes, resource nodes and at least one behavioral interaction relationship. The processing module is used to vectorize the nodes in the recommendation graph using graph embedding learning methods, and generate user preference representation vectors and resource feature representation vectors. The processing module is also used to integrate the contextual information of the current learning task, calculate the contribution weight of different interaction relationship paths in the recommendation graph to the recommendation decision through a graph attention network, and thereby dynamically adjust the user preference representation vector. The processing module is also used to periodically monitor the amount of preference drift between the user's current cognitive state and historical cognitive state; The execution module is used to trigger the weight reallocation mechanism in the graph attention network when the detected preference drift amount is greater than a preset threshold, and generate a dynamically adjusted teaching resource recommendation list based on the matching degree between the updated user preference representation vector and the resource feature representation vector.

Images