Evaluation task recommendation method

Abstract

The invention relates to an evaluation task recommendation method, and the method comprises the steps: constructing a user preference vector and a plurality of evaluation task feature vectors based on the pre-obtained basic information and evaluation task information of a target teacher; determining an initial recommendation list according to the user preference vector and the feature vector of each evaluation task; and performing optimization processing on the initial recommendation list by adopting a multi-target optimization model constructed based on a task diversity index and an accuracy index to obtain a target recommendation task. According to the method, the initial recommendation list is optimized by adopting the multi-target optimization model constructed based on the task diversity index and the accuracy index, so that the recommendation result can be further optimized on the basis of recommendation accuracy. According to the model, two key factors of accuracy and diversity are considered at the same time, and the problem that recommendation tasks are too single due to the fact that only accuracy is paid attention to or recommendation accuracy is sacrificed due to the fact that only diversity is pursued is solved.

Description

Technical Field

[0001] The present application relates to the field of intelligent assessment technology, and in particular to a method for recommending assessment tasks. Background Art

[0002] In the field of education, as the importance of teacher professional development continues to increase, intelligent support service platforms for teacher professional development have emerged. These platforms aim to provide personalized professional development support for teachers. One of their important functions is to intelligently recommend assessment tasks that are in line with their own development. However, there are many challenges in achieving accurate and effective assessment task recommendations.

[0003] In the existing intelligent support service scenarios for teacher professional development, the recommendation of teacher assessment tasks usually adopts some more traditional methods. For example, the most popular and latest assessment tasks with high ranking are recommended to users, but this recommendation method is not accurate enough. Summary of the invention

[0004] Based on this, it is necessary to provide an evaluation task recommendation method that can accurately recommend evaluation tasks in response to the above technical problems.

[0005] In a first aspect, the present application provides a method for recommending an assessment task, the method comprising:

[0006] Based on the pre-acquired basic information of the target teacher and the assessment task information, a user preference vector and multiple assessment task feature vectors are constructed;

[0007] Determine the initial recommendation list based on the user preference vector and the feature vectors of each evaluation task;

[0008] Based on a pre-built multi-objective optimization model, the initial recommendation list is optimized to obtain the target recommendation task; the multi-objective optimization model is built based on task diversity indicators and accuracy indicators.

[0009] In one of the embodiments, the above-mentioned determining the initial recommendation list based on the user preference vector and the feature vector of each evaluation task includes: using the cosine similarity algorithm to calculate the similarity between the user preference vector and the feature vector of each evaluation task; sorting each evaluation task according to each similarity to determine the initial recommendation list.

[0010] In one of the embodiments, the above-mentioned determining the initial recommendation list based on the user preference vector and the feature vectors of each assessment task includes: generating a target teacher task affinity graph based on the user preference vector and the feature vectors of each assessment task; the target teacher task affinity graph is used to characterize the degree of matching between the target teacher and multiple assessment tasks; and generating the initial recommendation list based on the target teacher task affinity graph.

[0011] In one of the embodiments, the target teacher task affinity graph is generated based on the user preference vector and the feature vectors of each assessment task, including: performing node feature initialization processing on the user preference vector and the feature vectors of each assessment task, respectively, to obtain a first feature vector of the teacher node and second feature vectors of multiple assessment task nodes; obtaining an initial teacher task affinity graph based on the distance between the first feature vector and each second feature vector; and updating the initial teacher task affinity graph based on a graph neural network to obtain a target teacher task affinity graph.

[0012] In one of the embodiments, the above-mentioned initial teacher task affinity graph is obtained based on the distance between the first eigenvector and each second eigenvector, including: inputting the distance between the first eigenvector and each second eigenvector into the distance function for distance calculation processing to obtain the distance result; the distance result is used to characterize the distance between the teacher node and each assessment task node; the distance result is input into the edge weight function for calculation processing to obtain edge weight data; the edge weight data is used to characterize the affinity between the teacher node and the assessment task node; and the initial teacher task affinity graph is constructed according to the edge weight data.

[0013] In one of the embodiments, the above-mentioned generating an initial recommendation list based on the target teacher task affinity graph includes: calculating the affinity between the teacher node and each assessment task node in the target teacher task affinity graph to obtain multiple affinity scores; sorting multiple assessment tasks according to the multiple affinity scores to obtain the initial recommendation list.

[0014] In one of the embodiments, the above-mentioned graph neural network includes an input layer and an update layer. Based on the graph neural network, the initial teacher task affinity graph is optimized to obtain a target teacher task affinity graph, including: inputting the first eigenvector and each second eigenvector into the input layer for feature vector mapping processing to obtain a third eigenvector; the dimension of the third eigenvector corresponds to the dimension of the input layer; inputting the third eigenvector into the update layer for feature update processing to obtain an update node; the update node includes a teacher update node and a task update node; according to the teacher update node and the task update node, the edge weight data in the initial teacher task affinity graph is adjusted to obtain the target teacher task affinity graph.

[0015] In one of the embodiments, the edge weight data in the initial teacher-task affinity graph is adjusted and processed according to the teacher update nodes and the task update nodes to obtain the target teacher-task affinity graph, including: calculating the cosine similarity of the teacher update nodes and the task update nodes to obtain the similarity data; determining the recommendation score according to the similarity data; the recommendation score is used to characterize the matching degree between the target teacher and each assessment task; and adjusting the edge weight data according to the recommendation score to obtain the target teacher-task affinity graph.

[0016] In one of the embodiments, the above-mentioned construction of a user preference vector and multiple assessment task feature vectors based on the pre-acquired basic information and assessment task information of the target teacher includes: performing a first feature encoding process on the basic information to obtain a user preference vector; performing a second feature encoding process on the assessment task information to obtain feature vectors of each assessment task.

[0017] In one of the embodiments, the basic information includes discrete features and continuous features; performing a first feature encoding process on the basic information to obtain a user preference vector includes: performing data conversion processing on the continuous features according to a pre-established binning strategy to obtain categorical features; the binning strategy includes one of equal-width binning, equal-frequency binning and custom binning; performing unique-hot encoding processing on the categorical features and discrete features respectively to obtain a user preference vector.

[0018] In one embodiment, the above-mentioned evaluation task information includes multiple operation types; the evaluation task information is subjected to a second feature encoding process to obtain a feature vector of each evaluation task, including: performing a one-hot encoding process on each operation type to obtain a binary vector; performing feature extraction process on the binary vector to obtain a feature set; performing feature conversion process on the feature set to obtain a feature vector of each evaluation task.

[0019] In one of the embodiments, the above-mentioned optimization processing of the initial recommendation list based on the pre-built multi-objective optimization model to obtain the target recommendation task includes: using an adaptive step gradient algorithm to adjust the learning rate of the multi-objective optimization model to obtain the target weight parameter; adjusting the initial recommendation list according to the target weight parameter to obtain the target recommendation task.

[0020] In one of the embodiments, the adaptive step gradient algorithm is used to adjust the learning rate of the multi-objective optimization model to obtain the target weight parameters, including: obtaining the initial weight parameters; in each iterative solution process, calculating the current gradient of the multi-objective optimization model; updating the initial weight parameters according to the current gradient and the fixed step size to obtain the intermediate weight parameters; when the intermediate weight parameters meet the convergence conditions, determining the target weight parameters.

[0021] In a second aspect, the present application also provides an evaluation task recommendation device. The device comprises:

[0022] A vector construction module is used to construct a user preference vector and multiple evaluation task feature vectors based on the pre-acquired basic information of the target teacher and the evaluation task information;

[0023] A list determination module is used to determine the initial recommendation list based on the user preference vector and the feature vectors of each evaluation task;

[0024] The task determination module is used to optimize the initial recommendation list based on a pre-built multi-objective optimization model to obtain the target recommendation task; the multi-objective optimization model is built based on task diversity indicators and accuracy indicators.

[0025] In a third aspect, the present application further provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the first aspect when executing the computer program.

[0026] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the first aspect are implemented.

[0027] In a fifth aspect, the present application further provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps of the first aspect are implemented.

[0028] The above-mentioned assessment task recommendation method constructs a user preference vector and multiple assessment task feature vectors based on the pre-acquired basic information of the target teacher and the assessment task information; determines the initial recommendation list according to the user preference vector and the feature vectors of each assessment task; and optimizes the initial recommendation list using a multi-objective optimization model constructed based on task diversity indicators and accuracy indicators to obtain the target recommendation task. The present application uses a multi-objective optimization model constructed based on task diversity indicators and accuracy indicators to optimize the initial recommendation list, which can further optimize the recommendation results based on the recommendation accuracy. The model takes into account the two key factors of accuracy and diversity at the same time, avoiding the problem of focusing only on accuracy and resulting in too single recommended tasks, or only pursuing diversity and sacrificing recommendation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A diagram of an application environment of a method for recommending an evaluation task in an embodiment;

[0030] Figure 2 A flowchart of a method for recommending an evaluation task in one embodiment;

[0031] Figure 3 A schematic diagram of a process for determining an initial recommendation list in one embodiment;

[0032] Figure 4 A schematic diagram of a process for generating a target teacher task affinity diagram in one embodiment;

[0033] Figure 5 A schematic diagram of a process for obtaining an initial teacher task affinity diagram in one embodiment;

[0034] Figure 6 A schematic diagram of a process for generating an initial recommendation list in one embodiment;

[0035] Figure 7 A schematic diagram of a process for constructing a user preference vector and an evaluation task feature vector in one embodiment;

[0036] Figure 8 is a schematic diagram of a process of obtaining a user preference vector in one embodiment;

[0037] Fig. 9 A schematic diagram of a process for obtaining a feature vector of each evaluation task in an embodiment;

[0038] Fig.10 A schematic diagram of a process for obtaining a target recommendation task in one embodiment;

[0039] Fig.11 A schematic diagram of a process for obtaining a target weight parameter in one embodiment;

[0040] Fig.12 It is a structural block diagram of an evaluation task recommendation device in one embodiment;

[0041] Fig.13 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0043] First of all, before specifically introducing the technical solution of the embodiments of the present application, the technical background based on the embodiments of the present application is introduced.

[0044] In the intelligent support service platform for teacher professional development, it is necessary to intelligently recommend assessment tasks that are in line with the teacher's own development for the three stages of basic education, namely primary school, junior high school and high school, based on key information such as teacher basic information, assessment task model library, ability analysis model, and teacher digital portrait. Based on the formed teacher ability map, teacher teaching and research are carried out to establish a closed loop of "teacher ability assessment, teacher portrait, problem-oriented teaching and research, and teacher ability assessment". Through teacher ability assessment, accurate teacher portraits are achieved. Teachers use portraits to self-diagnose and evaluate their teaching and research performance, recall and review teaching and research activities, plan professional development routes, optimize classroom teaching, and link teaching and research resource services to achieve adaptive and accurate recommendations. Summarize the model, methods, and strategies for accurate teacher teaching and research driven by teacher portraits.

[0045] The current intelligent support service platform for teacher professional development has the following problems:

[0046] Extraction of teacher features under the aggregation of a large amount of assessment data (data sparsity problem): After the teacher group completes a large number of set assessment tasks, this information includes teacher basic information, assessment task model library, ability analysis model, teacher digital portrait and other aggregate data. The accompanying data generated by the assessment tasks completed by teachers may be very sparse, making it difficult to directly use collaborative filtering algorithms. Deep learning models must be introduced to extract and calculate the characteristics of teachers so that high-level features of the data can be discovered by learning a low-dimensional representation of the input data (encoding) and trying to reconstruct the input data (decoding).

[0047] Recommendations based on teachers’ basic information and data (cold start problem): For newly added teachers or newly designed assessment tasks, there is a lack of sufficient historical data for recommendations. In the initial stage, the platform will match some teachers and assessment tasks and generate data, but as new teachers and assessment tasks continue to join, there will be problems such as the inability to match assessment tasks to teachers, and the existing features of teachers will no longer be updated and grown. At this time, it is necessary to combine content recommendation and expert rules to solve this cold start problem.

[0048] Recommend assessment tasks that are in line with teachers' development (balance between diversity and accuracy of assessment tasks): With the increasing number of new teachers and assessment tasks, teacher feature data may tend to develop in a single direction, and the diversity and accuracy balance of assessment tasks will also tend to be chaotic and unstable as information entropy increases. How to balance recommendation accuracy and recommendation diversity in the recommendation system is a challenge. Using a hybrid recommendation system that combines content recommendation and collaborative filtering methods may be an effective strategy.

[0049] The above questions are raised in order to make more accurate recommendations, which is the key to solving the intelligent recommendation evaluation task. What we ultimately want to solve is the accurate recommendation of the intelligent evaluation task, and these three questions belong to three related subset problems.

[0050] These three issues directly affect the performance and user experience of the recommendation system.

[0051] Data sparsity problem: In education and assessment systems, there are usually a large number of possible assessment tasks and teachers, but each teacher's interaction with the assessment task (such as completing or scoring the assessment task) is relatively small. This leads to the sparsity of the data matrix, that is, most of the possible teacher-task combinations have no existing data.

[0052] The additional problem with data dimensionality is that it is very sparse, and the effect of using it in collaborative filtering algorithms is very poor. Data sparsity makes it difficult for recommendation systems to learn enough information from limited user behavior data, and thus it is difficult to accurately predict teachers' preferences or performance on unknown assessment tasks.

[0053] Cold start problem: The cold start problem is that there is not enough historical data for new teachers or newly created assessment tasks to make effective recommendations. For new teachers, the system may not understand their specific needs and preferences; for new tasks, the system also lacks user feedback and performance data.

[0054] The balance between diversity and accuracy:

[0055] Accuracy ensures that the recommendation system can provide assessment tasks that may be of interest or useful to teachers, thereby improving task completion rates and teacher satisfaction.

[0056] Diversity ensures that recommendations are not limited to task types that teachers are already familiar with or frequently exposed to. This can help teachers discover new potentially valuable assessment tasks, promote their all-round development and stimulate new interests.

[0057] Only when these three problems are related and solved in a coordinated manner and finally pushed to the teacher are they appropriate situational assessment tasks.

[0058] In the scenario of intelligent support services for teacher professional development, intelligent recommendation of teacher assessment tasks generally uses resources such as the hottest and latest assessment tasks with high rankings to recommend to users. These scenarios do not use sparse data feature extraction technology in the context of big data, and it is possible that new expanded assessment tasks cannot be effectively recommended. There is no effective solution to the diversity and accuracy balance of assessment tasks. In the field of teacher professional development, there is no effective feature extraction based on the industry standards for teacher digital literacy, and no intelligent recommendation of assessment tasks.

[0059] The following is an analysis of some common shortcomings of existing technologies that are improved:

[0060] Lack of in-depth feature extraction: Traditional recommendation systems usually rely on surface-level user and item features for matching, such as basic user information and simple item attributes. This approach ignores the deeper patterns and connections behind user behavior, limiting the accuracy and personalization of the recommendation system. Failure to fully explore and utilize user behavior data and complex features may result in inaccurate or inappropriate recommendation results.

[0061] Limited ability to process high-dimensional sparse data: Traditional methods do not work well when processing data with high-dimensional sparse characteristics, such as user interaction records for evaluation tasks. This sparsity makes it difficult for recommendation systems to accurately predict user preferences and needs. High-dimensional sparse data is difficult to process effectively with simple models, resulting in reduced performance of recommendation systems.

[0062] Cold start problem: New users or new items lack sufficient interaction records, and traditional recommendation systems find it difficult to provide effective recommendations for them. Recommendation models that rely on historical interaction data cannot work effectively when facing new users or new items, resulting in poor user experience.

[0063] The traditional methods are as follows:

[0064] 1. Collaborative filtering (pure algorithm mode, no AI model introduced, simple algorithm formula).

[0065] Disadvantages: When the number of users or tasks is very large, calculating the similarity between users or tasks becomes very time-consuming, and the sparsity of the data will seriously affect the accuracy of the similarity calculation.

[0066] 2. Matrix decomposition (Matrix decomposition, such as singular value decomposition and alternating least squares, is another common method for dealing with sparse data).

[0067] Disadvantages: Sufficient observation data is required to train the model, and the computational complexity is high.

[0068] 3. Filling technology.

[0069] Impute missing user-item ratings using the mean, median, most frequent value, or more complex predictive models. However, this approach may introduce bias and is not suitable for all scenarios.

[0070] 4. Content-based recommendations.

[0071] When collaborative filtering encounters cold start or data sparsity problems, content-based recommendation can provide an alternative. This method makes recommendations based on the user's previous feedback on the task content (such as keywords, categories, descriptions, etc.), rather than based on the user's or other users' rating history, and lacks intelligent model support.

[0072] Based on this, the present application provides an evaluation task recommendation method, aiming to solve the above technical problems.

[0073] The evaluation task recommendation method provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The terminal constructs a user preference vector and multiple evaluation task feature vectors based on the basic information and evaluation task information of the target teacher acquired in advance; determines the initial recommendation list according to the user preference vector and the characteristic vectors of each evaluation task; and optimizes the initial recommendation list using a multi-objective optimization model constructed based on task diversity indicators and accuracy indicators to obtain the target recommendation task. Among them, the terminal 102 can be but is not limited to various personal computers, laptops, smart phones and tablet computers. The server 104 can be implemented with an independent server or a server cluster consisting of multiple servers.

[0074] In an exemplary embodiment, Figure 2 As shown, the present application provides a method for recommending evaluation tasks, which is applied to Figure 1 The terminal in is used as an example to illustrate, including the following steps:

[0075] S201, constructing a user preference vector and multiple assessment task feature vectors based on the pre-acquired basic information of the target teacher and the assessment task information.

[0076] Among them, the user preference vector is used to describe the target teacher's preference characteristics in the assessment task. It is constructed based on the target teacher's basic information, which may include the teacher's subject, teaching years, preferred teaching stage, etc. By encoding and quantifying this information in a specific way (such as one-hot encoding), a vector is formed. This vector can mathematically represent the teacher's preference characteristics so that it can be compared and calculated with the assessment task feature vector.

[0077] Assessment task feature vector: used to describe the characteristics of each assessment task. These features are constructed based on assessment task information, which may include the task stage (such as elementary school, junior high school, high school), task subject (such as mathematics, Chinese, science, etc.), task difficulty (such as easy, medium, difficult), and other attributes related to the task content (such as whether it contains specific teaching methods, teaching objectives, etc.). By encoding these task features (such as one-hot encoding), a vector that can represent the assessment task features is formed, which is used for matching calculations with user preference vectors.

[0078] In an embodiment of the present application, a user preference vector and multiple assessment task feature vectors are constructed based on the pre-acquired basic information and assessment task information of the target teacher.

[0079] For the basic information of the target teacher, such as the subject is mathematics, the teaching years are 5 years, and the preferred junior high school teaching stage, this information is encoded. If one-hot encoding is used, for the subject, assuming that there are only three subject options: mathematics, Chinese, and English, then the encoding vector corresponding to the mathematics subject is [1,0,0]; the teaching years of 5 years are encoded as the vector corresponding to the intermediate level [0,1,0] according to the pre-defined binning strategy (such as 1-3 years for elementary, 4-6 years for intermediate, 7-10 years for advanced); the preferred junior high school teaching stage is also encoded similarly, such as [0,1,0] (assuming there are three stages: elementary school, junior high school, and high school). Combining these encoding vectors will get the user preference vector of the target teacher.

[0080] For the assessment task information, take a middle school mathematics assessment task of medium difficulty as an example. The task stage is middle school and is coded as [0,1,0], the task subject is mathematics and is coded as [1,0,0], and the task difficulty is medium and is coded as [0,1,0]. Combining these coding vectors will get the feature vector of the assessment task. This feature vector construction is performed for all assessment tasks.

[0081] S202: Determine an initial recommendation list based on the user preference vector and the feature vectors of each evaluation task.

[0082] Among them, the initial recommendation list is a preliminary ranking list of assessment tasks obtained after preliminary calculation (calculated based on the similarity between the user preference vector and the assessment task feature vector). The assessment tasks in this list are selected from all assessment tasks and sorted according to the degree of match with the teacher's preferences (i.e. similarity). The assessment tasks ranked at the top are considered to be more matched with the teacher's preferences in the preliminary calculation, but this list has not yet been finally optimized.

[0083] An implementation method of an embodiment of the present application: the terminal can generate an affinity graph according to the user preference vector and the feature vectors of each evaluation task, and then obtain an initial recommendation list according to the affinity graph.

[0084] Another implementation method of the embodiment of the present application: the terminal can use a cosine similarity algorithm to calculate the similarity between the user preference vector and the feature vector of each evaluation task; according to each similarity, the evaluation tasks are sorted to determine an initial recommendation list.

[0085] Specifically, the assessment tasks are sorted according to the similarities to determine the initial recommendation list. The similarity values ​​of all the assessment tasks and the target teacher calculated are sorted, and the assessment tasks with higher similarity values ​​are ranked at the front of the initial recommendation list. For example, if the similarity value of assessment task 1 is 0.8 and the similarity value of assessment task 2 is 0.6, then in the initial recommendation list, assessment task 1 will be ranked ahead of assessment task 2.

[0086] S203, based on the pre-built multi-objective optimization model, the initial recommendation list is optimized to obtain the target recommendation task.

[0087] Among them, the multi-objective optimization model is constructed based on the task diversity index and the accuracy index. The multi-objective optimization model is used to simultaneously consider multiple objectives (task diversity index and accuracy index in this application) for optimization. This model aims to balance the two important goals of accuracy and diversity of recommendations. The accuracy index is used to measure the degree of match between the recommended evaluation tasks and the actual needs of teachers, for example, it can be measured by the difference between the predicted score and the actual score (such as the root mean square error). The diversity index is used to measure the degree of difference between the evaluation tasks in the recommendation list, for example, it can be measured by the similarity between the items in the recommendation list (such as the cosine distance based on the project features). By adjusting the parameters in the model (such as the weight parameters), a suitable balance can be found between accuracy and diversity, thereby optimizing the recommendation results and obtaining the final target recommendation task.

[0088] In the embodiment of the present application, in the multi-objective optimization model, the task diversity index uses the cosine distance based on the project characteristics to calculate the diversity between the evaluation tasks in the recommendation list, and the accuracy index uses the root mean square error to measure the accuracy of the recommendation. By adjusting the parameters in the model (such as weight parameters), balance accuracy and diversity. For example, when more attention is paid to accuracy, the weight of the accuracy index in the model can be increased, and vice versa, the weight of the diversity index can be increased.

[0089] Furthermore, according to the optimization results of the model, the assessment tasks in the initial recommendation list are reordered or screened to obtain the final target recommended tasks. These target recommended tasks have high accuracy and can well match the needs of teachers, and have a certain degree of diversity, which helps teachers get in touch with different types of assessment tasks and promotes their professional development.

[0090] The above-mentioned assessment task recommendation method constructs a user preference vector and multiple assessment task feature vectors based on the pre-acquired basic information of the target teacher and the assessment task information; determines the initial recommendation list according to the user preference vector and the feature vectors of each assessment task; and optimizes the initial recommendation list using a multi-objective optimization model constructed based on task diversity indicators and accuracy indicators to obtain the target recommendation task. The present application uses a multi-objective optimization model constructed based on task diversity indicators and accuracy indicators to optimize the initial recommendation list, which can further optimize the recommendation results based on the recommendation accuracy. The model takes into account the two key factors of accuracy and diversity at the same time, avoiding the problem of focusing only on accuracy and resulting in too single recommended tasks, or only pursuing diversity and sacrificing recommendation accuracy.

[0091] In an exemplary embodiment, based on the above embodiment, see Figure 3 The embodiment of the present application relates to a process of determining an initial recommendation list according to a user preference vector and feature vectors of each evaluation task, including the following steps:

[0092] S301, generating a target teacher task affinity diagram based on the user preference vector and the feature vectors of each assessment task.

[0093] Among them, the target teacher task affinity graph is used to characterize the matching degree between the target teacher and multiple assessment tasks. In this graph, the nodes can be divided into two categories, one for the target teacher and the other for the assessment task. The edge between the nodes represents the relationship between the teacher and the assessment task, and the weight of the edge represents the matching degree between them. This matching degree is calculated by the user preference vector and the assessment task feature vector. It comprehensively considers various preference information of teachers (such as subject, teaching years, teaching stage preference, etc.) and various characteristics of assessment tasks (such as task stage, subject, difficulty, etc.), and can intuitively reflect the affinity of each assessment task with the target teacher.

[0094] In the embodiment of the present application, first, for each evaluation task, the relationship between its feature vector and the user preference vector of the target teacher is calculated. Assuming that cosine similarity calculation is adopted, the target teacher is taken as a node, and each evaluation task is taken as a node, the calculated cosine similarity value is used as the weight of the edge between the target teacher node and the evaluation task node, and the target teacher task affinity graph is constructed. For example, assuming that there are three evaluation tasks, the cosine similarities calculated with the target teacher are 0.7, 0.5 and 0.6 respectively, then in the affinity graph, the edge weights between the target teacher node and the three evaluation task nodes are 0.7, 0.5 and 0.6 respectively.

[0095] S302, generating an initial recommendation list based on the target teacher task affinity diagram.

[0096] In the embodiment of the present application, in the constructed target teacher task affinity graph, the evaluation task nodes are sorted according to the edge weight (i.e., the cosine similarity value between the target teacher and the evaluation task). The greater the edge weight, the higher the matching degree between the evaluation task and the target teacher.

[0097] The sorted assessment tasks are arranged in order to form an initial recommendation list. For example, in the affinity graph, the order of the assessment tasks after sorting from large to small according to the edge weight is Task 1, Task 3, Task 2, so the initial recommendation list is [Task 1, Task 3, Task 2]. The assessment tasks in this initial recommendation list will serve as the basis for subsequent optimization to obtain the final recommended tasks, thereby providing the target teachers with assessment task recommendations that better meet their needs.

[0098] The embodiment of the present application can comprehensively consider the complex relationship between the target teacher and multiple assessment tasks by constructing a target teacher task affinity graph. This construction method is not a simple one-to-one matching calculation, but a combination of various characteristics of teachers and assessment tasks, making the calculation of the matching degree more accurate and comprehensive, laying a solid foundation for generating a high-quality initial recommendation list.

[0099] In an exemplary embodiment, based on the above embodiment, see Figure 4 The embodiment of the present application relates to a process of generating a target teacher task affinity graph based on a user preference vector and feature vectors of each assessment task, including the following steps:

[0100] S401, performing node feature initialization processing on the user preference vector and each evaluation task feature vector respectively, to obtain a first feature vector of the teacher node and second feature vectors of multiple evaluation task nodes.

[0101] Among them, the first feature vector is the feature vector related to the teacher after node feature initialization. It is converted from the user preference vector and contains the processed teacher-related feature information. It will be used as the feature input representing the teacher node in constructing the teacher task affinity graph and subsequent graph neural network processing.

[0102] Second eigenvector: The eigenvector related to the assessment task after node feature initialization. It is converted from the assessment task eigenvector and contains the processed assessment task-related feature information. It is used as the feature input of the assessment task node in constructing the teacher task affinity graph and graph neural network processing.

[0103] In the embodiment of the present application, for the user preference vector, for example, the basic information of the target teacher is that the subject is mathematics, the teaching years are 5 years, and the junior high school teaching stage is preferred, after a specific node feature initialization process (which may include coding adjustment, feature extraction and other operations), it is converted into the first feature vector of the teacher node. Similarly, for each assessment task feature vector, such as a junior high school mathematics assessment task of medium difficulty, the feature vector composed of its original task stage, subject, difficulty and other information is obtained after the node feature initialization process to obtain the second feature vector of the assessment task node.

[0104] S402, obtaining an initial teacher task affinity graph based on the distance between the first eigenvector and each second eigenvector.

[0105] The distance in the embodiment of the present application can be a certain defined distance measurement method, such as Euclidean distance or a distance calculated based on a specific rule. The weight of the edge between the teacher node and the evaluation task node in the initial teacher task affinity graph is determined according to the calculated distance. For example, if the first eigenvector is closer to the second eigenvector of a certain evaluation task node, it means that the teacher is initially considered to have a higher correlation with the evaluation task, and the corresponding edge weight may be larger.

[0106] S403, based on the graph neural network, the initial teacher task affinity graph is updated to obtain the target teacher task affinity graph.

[0107] Among them, graph neural network: a neural network that specializes in processing graph structure data. In this application, it is used to optimize the initial teacher task affinity graph. It contains structures such as input layer and update layer, which can update the structure and node features of the graph by learning the feature information of nodes and edges in the graph, so as to obtain a more accurate relationship representation.

[0108] Specifically, the graph neural network includes an input layer and an update layer. Based on the graph neural network, the initial teacher task affinity graph is optimized to obtain the target teacher task affinity graph, including:

[0109] The first eigenvector and each second eigenvector are input into the input layer for feature vector mapping processing to obtain a third eigenvector; the dimension of the third eigenvector corresponds to the dimension of the input layer; the third eigenvector is input into the update layer for feature update processing to obtain an update node; the update node includes a teacher update node and a task update node; according to the teacher update node and the task update node, the edge weight data in the initial teacher task affinity graph is adjusted to obtain a target teacher task affinity graph.

[0110] Furthermore, the above-mentioned edge weight data in the initial teacher-task affinity graph is adjusted and processed according to the teacher update nodes and the task update nodes to obtain the target teacher-task affinity graph, including: calculating the cosine similarity of the teacher update nodes and the task update nodes to obtain the similarity data; determining the recommendation score according to the similarity data; the recommendation score is used to characterize the matching degree between the target teacher and each assessment task; and adjusting the edge weight data according to the recommendation score to obtain the target teacher-task affinity graph.

[0111] Among them, feature vector mapping processing (input layer): the first feature vector and each second feature vector are input into the input layer for feature vector mapping processing to obtain the third feature vector. The dimension of the third feature vector corresponds to the dimension of the input layer. In this process, the input layer transforms the input feature vector through its internal parameters and calculation mechanism, so that the output third feature vector can be better processed by the subsequent part of the network.

[0112] Feature update processing (update layer): The third feature vector is input into the update layer for feature update processing to obtain update nodes. Update nodes include teacher update nodes and task update nodes. The update layer updates the input third feature vector based on its own learning algorithm and parameters, and mines deeper feature information of teachers and assessment tasks, thereby obtaining updated teacher and assessment task nodes.

[0113] Edge weight adjustment processing: Calculate the cosine similarity between the teacher update node and the task update node to obtain the similarity data.

[0114] The recommendation score is determined based on the similarity data. This recommendation score can accurately reflect the degree of match between the target teacher and each assessment task after the graph neural network optimization. The edge weight data is adjusted according to the recommendation score to obtain the target teacher task affinity graph. For example, if the recommendation score is high, it means that the teacher and the assessment task have a high degree of match, then the weight of the corresponding edge in the target teacher task affinity graph will increase, and vice versa. The target teacher task affinity graph obtained in this way can more accurately reflect the relationship between the teacher and the assessment task, and provide a more reliable basis for the subsequent assessment task recommendation.

[0115] In the embodiment of the present application, the initial teacher-task affinity graph is optimized by using a graph neural network, which can deeply mine the characteristic information of teachers and assessment tasks, thereby more accurately representing the relationship between the target teacher and the assessment task. Compared with the affinity graph constructed by traditional methods, the edge weights in the target teacher-task affinity graph can more truly reflect the degree of match between the teacher and the assessment task.

[0116] In an exemplary embodiment, based on the above embodiment, see Figure 5 The embodiment of the present application relates to a process of obtaining an initial teacher task affinity graph according to the distance between the first eigenvector and each second eigenvector, comprising the following steps:

[0117] S501, input the distance between the first eigenvector and each second eigenvector into a distance function to perform distance calculation processing to obtain a distance result. The distance result is used to represent the distance between the teacher node and each evaluation task node.

[0118] Edge weight data: Data obtained after the distance result is calculated and processed by the edge weight function. It is used to characterize the affinity between the teacher node and the assessment task node. The larger the edge weight data, the higher the affinity between the teacher and the assessment task, that is, the closer the relationship between the two in the initial teacher-task affinity graph.

[0119] In the embodiment of the present application, the terminal inputs the distance between the first eigenvector and each second eigenvector into the above distance function to perform distance calculation processing to obtain a distance result. This distance result can quantify the distance between the teacher node and the evaluation task node in the feature space. For the second eigenvector of each evaluation task node, the distance calculation is performed with the first eigenvector in this way to obtain the corresponding distance results.

[0120] S502, input the distance result into the edge weight function for calculation and processing to obtain edge weight data; the edge weight data is used to characterize the affinity between the teacher node and the assessment task node.

[0121] In the embodiment of the present application, the edge weight function can design different calculation rules according to actual needs. When the distance result is small, the edge weight data is large, indicating that the affinity between the teacher node and the assessment task node is high; conversely, when the distance result is large, the edge weight data is small, indicating that the affinity is low. In this way, each distance result is converted into corresponding edge weight data.

[0122] S503, construct an initial teacher task affinity graph based on edge weight data.

[0123] In an embodiment of the present application, the teacher node is taken as a central node, each evaluation task node is taken as other nodes, and the edge weight data is taken as the weight of the edge connecting the teacher node and the evaluation task node. For example, if there are three evaluation task nodes, and the corresponding edge weight data are w1, w2, and w3 respectively, then edges are constructed between the teacher node and these three evaluation task nodes, and the edge weights are w1, w2, and w3 respectively, thereby constructing an initial teacher task affinity graph. This initial teacher task affinity graph preliminarily characterizes the relationship between the target teacher and each evaluation task, and provides a basic relationship structure for subsequent further optimization (such as optimization through graph neural network).

[0124] The embodiment of the present application constructs an initial teacher-task affinity graph by converting the characteristic vector distance between the teacher and the assessment task into edge weight data, which can accurately quantify the relationship between the teacher and the assessment task. This quantification method is based on the distance calculation of the characteristic vector. Compared with the traditional simple association method, it can more carefully reflect the differences and connections between teachers and assessment tasks in many aspects of characteristics, and provide a more accurate initial relationship basis for subsequent assessment task recommendations.

[0125] In an exemplary embodiment, based on the above embodiment, see Figure 6 The embodiment of the present application relates to a process of generating an initial recommendation list according to a target teacher task affinity graph, comprising the following steps:

[0126] S601, calculating the affinity between the teacher node and each assessment task node in the target teacher task affinity graph to obtain multiple affinity scores.

[0127] Among them, the affinity score is the score obtained by calculating the affinity between the teacher node and each assessment task node in the target teacher task affinity graph. It quantifies the affinity between the teacher and each assessment task and is an important basis for determining the ranking position of the assessment task in the initial recommendation list.

[0128] In the embodiment of the present application, the calculation of affinity can be based on the weight of the edge in the target teacher task affinity graph and other relevant information (such as the attributes of the node). For example, if the weight of the edge directly represents the affinity, then the weight value of the edge can be directly obtained as the affinity score. Alternatively, if other factors are taken into account, such as the comprehensive relationship between the feature vector of the node and the edge weight, a custom calculation method can be used.

[0129] S602, sorting multiple evaluation tasks according to multiple affinity scores to obtain an initial recommendation list.

[0130] In an embodiment of the present application, the multiple affinity scores calculated are sorted in order from high to low (or from low to high, according to the definition of affinity score, if the higher the score, the higher the affinity, then sort from high to low). For example, if there are three assessment tasks, and the corresponding affinity scores are 0.8, 0.6, and 0.7, respectively, then the order obtained after sorting from high to low is assessment task 1, assessment task 3, and assessment task 2, and the assessment tasks corresponding to this order form an initial recommendation list, that is, [assessment task 1, assessment task 3, assessment task 2]. The assessment tasks in this initial recommendation list will serve as the basis for subsequent further optimization (such as through methods such as multi-objective optimization models) to ultimately obtain recommended tasks suitable for target teachers.

[0131] The embodiment of the present application generates an initial recommendation list by calculating and sorting the affinity between the teacher and the assessment task in the target teacher task affinity graph, which can effectively perform a preliminary screening of the assessment tasks based on the affinity relationship between the teacher and the assessment task. This affinity-based screening method makes the assessment tasks in the initial recommendation list more likely to meet the needs of the teacher, providing a good initial foundation for subsequent optimization recommendations.

[0132] More specifically, in order to further reduce the impact of cold start on the accuracy of assessment task recommendation, a "teacher-task affinity graph" is established, which combines the concepts of knowledge graph and graph network to predict possible preferences by associating the potential characteristics of teachers and the intrinsic properties of assessment tasks. Algorithm model: A graph-based neural network is constructed, in which nodes represent teachers and assessment tasks, and edges represent potential associations.

[0133] How it works: The algorithm learns the underlying connections between teachers and assessment tasks, and can make effective recommendations even in the absence of historical interaction data.

[0134] Algorithm model creation steps:

[0135] 1. Node feature initialization.

[0136] The characteristics of the teacher node are composed of their teaching field, teaching experience, etc., which can be encoded through embedding technology. The characteristics of the assessment task node are composed of task type, target stage, difficulty level, etc., which are also encoded through embedding technology.

[0137] 2. Edge weight calculation.

[0138] For each possible teacher and assessment task pair, a preliminary affinity weight is calculated, which can be based on the degree of match between their shared features.

[0139] The algorithm formula is as follows:

[0140]

[0141] Among them, j is the weight of the edge between teacher i and task j, γ is the scaling parameter, d() is the distance function between features, and f teacher and f taskj are the feature vectors of teacher i and task j respectively.

[0142] 3. Graph Neural Network,Formula (2) uses the graph network to update node features and integrate the affinity information between teachers and tasks.

[0143]

[0144] Previous node feature f i By its adjacent node features f j The update is performed by a weighted combination of , where w is the weight matrix and σ is the nonlinear activation function.

[0145] 4. Recommendation generation formula (3):

[0146]

[0147] The final recommendation score is calculated using the updated teacher node features and task node features.

[0148] For example, teacher A’s specialty is algebra, and the eigenvector is

[0149] Task 1 requires algebraic skills, eigenvectors

[0150] Calculate affinity, the affinity w between teacher A and task 1 A1 will be 1, since they match exactly.

[0151] Theoretically, a 100% recommendation success accuracy rate can be achieved.

[0152] In actual practice, the actual results will fluctuate within the theoretical range.

[0153] The matching formula (4) between teacher A and task 1 is as follows:

[0154]

[0155] This means that teacher A has the highest matching degree for task 1, so the recommendation score is A1 =1.

[0156] When a set of teachers and tasks are matched, all our matches are very close to perfect matches, such as teacher B's match w for task 2 B2 =0.98, indicating an almost perfect match.

[0157] If we calculate affinities for all teachers and tasks, and all affinities are close to 1, we can say that the average accuracy of the recommender system is very high.

[0158] By adding the "teacher-task affinity graph" model to the cold start problem, the accuracy data of the results is summarized from the obtained data verification. In the absence of a large amount of historical interaction data, the recommendation system of the "teacher-task affinity graph" can provide highly accurate recommendations. The efficiency of this method lies in the ability to quickly generate recommendations by matching expertise and needs, while its accuracy is reflected in the ability to capture and utilize detailed professional associations between teachers and tasks. This is because the cold start problem itself is caused by the lack of a large amount of historical interaction data.

[0159] In an exemplary embodiment, based on the above embodiment, see Figure 7 The embodiment of the present application relates to a process of constructing a user preference vector and multiple evaluation task feature vectors based on pre-acquired basic information of a target teacher and evaluation task information, including the following steps:

[0160] S701, performing a first feature encoding process on the basic information to obtain a user preference vector.

[0161] Among them, the first feature encoding process: This is an encoding operation for the basic information of the target teacher. It converts the basic information of the teacher (such as subject, teaching years, preferred teaching stage, etc.) into a vector form according to specific rules to obtain the user preference vector. This encoding process may involve operations such as data type conversion and feature quantization in order to accurately represent the teacher's preference characteristics.

[0162] In the embodiment of the present application, a first feature encoding process is performed on the basic information to obtain a user preference vector.

[0163] For example, for the subjects in the basic information of the target teacher, if there are three subject types, namely mathematics, Chinese and English, the first feature encoding can be processed by one-hot encoding. Assume that the mathematics subject is coded as [1,0,0], Chinese is [0,1,0], and English is [0,0,1]. For the teaching years, if the encoding is carried out according to the rule that 1-3 years are elementary (coded as [1,0,0]), 4-6 years are intermediate (coded as [0,1,0]), and 7-10 years are advanced (coded as [0,0,1]). For the preferred teaching stage, such as elementary school, junior high school, and high school, one-hot encoding can also be used. For example, the junior high school stage is coded as [0,1,0]. Then these codes are combined to obtain the user preference vector of the target teacher. For example, a teacher who teaches mathematics, has intermediate teaching years, and prefers the junior high school teaching stage may have a user preference vector of [1,0,0,0,1,0,0,1,0].

[0164] S702, performing a second feature encoding process on the evaluation task information to obtain a feature vector of each evaluation task.

[0165] Among them, the second feature encoding process: This is an encoding operation performed on the assessment task information. It converts the assessment task information (such as the task stage, subject, difficulty, etc.) into a vector form according to specific rules to obtain the feature vector of each assessment task. This encoding process helps to quantify the various features of the assessment task in a unified manner, which facilitates the subsequent calculation related to the user preference vector.

[0166] In the embodiment of the present application, the evaluation task information is subjected to a second feature coding process to obtain a feature vector of each evaluation task. Taking an evaluation task as an example, it is assumed that the task stage is junior high school, the task subject is mathematics, and the task difficulty is medium. For the task stage, a unique hot coding is used, and the junior high school can be encoded as [0,1,0] (assuming that the elementary school is [1,0,0] and the high school is [0,0,1]); for the task subject, the mathematics subject is encoded as [1,0,0] (assuming that Chinese is [0,1,0] and English is [0,0,1]); for the task difficulty, the medium difficulty is encoded as [0,1,0] (assuming that easy is [1,0,0] and difficult is [0,0,1]). These codes are combined to obtain the feature vector of the evaluation task [0,1,0,1,0,0,0,1,0]. For each evaluation task, this second feature coding processing method is used to obtain the respective evaluation task feature vectors.

[0167] In the embodiment of the present application, by performing the first feature coding process and the second feature coding process on the basic information of the target teacher and the assessment task information, various types of information can be accurately quantified into vector forms. This accurate quantification method helps to more accurately reflect the teacher's preferences and the characteristics of the assessment task in subsequent calculations (such as calculating the matching degree between the teacher and the assessment task, etc.).

[0168] In an exemplary embodiment, based on the above embodiment, see Figure 8 The basic information of the embodiment of the present application includes discrete features and continuous features; performing a first feature encoding process on the basic information to obtain a user preference vector includes the following steps:

[0169] S801, performing data conversion processing on continuous features according to a pre-defined binning strategy to obtain categorical features.

[0170] Among them, discrete features: This is a data type in the basic information of the target teacher, and its value is a discrete, finite category. For example, the teacher's subject (such as mathematics, Chinese, English, etc.), the preferred teaching stage (such as elementary school, junior high school, high school), etc. The values ​​of these features are independent and limited categories, and there is no intermediate transition value.

[0171] Continuous feature: This is another data type in the target teacher's basic information. Its value is a numerical value that changes continuously within a certain range. For example, the teacher's teaching years can be a specific number, such as 1 year, 3.5 years, 5 years, etc. The value range is a continuous interval, and theoretically any real number within this interval can be taken.

[0172] Binning strategy: This is a method of data conversion processing on continuous features, the purpose is to convert continuous features into categorical features. Binning strategies include equal width binning, equal frequency binning and custom binning.

[0173] Equal-width binning: Divide the value range of a continuous feature into several intervals of equal width, and each interval is a category. For example, for the continuous feature of teaching years, if the value range is 0-10 years, equal-width binning can divide it into 0-2 years, 2-4 years, 4-6 years, 6-8 years, and 8-10 years, and each interval is a category.

[0174] Equal frequency binning: Divide continuous features into several intervals according to the distribution frequency of the data, so that the number of data in each interval is roughly equal. For example, if the teaching years of teachers are divided into equal frequency bins, several intervals may be divided according to the distribution of the teaching years of all teachers, and each interval contains roughly the same number of teachers.

[0175] Custom binning: Artificially define the intervals for continuous features based on specific business needs or data characteristics. For example, the teaching years can be divided into intervals based on the school's specific policies or the special definition of the teacher's development stage.

[0176] Categorical features: This is the type of feature obtained after the continuous feature is converted through the binning strategy. It is a discretized feature that converts the continuous value range into a finite number of categories for subsequent coding. For example, the categorical features of the teaching years after the binning strategy conversion may be "primary (0-3 years)", "intermediate (3-6 years)", "advanced (6-10 years)" and other categories.

[0177] In the embodiment of the present application, data conversion processing is performed on continuous features according to a pre-defined binning strategy to obtain classification features.

[0178] Specifically, assume that the continuous feature in the basic information of the target teacher is the teaching years, and the value range is 0-10 years. If the equal-width binning strategy is adopted and the bin width is set to 3 years, the classification features can be obtained: "0-3 years", "3-6 years", "6-9 years", "9-10 years". If the equal-frequency binning strategy is adopted, it is assumed that according to the data distribution of the teaching years of all teachers, it is divided into four intervals so that the number of teachers in each interval is roughly equal, and the corresponding classification features can also be obtained. If a custom binning strategy is adopted, for example, according to the school's division of the years of teacher title evaluation, the teaching years are divided into classification features such as "0-2 years (primary)", "2-5 years (early intermediate)", "5-8 years (late intermediate)", and "8-10 years (senior)".

[0179] S802, performing one-hot encoding processing on the categorical features and the discrete features respectively to obtain a user preference vector.

[0180] In the embodiment of the present application, the categorical features and the discrete features are respectively processed by one-hot encoding to obtain a user preference vector.

[0181] For example, the discrete features of the target teacher are that the subject is mathematics, the preferred teaching stage is junior high school, and the continuous feature (converted into categorical features) is that the teaching years are "3-6 years (intermediate)". For the subject "mathematics", assuming that the subject has three options: mathematics, Chinese, and English, its one-hot encoding is [1,0,0]; for the teaching stage "junior high school", assuming that the teaching stage has three options: elementary school, junior high school, and high school, its one-hot encoding is [0,1,0]; for the teaching years "3-6 years (intermediate)", assuming that the classification of teaching years has three options: "0-3 years (primary)", "3-6 years (intermediate)", and "6-10 years (advanced)", its one-hot encoding is [0,1,0]. Combining these one-hot encodings, the user preference vector is [1,0,0,0,1,0,0,1,0].

[0182] The embodiment of the present application can effectively process the discrete features and continuous features in the basic information of the target teacher. The continuous features are converted into categorical features through the binning strategy, and then the unique hot encoding is uniformly performed, so that different types of features can be converted into vector form in an appropriate manner, so that the user preference vector can be constructed more comprehensively and accurately.

[0183] In an exemplary embodiment, based on the above embodiment, see Fig. 9 , the evaluation task information of the embodiment of the present application includes multiple operation types; the evaluation task information is subjected to a second feature encoding process to obtain a feature vector of each evaluation task, including:

[0184] S901, perform one-hot encoding processing on each operation type to obtain a binary vector.

[0185] Among them, binary vector: This is the vector form obtained after one-hot encoding, and its elements only contain 0 and 1. Each binary vector corresponds to an operation type. This vector form can intuitively indicate whether a certain operation type exists in the evaluation task.

[0186] In the embodiment of the present application, each operation type is processed by one-hot encoding to obtain a binary vector. Assume that the operation types in the assessment task information include four types of operation: answering multiple-choice questions, answering short-answer questions, practical operations, and group collaboration. For answering multiple-choice questions, after one-hot encoding, a binary vector [1,0,0,0] is obtained; the binary vector corresponding to answering short-answer questions is [0,1,0,0]; the corresponding binary vector for practical operations is [0,0,1,0]; and the corresponding binary vector for group collaboration is [0,0,0,1]. In this way, each operation type is converted into a 4-dimensional binary vector.

[0187] S902, performing feature extraction processing on the binary vector to obtain a feature set.

[0188] In the embodiment of the present application, feature extraction is performed on the binary vector obtained above. The number in each binary vector can be counted to obtain a eigenvalue. For [1,0,0,0], the number is 1; for [0,1,0,0], the number is also 1; for [0,0,1,0], the number is 1; for [0,0,0,1], the number is 1. At the same time, other features such as the position in the vector can also be analyzed. These eigenvalues ​​(such as the number, the position, etc.) are combined to form a feature set. Assume that this feature set is {"the number of 1 is 1", "the positions of 1 are 1, 2, 3, 4 respectively"}.

[0189] S903, perform feature conversion processing on the feature set to obtain feature vectors of each evaluation task.

[0190] In an embodiment of the present application, the features in the feature set are quantized. The feature "the number of 1s is 1" is quantized to the value 1. For the feature "the positions of 1s are 1, 2, 3, and 4 respectively", an encoding method can be used, such as converting it to [1, 1, 1, 1] (indicating that it has appeared in all four positions). These quantized features are then combined to obtain an evaluation task feature vector, such as [1, 1, 1, 1]. Each evaluation task is processed according to these steps to obtain its own evaluation task feature vector.

[0191] The embodiment of the present application can fully utilize the operation type information to construct the evaluation task feature vector by performing unique hot encoding, feature extraction and feature conversion on multiple operation types in the evaluation task information. Compared with the traditional method, it can more comprehensively reflect the characteristics of the evaluation task and provide richer information for accurately evaluating the compatibility of the evaluation task with the teacher.

[0192] More specifically, in order to solve the data sparsity problem and extract teacher ability features from the provided data, we can adopt the following mathematical modeling approach. This process involves encoding discrete data, feature engineering, and finally feature extraction to form a feature set that can be used for further analysis or modeling.

[0193] For the discrete features in the basic information of teachers (such as gender, ethnicity, religious beliefs, political affiliation, native place, age, grade, subject, position, left-click-button, drag-option, right-click-page, left-click-text box, left-click-text box, left-click-button, scroll wheel click-web page, scroll wheel click-text, input-text box, cursor-picture, cursor-text, left-click-area, etc.), we use one-hot encoding for processing.

[0194] It should be noted that the one-hot encoding used in the processing is used to process discrete feature data, not continuous features. One-hot encoding converts each category value of a categorical variable into a new binary column, which is suitable for processing categorical data, such as gender, ethnicity, etc. However, for continuous features, directly applying one-hot encoding is usually not a suitable method, because continuous features may contain a large number of unique values, and direct one-hot encoding will result in too high a feature dimension and be difficult to process.

[0195] When dealing with continuous features, the actual binning technique is used, which indirectly uses one-hot encoding to process continuous features. Binning involves dividing the continuous variable into multiple intervals, each interval representing a "bucket" or "box", and then one-hot encoding can be applied to these buckets. In this way, the continuous features are converted into a limited number of discrete categories, each category corresponding to a specific interval. This method retains a certain amount of continuity information and is easy to process by the model.

[0196] Here are the steps to process continuous features through binning and one-hot encoding:

[0197] Step 1: Determine the binning strategy: Equal width binning: Divide the range of continuous features into several segments of equal width. Equal frequency binning: Divide continuous features into several segments, each containing the same number of observations. Custom binning: Customize segment boundaries based on business understanding or feature distribution.

[0198] Step 2, apply binning: Use tools such as pandas.cut or pandas.qcut to bin the data. The pandas library in Python provides this function.

[0199] Step 3: One-hot encoding application: Convert the binning results into one-hot encoding format. Each bin corresponds to a feature column, where a value of 1 indicates that the sample is in the bin, and 0 indicates that it is not.

[0200] The following features usually describe the operating behaviors of teachers when using a certain system or software. These data may be collected in specific situations to analyze user behavior or interaction habits: left-click-button, drag-option, right-click-page, left-click-text box (this item is repeated and should only be counted once), scroll wheel click-web page, scroll wheel click-text, input-text box, cursor-picture, cursor-text, left-click-area.

[0201] For each feature x i , where i represents the index of the feature. If feature x i There are i different category values, the feature will be converted into a length of n i A binary vector of .

[0202] The task assessment feature vectorization of teacher micro-ability assessment (task assessment feature vectorization is the feature operation vectorization below). In the situational assessment task, the behavior is encoded and extracted, and the feature operations such as left-click-button, drag-option, right-click-page, left-click-text box, left-click-text box, left-click-button, scroll wheel click-web page, scroll wheel click-text, input-text box, cursor-picture, cursor-text, left-click-area are vectorized. We regard each operation type as a feature and convert the frequency or existence of the operation into a numerical value to construct a feature vector.

[0203] In the teacher micro-competency assessment task, feature vectorization, encoding, and extraction are important steps in data processing, which are used to convert the original user behavior data into a format that can be understood and processed by the machine learning model. The following is a detailed explanation of these concepts and the order of execution:

[0204] For the processing of behavioral data, the usual order is:

[0205] Step 1: Encoding. First, encode the behavioral data to convert each discrete operation type into a numerical representation. This includes determining whether to use frequency counts or one-hot encoding to represent each operation.

[0206] Step 2: Extraction: Useful features are extracted from the encoded data, such as counting the frequency of each operation or constructing features about the operation context (such as the time the operation takes in the session).

[0207] Step 3: Vectorize. Convert the extracted feature set into a whole feature vector. If one-hot encoding is used, it may have been partially vectorized; if it is frequency or other numerical values, the final feature vector will contain all these numerical features.

[0208] The whole process converts the teacher's behavioral operations into numerical feature vectors that can be directly processed by machine learning algorithms, which is necessary for further analysis, pattern recognition, or predictive model building.

[0209] Regarding the relationship between eigenvalue and operating frequency, there are two points that need to be explained:

[0210] First, operation frequency: refers to the number of times a specific operation occurs in a given time period or task. For example, if the teacher left-clicked the button 5 times, dragged the option 2 times, and right-clicked the page 1 time in completing a task, then the frequency of these operations can be directly recorded and used as a feature vector:

[0211] Feature vector: [5,2,1]. This method is simple and direct, and is very suitable for application scenarios that need to quantify the frequency of user interaction.

[0212] Second, operation existence: focuses on whether the operation occurred, not how many times it occurred. This is often used in scenarios where it is important to mark whether an action was performed, not how many times it was performed. For example, if we only care about whether an operation was performed, not the frequency of the operation.

[0213] In machine learning modeling, using these feature vectors, classifiers or regression models can be trained to predict user behavior, completion efficiency, or other important output variables. Analyzing operation frequency and presence can reveal user habits, preferences, and possible areas for improvement.

[0214] Given a sequence of operations O = {o1, o2, ..., o m}, where m represents the total number of operation types, operation O j The frequency f j It can be expressed as formula (5):

[0215]

[0216] Where a represents the operation type, b represents the number of times the characteristic operation occurs, and c represents the total number of operations. Oj is the operation instance. represents the specific j-th operation instance. In other words, if a is a general identifier such as "left-click button", "drag option", etc., then Oj is a specific instance of these operation types.

[0217] Explain the specific role of Oj:

[0218] First, operation instance: In a specific context, Oj can refer to a specific operation event, such as a single behavior instance of a teacher using software for assessment. Here, j is used to index these single events.

[0219] Second, an element in the data set: When collecting data, each Oj can represent a row or a record in the data set, which specifically records the detailed information of a certain operation, such as the operation type, operation time, and the object of the operation.

[0220] Third, the components of the feature vector: If each operation type a is further subdivided or concretized in the analysis, Oj may refer to the numerical representation of a certain dimension or characteristic corresponding to the specific operation a.

[0221] Application scenario examples:

[0222] The teacher's behavior data includes various types of operations, such as clicking, dragging, etc. Each operation a represents an operation type, and each Oj records the specific circumstances of each operation in detail. For example: a = "left click button".

[0223] O1 may represent: In Task 1, the teacher left - clicked the "Submit" button at time t1.

[0224] O2 may represent: In Task 2, the teacher left - clicked the "Next Page" button at time t2.

[0225] When processing these data, the frequency of each operation type a occurring within a given time period can be calculated to construct a feature vector, where the frequency of each Oj becomes part of the vector.

[0226] Therefore, if a is the type of operation, Oj generally refers to a specific instance or event of that operation type. This distinction is very important when performing tasks such as data analysis, feature engineering, and machine learning model training, because it helps us understand the structure and content of the data in detail, thus processing and analyzing the data more precisely.

[0227] For the feature extraction and fusion of high - dimensional sparse data in teacher evaluation data, after encoding and vectorizing all teachers' basic information and user operation information, we fuse all the obtained feature vectors into a comprehensive feature matrix X, where each row of the matrix represents a teacher and each column represents a feature. For the comprehensive feature matrix X, we use principal component analysis (PCA) for dimensionality reduction. Let X be an M * N matrix, where M is the number of samples (number of teachers) and N is the number of features. The goal of PCA is to find an N * K transformation matrix W (where K < N) such that the variance of XW is maximized. The transformed feature matrix X' can be expressed as in Equation (6):

[0228] X′ = XW (6)

[0229] where the column vectors of W are the eigenvectors corresponding to the top K largest eigenvalues of the covariance matrix of X.

[0230] At this time, for feature extraction, dimensionality reduction operations on high - dimensional data are required to explore low - dimensional feature data, and feature extraction and data dimensionality reduction are performed through autoencoders. Autoencoders attempt to learn a compressed data representation (encoder part) through a neural network, and then reconstruct the original data through the following formula (decoder part). The encoder and decoder can be trained by minimizing the reconstruction error. Given the input x, the encoded representation is h, and the reconstructed input is x', as shown in Equations (7) and (8):

[0231] h = σ(W e x + b e ) (7〕

[0232] x′ = σ(W a h + b a ) (8

[0233] where, W eand W d are the weights of the encoder and decoder, b e and b d is the bias term and σ is the activation function.

[0234] Through the above steps, we can effectively process and utilize the sparse data in teacher basic information, assessment task information, and user operation information to extract meaningful data to provide support for subsequent recommendations and analysis.

[0235] 2) Recommendation based on the teacher’s basic information and data, solution to the cold start problem: To solve the cold start problem, that is, how to make effective recommendations in the absence of sufficient teacher user or situational assessment task data, specific strategies and algorithms are needed. Here we propose a solution that combines content recommendation and teacher user or situational assessment task feature embedding to reduce the impact of cold start through a mathematical model. The cold start problem is directly related to the feature matrix generated in the previous step. In the recommendation system, the feature matrix is ​​a key tool for describing users (such as teachers) and items (such as assessment tasks). When it comes to new teachers or new assessment tasks, the main reason for the cold start problem is that these new entities lack sufficient data in the feature matrix, making it difficult to make effective recommendations.

[0236] Content recommendation relies on the attributes or descriptions of items to find items that teacher users may be interested in, and does not directly rely on human-computer interaction in the contextual assessment scenario activities of teacher users. Given a teacher user u and a contextual assessment task i, we can make recommendations by calculating the similarity between the teacher user preference vector Pu and the contextual assessment task feature vector qi. The teacher user preference vector can be obtained by scoring a series of contextual assessment task operation attributes, while the contextual assessment task feature vector comes from the description or attributes of a specific operation.

[0237] When constructing the teacher user preference vector and the situational assessment task feature vector, the scoring of operational attributes usually involves the analysis of historical data and the quantitative processing of these data. The operational attributes here can include the teacher's behavior pattern in the assessment task, the frequency of interaction, the preference for multiple-choice questions, etc. These scores are usually obtained based on the teacher's past behavior and feedback.

[0238] First, collect rating data: Rating data usually comes from two main sources:

[0239] Direct feedback: After completing the assessment task, teachers may be asked to directly evaluate the difficulty, relevance, or satisfaction of the task. These direct evaluations can be used as the basis for constructing the preference vector.

[0240] Indirect feedback: Infer teachers’ preferences for tasks by analyzing their interaction data with assessment tasks (e.g., time to complete a task, accuracy, number of retries, etc.). For example, if a teacher frequently selects and successfully completes a certain type of assessment task, it can be inferred that he / she has a higher preference for this type of task.

[0241] Second, processing and conversion of scores: Converting the collected data into scores usually involves the following steps:

[0242] Quantitative processing: Convert teachers' direct feedback and indirect behavioral data into numerical scores.

[0243] Weighted treatment: Different types of feedback or behaviors are given different weights based on their importance, such as a teacher's rating of task satisfaction may be more important than completion speed.

[0244] Third, construct preference and feature vectors:

[0245] Teacher preference vector pu: combines the teacher's rating and other characteristics (such as teaching experience, professional background, etc.) to form a comprehensive preference description vector.

[0246] Assessment task feature vector qi: Each assessment task constructs a feature vector based on its attributes (such as task type, difficulty, target stage) and the teacher’s average score.

[0247] Fourth, the relationship between recommendation score and rating:

[0248] In the recommendation system, the similarity between the teacher's preference vector and the feature vector of the assessment task is usually used to calculate the recommendation score. This recommendation score refers to the interest or preference of a given teacher in a specific assessment task, which is usually calculated using a metric such as cosine similarity. The similarity between pu and qi is used to sort or select the assessment tasks that are most suitable for the teacher.

[0249] Therefore, the scores of the operational attributes of the assessment tasks are obtained from the teacher's behavior and feedback. These scores not only reflect the teacher's direct feedback on the assessment tasks, but also include information indirectly obtained from their interactive behaviors. These scores are directly related to the recommendation scores because they are used to construct vectors that reflect the teacher's preferences and task characteristics. These vectors are at the core of the recommendation system's decision-making process. In this way, the recommendation system can more accurately match the teacher's needs and assessment tasks, thereby providing personalized recommendations.

[0250] The similarity can be calculated using cosine similarity, as shown in formula (9):

[0251]

[0252] In order to solve the cold start problem, a machine learning model can be used to learn low-dimensional dense feature embeddings of teacher, expert user (expert actually refers to excellent teachers) and contextual assessment task features. These embeddings can capture the complex relationships between teachers, expert users and contextual assessment tasks, and can be used to predict unknown preferences. Given a teacher user u and a contextual assessment task i, the user embedding is eu, the contextual assessment task embedding is, and the recommendation score can be calculated by the dot product of the embeddings as shown in formula (10):

[0253] score(u,i)=e u ·e i (10)

[0254] Explanation of recommendation score: In this system, the recommendation score is used to quantify the attractiveness or relevance of an item (such as a contextual assessment task) to a user (such as a teacher). When we talk about contextual assessment task embedding ei and calculating the recommendation score through the dot product of the embedding, our goal is to evaluate the suitability of a specific task for a specific teacher. The higher the recommendation score, the better the task matches the teacher's interests and needs. When faced with new teachers or new assessment tasks (i.e., the cold start problem), there is a lack of sufficient data to directly calculate similarities or generate recommendation scores. Content-based recommendations can also be used to generate recommendations using descriptive features of users or items (such as the teacher's teaching background or the subject area of ​​the task). Recommendations are generated by matching the teacher's expertise and the requirements of the task, rather than relying on historical interaction data.

[0255] Through these methods, the recommendation system can still make effective recommendations even with less data, helping new teachers find suitable assessment tasks or finding suitable teachers for new assessment tasks. These strategies alleviate the impact of the cold start problem through a variety of technical means, making the recommendation system more robust and effective.

[0256] These embeddings can be learned by optimizing the following objective optimization model (Note: By optimizing the objective function, the optimal user embedding eu and task embedding ei can be obtained. This objective optimization model is commonly used in machine learning, especially in the matrix decomposition technology of recommendation systems), the objective optimization model is shown in the following formula and aims to minimize the error between the predicted score and the actual interaction (such as rating), as shown in formula (11):

[0257]

[0258] Where D is the set of human-computer interactions among known teachers, expert users, and situational assessment tasks, y ui is the actual evaluation score of situational evaluation task i performed by teacher user u, and λ is the regularization parameter.

[0259] In the context of recommender systems, when we refer to the set D as “the set of known interactions between teachers, expert users and contextual evaluation tasks”, this usually means that it contains all recorded interaction data between users (teachers and experts) and tasks.

[0260] In an exemplary embodiment, based on the above embodiment, see Fig.10 The embodiment of the present application relates to a process of optimizing an initial recommendation list based on a pre-built multi-objective optimization model to obtain a target recommendation task, including the following steps:

[0261] S1001, using an adaptive step gradient algorithm to adjust the learning rate of the multi-objective optimization model to obtain a target weight parameter.

[0262] Among them, the adaptive step gradient algorithm is used to adjust the learning rate in the multi-objective optimization model, and then adjust the weight parameters. This algorithm dynamically adjusts the step size (i.e., learning rate) according to the current gradient. Compared with the fixed step size algorithm, it can search more effectively in the direction of the optimal solution, avoiding the problem of slow convergence or convergence to the local optimal solution during the optimization process.

[0263] Learning rate: In optimization algorithms (such as adaptive step gradient algorithms), the learning rate is an important parameter. It determines the update step size of the weight parameters at each iteration. A larger learning rate may allow the algorithm to quickly approach the optimal solution during the optimization process, but it may also cause the optimal solution to be missed and oscillate near the optimal solution; a smaller learning rate can search for the optimal solution more finely, but it may lead to too slow convergence.

[0264] In the embodiment of the present application, the adaptive step gradient algorithm dynamically adjusts the learning rate according to the current gradient situation in each iteration, so as to update the weight parameters more effectively. For example, during the optimization process, if the current gradient is found to be large, it means that the objective function changes rapidly in the weight parameter space, and the adaptive step gradient algorithm may appropriately reduce the learning rate to avoid missing the optimal solution when updating the weight parameters; if the current gradient is small, it means that it is close to the optimal solution area, and the algorithm may appropriately increase the learning rate to speed up the convergence speed. In this way, the learning rate is continuously adjusted and the weight parameters are updated to finally obtain the target weight parameters.

[0265] S1002: Adjust the initial recommendation list according to the target weight parameter to obtain the target recommendation task.

[0266] In the embodiment of the present application, the obtained target weight parameters are substituted into the multi-objective optimization model to re-evaluate and sort the assessment tasks in the initial recommendation list. For example, the proportion of accuracy and diversity in the assessment of the assessment tasks is adjusted according to the target weight parameters, and the assessment tasks that better meet the needs of the target teachers (in terms of multiple objectives such as accuracy and diversity) are ranked first, thereby obtaining the target recommended tasks.

[0267] The embodiment of the present application adjusts the weight parameters of the multi-objective optimization model through an adaptive step gradient algorithm, and can dynamically adjust the learning rate according to the actual situation during the optimization process, thereby more effectively updating the weight parameters. This adaptive method avoids the problem of slow convergence speed or convergence to a local optimal solution that may occur in the traditional fixed step method, and improves the efficiency and accuracy of weight parameter optimization.

[0268] In an exemplary embodiment, based on the above embodiment, see Fig.11 The embodiment of the present application involves using an adaptive step gradient algorithm to adjust the learning rate of a multi-objective optimization model to obtain a target weight parameter, including:

[0269] S1101, obtaining initial weight parameters.

[0270] Among them, the initial weight parameter: This is the weight parameter set by the multi-objective optimization model before starting iterative optimization. These initial weight parameters may be obtained based on experience or some preliminary settings, but they are not necessarily the most suitable parameters for optimizing the model to obtain the best recommendation task, and need to be adjusted through the adaptive step gradient algorithm.

[0271] In the embodiment of the present application, the initial weight parameters are first obtained. These initial weight parameters are the initial settings when the multi-objective optimization model starts to optimize. For example, for a multi-objective optimization model that considers both accuracy and diversity, the initial weight parameters may be set to a weight of 0.5 for the accuracy target and a weight of 0.5 for the diversity target, indicating that the two targets are given equal importance in the initial state. These initial weight parameters can be set based on experience, prior knowledge, or some preliminary experimental results.

[0272] S1102, in each iterative solution process, calculating the current gradient of the multi-objective optimization model.

[0273] Among them, the current gradient: In the multi-objective optimization model, the current gradient is the derivative of the objective function with respect to the weight parameter, which indicates the rate of change of the objective function under the current weight parameter value. By calculating the current gradient, we can understand the direction of change of the objective function in the weight parameter space, thus providing a basis for updating the weight parameter.

[0274] In the embodiment of the present application, for example, it is assumed that the objective function of the multi-objective optimization model is J(x), where x is a weight parameter vector. The current gradient is calculated by the derivation formula. This current gradient reflects the rate of change of the objective function under the current weight parameter value, and provides directional information for updating the weight parameter.

[0275] S1103, updating the initial weight parameters according to the current gradient and the fixed step size to obtain the intermediate weight parameters.

[0276] Among them, fixed step size: when using the gradient algorithm to update the weight parameters, the fixed step size is a pre-set constant used to determine the step size each time the weight parameters are updated. In this application, the fixed step size is used together with the current gradient to update the initial weight parameters to obtain the intermediate weight parameters.

[0277] Intermediate weight parameters: These are the weight parameters calculated based on the initial weight parameters, current gradients, and fixed step sizes during the iteration of the adaptive step size gradient algorithm. These intermediate weight parameters are the intermediate results of the stepwise iteration toward the target weight parameters, and further judgment is needed to determine whether the convergence conditions are met to determine whether the target weight parameters are reached.

[0278] Convergence condition: This is a condition used to determine whether the iterative process is finished. In an embodiment of the present application, the convergence condition is used to determine whether the intermediate weight parameter is close enough to the optimal solution (i.e., the target weight parameter). The convergence condition can be defined based on the change amplitude of the intermediate weight parameter (such as the difference between the weight parameter between two iterations is less than a certain threshold) or based on the value of the objective function (such as the value of the objective function changes very little after multiple iterations).

[0279] In the embodiment of the present application, the update formula represents updating the initial weight parameter with a fixed step size in the opposite direction of the current gradient (because the optimization is to be made in the direction of reducing the objective function) to obtain the intermediate weight parameter.

[0280] S1104: When the intermediate weight parameters meet the convergence conditions, determine the target weight parameters.

[0281] In the embodiment of the present application, if the convergence condition is defined as the norm of the difference between the intermediate weight parameters of two adjacent iterations is less than a certain threshold, when the intermediate weight parameter meets this condition, it is considered to have converged to the target weight parameter. This target weight parameter enables the multi-objective optimization model to better balance the relationship between different objectives.

[0282] The embodiment of the present application uses the optimized target weight parameters to adjust the initial recommendation list to obtain the target recommendation task, which can better balance multiple goals (such as accuracy and diversity) in the evaluation task recommendation. The target recommendation task obtained in this way better meets the needs of the target teacher and improves the quality and effectiveness of the recommendation task.

[0283] More specifically, we recommend assessment tasks that are in line with teachers' own development, and solve the problem of balancing the diversity and accuracy of assessment tasks:

[0284] In the intelligent support service platform for teacher professional development, with the increasing number of new teachers and assessment tasks, teacher characteristic data may tend to develop in a single direction, and the diversity and accurate balance of assessment tasks will tend to be chaotic and unstable as the information entropy increases.

[0285] The following is a mathematical algorithm model design for this problem:

[0286] The situational assessment task is aimed at quantifying the accuracy and diversity of teacher recommendations. First, we need to define mathematical quantitative indicators of accuracy and diversity.

[0287] Among them, recommendation accuracy can be measured by the difference between the predicted rating and the actual rating, using the root mean square error (RMSE).

[0288] In a recommendation system, a predicted rating is an estimate of a user's potential preference or evaluation of an item (such as a contextual assessment task) based on a model. This prediction is usually based on the user's historical data, user characteristics, item characteristics, and possible interactions between users and items. By analyzing the user's preference characteristics and the attributes of the item, these features are used to build a prediction model, such as a regression model, to estimate the user's rating of the item.

[0289] The known and unknown parts of the algorithm formula:

[0290] Known: Actual ratings, i.e., users’ true evaluations of items, usually from historical data. The dataset includes the set of all known user-item rating pairs, as well as the total number of known ratings.

[0291] Unknown: Predicted rating, which is what the model predicts the user will rate the item. This is calculated using any of the methods mentioned above.

[0292] RMSE is the root mean square error itself, which is a calculation result used to measure the difference between the predicted score and the actual score, as shown in formula (12):

[0293]

[0294] Among them, rui is the actual rating of item i by user u, is the predicted rating, D is the rating dataset, and N is the total number of ratings in the dataset.

[0295] The recommendation diversity can be measured by the similarity between items in the recommendation list. The cosine distance based on item features can be used, as shown in formula (13):

[0296]

[0297] Among them, i and j are the feature vectors of items i and j in the recommendation list respectively, and N is the length of the recommendation list.

[0298] Creation of a multi-objective optimization model between teacher groups and situational assessment task application model library.

[0299] The multi-objective optimization model is shown in formula (14), which aims to maximize both accuracy and diversity. This is mainly achieved by introducing the λ adjustment parameter, which is used to balance the weight between the two:

[0300] Objective=λ·Accuracy+(1-in)·Diversity (14)

[0301] λ is a weight parameter that needs to be set in advance. It determines the weight ratio between accuracy and diversity when optimizing the objective function.

[0302] Accuracy: The specific value of accuracy is unknown and needs to be calculated from actual user feedback data. For example, in a movie recommendation system, accuracy may be measured by comparing the user's actual rating with the predicted rating recommended by the system, such as using the mean square error (RMSE) or other correlation indicators.

[0303] Diversity: Although the calculation method of diversity is known, the specific diversity value also needs to be calculated based on the actual data of the items in the recommendation list. It reflects the degree of difference between different items in the recommendation list.

[0304] In this framework, by adjusting the value of λ, we can control whether the recommendation system prefers accuracy or diversity.

[0305] Combining content-based recommendations and collaborative filtering recommendations, the recommendation lists generated by each method are mixed according to different weights, which reflect the preference for accuracy or diversity. When calculating the similarity between users or the similarity between items, the diversity index is introduced as a factor to select users or items that are both similar and have certain differences for recommendation.

[0306] Through the above methods, we can improve the diversity of the recommendation list while ensuring the accuracy of the recommendation, enhance the user experience, promote users to discover new points of interest, and ultimately improve user satisfaction and user stickiness of the platform.

[0307] Content-based recommendations and collaborative filtering recommendations will produce different recommendation lists because they are based on different data sources and algorithmic principles. In order to combine content-based recommendations and collaborative filtering recommendations, we usually use a weighted average approach to balance accuracy and diversity. The key here is to determine the weights, which reflect the recommendation system's preference for accuracy or diversity.

[0308] First, determine the weight: Set λ (lambda) as the weight to balance the results of content-based recommendations and collaborative filtering recommendations. λ can be adjusted according to the needs of the system. A higher λ value means more emphasis on content-based recommendations, while a lower λ value emphasizes collaborative filtering results.

[0309] Second, generate a recommendation list: For each user, a recommendation list from a content-based recommendation system and a collaborative filtering recommendation system is calculated separately. Each recommendation item can have a score, which may be calculated based on similarity.

[0310] Third, fusion recommendation list: For each recommendation item, the content-based recommendation score and the collaborative filtering recommendation score are weighted and fused according to the weight. The final recommendation list is generated according to the final score sorting.

[0311] For example:

[0312] The content-based recommendation system identified that Assessment Task A and Assessment Task B are well aligned with the teacher's preferences because these tasks involve mathematical skills in which the teacher is specialized. The recommendation scores are 0.9 and 0.7, respectively.

[0313] The collaborative filtering recommendation system, based on the behavior of other similar teachers, recommended assessment tasks A and C. These tasks are popular among other teachers, with scores of 0.85 and 0.75 respectively.

[0314] Next, we will iterate the key λ of the above solution to find further performance optimization, and optimize it while ensuring that the accuracy and diversity of the recommendation system preferences are not lost.

[0315] In order to quickly find the optimal λ value to ensure that the Objective function reaches the maximum value, and creatively propose a novel method, we can use the "improved adaptive step size gradient ascent method" combined with the "heuristic learning rate adjustment". This method will not only focus on reducing the number of iterations, but also accurately explore the optimal solution through adaptive adjustment, thereby improving computational efficiency and accuracy. Traditional gradient ascent or descent methods are used to find the maximum or minimum value of a function, but when applied to the Objective function, they may be limited by a fixed learning rate, resulting in slow convergence or premature convergence to the local optimum. Our method will introduce an adaptive learning rate adjustment strategy to dynamically adjust the step size according to the progress of the current iteration, and combine heuristic methods to predict gradient changes to optimize the search path.

[0316] Algorithm optimization:

[0317] 1. Initialize weight parameters: Set the initial value of λ to 0.5 and the initial value of learning rate η to 0.1.

[0318] 2. Calculate the gradient As shown in formula (15), the gradient of the Objective function at the current λ value is calculated using the difference method:

[0319]

[0320] where ε is a small constant, such as 0.01.

[0321] 3. Heuristic learning rate adjustment: Adjust the learning rate according to the sign of the current gradient and the previous gradient:

[0322] If two consecutive gradients have the same sign (i.e. both positive or both negative), increase the learning rate to speed up convergence. If the gradients change sign, decrease the learning rate to avoid overshoot.

[0323] 4. Update λ: Use the updated learning rate and gradient to update λ as shown in formula (16):

[0324] λ=λ+η·g (16)

[0325] Where n is the learning rate, g is the step size, and λ is ensured to be in the range of [0, 1].

[0326] 5. Convergence detection: If the change of λ is less than a set threshold (such as 0.001), or the preset maximum number of iterations is reached, the iteration is stopped.

[0327] Next, we will conduct an effect evaluation, which mainly includes the following steps:

[0328] Convergence speed: Traditional fixed-step gradient methods may require more iterations to converge, especially when the behavior of the objective function varies inconsistently in different intervals. Traditional methods may have too large a step size when approaching the optimal solution, resulting in oscillation, or too small a step size and slow convergence.

[0329] The improved adaptive step size method addresses these problems by dynamically adjusting the learning rate. It can reduce the step size when the Objective function is close to the optimal solution to avoid overshoot, and increase the step size to accelerate convergence when it is far away from the optimal solution. This adaptive mechanism can significantly reduce the number of iterations required to find the optimal solution.

[0330] Iterations: Assuming a standard setting, a traditional fixed-step gradient ascent method may require 50 iterations to reach a satisfactory convergence state. Using adaptive step-size gradient ascent, the same conditions may only require 20 iterations, which represents a 60% reduction in the number of iterations.

[0331] Accuracy and diversity: When the traditional method keeps a fixed step size, it may stop optimizing before reaching a satisfactory balance between accuracy and diversity. Due to its adaptive nature, the improved method can adjust λ more accurately to achieve a better balance between accuracy and diversity. For example, it is possible to maintain accuracy above 98% while maintaining diversity above 80%.

[0332] Quantification of performance improvement: In actual applications, the traditional method takes an average of 10 minutes to find the optimal solution, while the improved method only takes 4 minutes due to the reduced number of iterations, which is 60% faster.

[0333] This shows that accuracy and diversity are not only not lost, but may be improved due to more refined λ adjustment.

[0334] The improved adaptive step size gradient ascent method of the embodiment of the present application is particularly suitable for dealing with complex multi-objective optimization problems, such as balancing diversity and accuracy in recommendation systems. It provides an efficient and more effective solution, especially in dynamically changing environments, such as online learning platforms or real-time recommendation systems. The efficiency and accuracy improvements of this method will directly translate into better user experience and system performance.

[0335] In an exemplary embodiment, based on the above embodiment, the method of the embodiment of the present application further includes the following steps:

[0336] Step 1: Perform data conversion on the continuous features in the basic information of the target teacher according to the pre-established binning strategy to obtain categorical features; perform unique-hot encoding on the categorical features and the discrete features in the basic information to obtain the user preference vector; perform step 2 if the target teacher is an existing teacher, and perform steps 3 to 7 if the target teacher is a new teacher;

[0337] Step 2: Use the cosine similarity algorithm to calculate the similarity between the user preference vector and the feature vector of each evaluation task; sort the evaluation tasks according to the similarities and determine the initial recommendation list;

[0338] Step 3: Perform one-hot encoding on each operation type in the evaluation task information to obtain a binary vector; perform feature extraction on the binary vector to obtain a feature set; perform feature conversion on the feature set to obtain a feature vector for each evaluation task;

[0339] Step 4: Perform node feature initialization processing on the user preference vector and each evaluation task feature vector respectively to obtain the first feature vector of the teacher node and the second feature vectors of multiple evaluation task nodes;

[0340] Step 5: Input the distance between the first eigenvector and each second eigenvector into the distance function for distance calculation to obtain the distance result; input the distance result into the edge weight function for calculation to obtain the edge weight data; construct the initial teacher task affinity graph according to the edge weight data;

[0341] Step 6: Input the first eigenvector and each second eigenvector into the input layer for eigenvector mapping processing to obtain a third eigenvector; the dimension of the third eigenvector corresponds to the dimension of the input layer; input the third eigenvector into the update layer for feature update processing to obtain an update node; the update node includes a teacher update node and a task update node;

[0342] Step 7: Calculate the cosine similarity between the teacher update node and the task update node to obtain similarity data; determine the recommendation score based on the similarity data; the recommendation score is used to characterize the matching degree between the target teacher and each assessment task; adjust the edge weight data based on the recommendation score to obtain the target teacher task affinity graph; generate the initial recommendation list based on the target teacher task affinity graph;

[0343] Step 8: Obtain the initial weight parameters; in each iterative solution process, calculate the current gradient of the multi-objective optimization model; update the initial weight parameters according to the current gradient and the fixed step size to obtain the intermediate weight parameters; when the intermediate weight parameters meet the convergence conditions, determine the target weight parameters;

[0344] Step 9: Adjust the initial recommendation list according to the target weight parameter to obtain the target recommendation task.

[0345] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0346] Based on the same inventive concept, the embodiment of the present application also provides an evaluation task recommendation device for implementing the evaluation task recommendation method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more evaluation task recommendation device embodiments provided below can refer to the limitations of the evaluation task recommendation method above, and will not be repeated here.

[0347] In one embodiment, Fig.12 As shown, a device 1200 for recommending evaluation tasks is provided, comprising:

[0348] A vector construction module 1201 is used to construct a user preference vector and multiple evaluation task feature vectors based on the pre-acquired basic information of the target teacher and the evaluation task information;

[0349] List determination module 1202, used to determine the initial recommendation list according to the user preference vector and the feature vectors of each evaluation task;

[0350] The task determination module 1203 is used to optimize the initial recommendation list based on a pre-built multi-objective optimization model to obtain a target recommendation task; the multi-objective optimization model is built based on a task diversity index and an accuracy index.

[0351] In one embodiment, the list determination module includes:

[0352] An affinity graph generating unit is used to generate a target teacher task affinity graph according to the user preference vector and the feature vectors of each assessment task; the target teacher task affinity graph is used to characterize the matching degree between the target teacher and multiple assessment tasks;

[0353] The initial list generation unit is used to generate an initial recommendation list according to the target teacher task affinity diagram.

[0354] In one embodiment, the target graph generation unit includes:

[0355] A vector generation subunit is used to perform node feature initialization processing on the user preference vector and each evaluation task feature vector, respectively, to obtain a first feature vector of the teacher node and second feature vectors of multiple evaluation task nodes;

[0356] An initial graph generating subunit, for obtaining an initial teacher task affinity graph according to the distance between the first eigenvector and each second eigenvector;

[0357] The target graph generation subunit is used to update the initial teacher task affinity graph based on the graph neural network to obtain the target teacher task affinity graph.

[0358] In one embodiment, the initial graph generating subunit is specifically used for:

[0359] The distance between the first eigenvector and each second eigenvector is input into the distance function to perform distance calculation processing to obtain a distance result; the distance result is used to represent the distance between the teacher node and each evaluation task node;

[0360] The distance result is input into the edge weight function for calculation and processing to obtain edge weight data; the edge weight data is used to characterize the affinity between the teacher node and the assessment task node;

[0361] Construct an initial teacher-task affinity graph based on edge weight data.

[0362] In one embodiment, the initial list generating unit includes:

[0363] The score determination subunit is used to calculate the affinity between the teacher node and each assessment task node in the target teacher task affinity graph to obtain multiple affinity scores;

[0364] The initial list generation subunit is used to sort multiple evaluation tasks according to multiple affinity scores to obtain an initial recommendation list.

[0365] In one embodiment, the vector building module includes:

[0366] A first encoding unit, used to perform a first feature encoding process on the basic information to obtain a user preference vector;

[0367] The second encoding unit is used to perform second feature encoding processing on the evaluation task information to obtain a feature vector of each evaluation task.

[0368] In one embodiment, the basic information includes discrete features and continuous features; the first encoding unit includes:

[0369] A data conversion subunit is used to perform data conversion processing on continuous features according to a pre-defined binning strategy to obtain categorical features; the binning strategy includes one of equal-width binning, equal-frequency binning and custom binning;

[0370] The encoding subunit is used to perform one-hot encoding on the categorical features and discrete features respectively to obtain the user preference vector.

[0371] In one embodiment, the assessment task information includes multiple operation types; the second encoding unit includes:

[0372] The second encoding subunit is used to perform one-hot encoding processing on each operation type to obtain a binary vector;

[0373] An extraction subunit is used to perform feature extraction processing on the binary vector to obtain a feature set;

[0374] The conversion subunit is used to perform feature conversion processing on the feature set to obtain the feature vector of each evaluation task.

[0375] In one embodiment, the task determination module includes:

[0376] An adjustment unit, used to adjust the learning rate of the multi-objective optimization model by using an adaptive step gradient algorithm to obtain a target weight parameter;

[0377] The task generation unit is used to adjust the initial recommendation list according to the target weight parameter to obtain the target recommendation task.

[0378] In one embodiment, the adjustment unit includes:

[0379] A parameter acquisition subunit is used to obtain initial weight parameters;

[0380] A gradient calculation subunit is used to calculate the current gradient of the multi-objective optimization model in each iterative solution process;

[0381] The update subunit is used to update the initial weight parameters according to the current gradient and fixed step size to obtain the intermediate weight parameters;

[0382] The parameter determination subunit is used to determine the target weight parameter when the intermediate weight parameter meets the convergence condition.

[0383] Each module in the above-mentioned assessment task recommendation device can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0384] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Fig.13 As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a method for recommending an evaluation task is implemented. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0385] Those skilled in the art will understand that Fig.13 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0386] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the steps of the method described in any one of the above embodiments are implemented.

[0387] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in any of the above embodiments are implemented.

[0388] In one embodiment, a computer program product is provided, including a computer program, which implements the steps of the method described in any of the above embodiments when executed by a processor.

[0389] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0390] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0391] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0392] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for recommending an evaluation task, characterized in that: The method comprises: Based on the pre-acquired basic information of the target teacher and the assessment task information, a user preference vector and multiple assessment task feature vectors are constructed; Determining an initial recommendation list according to the user preference vector and each of the evaluation task feature vectors; Based on a pre-built multi-objective optimization model, the initial recommendation list is optimized to obtain a target recommendation task; the multi-objective optimization model is built based on a task diversity index and an accuracy index.

2. The method according to claim 1, characterized in that Determining an initial recommendation list according to the user preference vector and each of the evaluation task feature vectors includes: Generate a target teacher task affinity graph according to the user preference vector and each of the assessment task feature vectors; the target teacher task affinity graph is used to characterize the matching degree between the target teacher and multiple assessment tasks; The initial recommendation list is generated according to the target teacher task affinity diagram.

3. The method according to claim 2, characterized in that The step of generating a target teacher task affinity graph according to the user preference vector and each of the assessment task feature vectors includes: Performing node feature initialization processing on the user preference vector and each of the evaluation task feature vectors respectively to obtain a first feature vector of the teacher node and second feature vectors of multiple evaluation task nodes; Obtaining an initial teacher task affinity graph according to the distance between the first eigenvector and each of the second eigenvectors; Based on the graph neural network, the initial teacher task affinity graph is updated to obtain the target teacher task affinity graph.

4. The method according to claim 3, characterized in that: The initial teacher task affinity graph is obtained according to the distance between the first feature vector and each of the second feature vectors, including: Inputting the distance between the first feature vector and each of the second feature vectors into a distance function to perform distance calculation processing to obtain a distance result; the distance result is used to characterize the distance between the teacher node and each of the evaluation task nodes; The distance result is input into the edge weight function for calculation and processing to obtain edge weight data; the edge weight data is used to characterize the affinity between the teacher node and the assessment task node; The initial teacher task affinity graph is constructed based on the edge weight data.

5. The method according to claim 4, characterized in that The step of generating the initial recommendation list according to the target teacher task affinity graph includes: Calculating the affinity between the teacher node and each of the assessment task nodes in the target teacher task affinity graph to obtain a plurality of affinity scores; The multiple evaluation tasks are sorted according to the multiple affinity scores to obtain the initial recommendation list.

6. The method according to claim 1, characterized in that The method of constructing a user preference vector and multiple evaluation task feature vectors based on the pre-acquired basic information of the target teacher and the evaluation task information includes: Performing a first feature encoding process on the basic information to obtain the user preference vector; A second feature encoding process is performed on the assessment task information to obtain a feature vector of each assessment task.

7. The method according to claim 6, characterized in that The basic information includes discrete features and continuous features; The performing a first feature encoding process on the basic information to obtain the user preference vector comprises: Performing data conversion processing on the continuous features according to a pre-established binning strategy to obtain categorical features; the binning strategy includes one of equal-width binning, equal-frequency binning and custom binning; The categorical features and the discrete features are respectively subjected to one-hot encoding processing to obtain the user preference vector.

8. The method according to claim 6, characterized in that The evaluation task information includes multiple operation types; the second feature encoding process is performed on the evaluation task information to obtain the feature vectors of each evaluation task, including: Performing one-hot encoding processing on each of the operation types to obtain a binary vector; Performing feature extraction processing on the binary vector to obtain a feature set; Perform feature conversion processing on the feature set to obtain feature vectors of each evaluation task.

9. The method according to claim 1, characterized in that: The optimizing process of the initial recommendation list based on the pre-built multi-objective optimization model to obtain the target recommendation task includes: Adopting an adaptive step-size gradient algorithm to adjust the learning rate of the multi-objective optimization model to obtain an objective weight parameter; The initial recommendation list is adjusted according to the target weight parameter to obtain the target recommendation task.

10. The method according to claim 9, characterized in that The adaptive step gradient algorithm is used to adjust the learning rate of the multi-objective optimization model to obtain the target weight parameter, including: Get the initial weight parameters; In each iterative solution process, calculating the current gradient of the multi-objective optimization model; According to the current gradient and the fixed step size, the initial weight parameter is updated to obtain an intermediate weight parameter; When the intermediate weight parameter meets the convergence condition, the target weight parameter is determined.

Images