A test question recommendation model training method, test question recommendation method and system

Abstract

The application belongs to the technical field of personalized recommendation, and discloses a training method of a test question recommendation model, a test question recommendation method and a system, which comprise the following steps: enhancing each node in a multi-relation interaction graph by using a preset multi-relation graph convolution, to obtain an enhanced embedding representation corresponding to the node; comparing the interaction relation representation by a sample to learn the preferences between different learners, and constructing a learner comparison loss function; obtaining a text label representation and a test question representation by comparing the label information of a test question and a learner who cannot do the test question, and constructing a test question comparison loss function according to the text label representation and the test question representation; constructing a learner full-correct test question loss function and a learner incorrect test question loss function; and performing model training according to the learner comparison loss function, the test question comparison loss function, the learner full-correct test question loss function and the learner incorrect test question loss function, to obtain a target recommendation model. The application improves the accuracy and comprehensiveness of personalized recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of personalized recommendation, in particular to a training method of a test question recommendation model, a test question recommendation method and system. BACKGROUND

[0002] The traditional teaching mode adopts indiscriminate teaching and blind question training, which increases the burden of learners. With the rise of smart education and intelligent auxiliary teaching, many online education platforms have appeared at home and abroad, and these platforms have been highly valued. Therefore, how to accurately recommend test questions suitable for learners has become a hot issue.

[0003] However, existing research only assumes that all learners and test questions have static and unchanging characteristics, and uses a fixed mode (such as an implicit vector) to model, without fully focusing on the dynamic characteristics of learners and test questions over time. Many recommendation algorithms are too dependent on answer history due to data sparsity and cold start problems, and lack understanding of the propagation and role of various interaction relationships between learners and their adjacent learners (neighbors) and test questions and their collaborative information, making it difficult to represent.

[0004] Therefore, it is urgent to solve how to reduce the data sparsity and answer history dependence of the model to improve the accuracy and comprehensiveness of personalized recommendation. SUMMARY

[0005] The purpose of the present application is to provide a training method of a test question recommendation model, a test question recommendation method and system to improve the accuracy and comprehensiveness of personalized recommendation.

[0006] To achieve the above purpose, in a first aspect, the present application provides a training method of a test question recommendation model, the training method comprising:

[0007] obtaining a multi-relation interaction graph; the multi-relation interaction graph is constructed with test questions and learners as nodes, and with a learner's poorly done test question interaction matrix and a learner's all-correct test question interaction matrix as edges; the learner's poorly done test question interaction matrix represents at least one knowledge point corresponding to the test question that the learner answered incorrectly, and the learner's all-correct test question interaction matrix represents all knowledge points corresponding to the test question that the learner answered correctly;

[0008] enhancing each node in the multi-relation interaction graph using a pre-set multi-relation graph convolution to obtain an enhanced embedding representation corresponding to the node; wherein the enhanced embedding representation includes a learner enhanced embedding representation and a test question enhanced embedding representation;

[0009] constructing different sample pair interaction representations according to the learner enhanced embedding representation and the test question enhanced embedding representation, comparing the preferences between different learners through the sample pair interaction representations, and constructing a learner comparison loss function.

[0010] constructing a question contrast loss function according to the text label representation and the question representation;

[0011] constructing a learner full-correct question loss function and a learner incorrect question loss function according to the learner enhanced embedding representation and the question enhanced embedding representation;

[0012] performing model training according to the learner contrast loss function, the question contrast loss function, the learner full-correct question loss function and the learner incorrect question loss function to obtain a target recommendation model.

[0013] In an embodiment, before the enhancing each of the nodes in the multi-relation interaction graph by using the preset multi-relation graph convolution to obtain the enhanced embedding representation corresponding to the node, the training method further comprises:

[0014] constructing a multi-relation adjacency matrix based on the learner incorrect question interaction matrix and the learner full-correct question interaction matrix;

[0015] enhancing each of the nodes in the multi-relation interaction graph by using the preset multi-relation graph convolution based on the multi-relation adjacency matrix to obtain the enhanced embedding representation corresponding to the node.

[0016] In an embodiment, the constructing a multi-relation adjacency matrix based on the learner incorrect question interaction matrix and the learner full-correct question interaction matrix comprises:

[0017] constructing a sparse interaction matrix according to the learner incorrect question interaction matrix and the learner full-correct question interaction matrix;

[0018] performing conversion on the sparse interaction matrix to obtain the multi-relation adjacency matrix.

[0019] In an embodiment, the enhancing each of the nodes in the multi-relation interaction graph by using the preset multi-relation graph convolution based on the multi-relation adjacency matrix to obtain the enhanced embedding representation corresponding to the node comprises:

[0020] obtaining an input vector of the node;

[0021] extracting an initial feature embedding vector of the node by using a preset nonlinear co-embedding function according to the input vector;

[0022] In a convolution layer of any layer of the multi-relation graph convolution, information of neighbor nodes of any node in the multi-relation interaction graph is aggregated through the multi-relation adjacency matrix to obtain an updated embedding vector of the any node;

[0023] Traverse each convolutional layer, and for each node, determine the average value of the updated embedding vectors corresponding to the node in all the convolutional layers to obtain the enhanced embedding representation corresponding to the node.

[0024] In one embodiment, constructing different sample-pair interaction relationship representations based on the learner-enhanced embedded representation and the question-enhanced embedded representation, comparing the sample-pair interaction relationship representations to learn preferences between different learners, and constructing a learner contrast loss function includes:

[0025] The learner enhanced embedding representation and the test question enhanced embedding representation are learned through the MLP layer to generate different sample pair interaction relationship representations; wherein the sample pair interaction relationship representations include positive sample pair interaction relationship representations, fine-grained negative sample pair interaction relationship representations and coarse-grained negative sample pair interaction relationship representations; the positive sample pair interaction relationship representations are used to represent the answer situation of the same learner on the same test question, including the interactive behavior of answering the test question poorly and answering all the test questions correctly; the fine-grained negative sample pair interaction relationship representations are used to represent the answer situation of different learners on the same test question, including the interactive behavior of answering the test question poorly; the coarse-grained negative sample pair interaction relationship representations are used to represent the answer situation of different learners on different test questions, including the interactive behavior of answering the test question poorly;

[0026] Constructing a fine-grained loss function according to the similarity between the positive sample pair interaction relationship representation and the fine-grained negative sample pair interaction relationship representation;

[0027] Constructing a coarse-grained loss function based on the similarity between the positive sample pair interaction relationship representation and the coarse-grained negative sample pair relationship representation;

[0028] The fine-grained loss function and the coarse-grained loss function are weighted to obtain the learner contrast loss function.

[0029] In one embodiment, the method of enhancing the embedding representation based on the learner, obtaining a text label representation and a test question representation by comparing the label information of the learning test question with the learner who did not perform well on the test question, and constructing a test question contrast loss function based on the text label representation and the test question representation, includes:

[0030] Extracting label information of the test question through a text encoder to obtain a text label representation;

[0031] According to the preset selection rule, the learner enhanced embedding representation is selected to obtain a learner embedding representation of a poorly done question;

[0032] The learner embedding representation of the poorly done question is converted into a question representation through a question encoder;

[0033] According to the similarity between the text label representation and the question representation, a text label loss function and a question loss function are constructed;

[0034] The text label loss function and the question loss function are averaged to obtain the question contrast loss function.

[0035] In an embodiment, the learner full-correct question loss function and the learner poorly done question loss function are constructed according to the learner enhanced embedding representation and the question enhanced embedding representation, comprising:

[0036] According to the learner enhanced embedding representation and the question enhanced embedding representation, a learner poorly done question score function and a learner full-correct question score function are determined through a fully connected neural network;

[0037] The learner poorly done question score function and the learner full-correct question score function are respectively constructed into an initial learner poorly done question loss function and an initial learner full-correct question loss function using a binary cross-entropy loss function;

[0038] A preset relative difficulty coefficient is optimized using a mean square error to obtain a learner poorly done question relative difficulty coefficient loss function and a learner full-correct question relative difficulty coefficient loss function; wherein the relative difficulty coefficient represents the mastery degree of the learner to the knowledge points in the question;

[0039] The learner full-correct question loss function is obtained according to the sum of the initial learner full-correct question loss function and the learner full-correct question relative difficulty coefficient loss function; and the learner poorly done question loss function is obtained according to the sum of the initial learner poorly done question loss function and the learner poorly done question relative difficulty coefficient loss function.

[0040] In a second aspect, an embodiment of the present application provides a question recommendation method, the recommendation method comprising:

[0041] A target multi-relation interaction graph is obtained; wherein the target multi-relation interaction graph is constructed by taking a target question and a to-be-recommended learner as nodes, and taking a to-be-recommended learner poorly done target question interaction matrix and a to-be-recommended learner full-correct target question interaction matrix as edges;

[0042] inputting the target multi-relation interaction graph into a target recommendation model obtained by the training method of the test question recommendation model to obtain a target recommended learner list corresponding to the target test question;

[0043] recommending the target test question to the learners in the target recommended learner list.

[0044] In a third aspect, an embodiment of the present application provides a training system of a test question recommendation model, and the training system comprises:

[0045] an interaction graph acquisition module configured to acquire a multi-relation interaction graph, wherein the multi-relation interaction graph is constructed by taking test questions and learners as nodes and taking a learner's interaction matrix of test questions that the learner does not answer correctly and a learner's interaction matrix of test questions that the learner answers correctly as edges, the learner's interaction matrix of test questions that the learner does not answer correctly represents at least one knowledge point corresponding to a test question that the learner answers incorrectly, and the learner's interaction matrix of test questions that the learner answers correctly represents all knowledge points corresponding to a test question that the learner answers correctly;

[0046] a node enhancement module configured to enhance each node in the multi-relation interaction graph by using a preset multi-relation graph convolution to obtain an enhanced embedding representation corresponding to the node, wherein the enhanced embedding representation comprises a learner enhanced embedding representation and a test question enhanced embedding representation;

[0047] a learner comparison module configured to construct different sample pair interaction representations according to the learner enhanced embedding representation and the test question enhanced embedding representation, compare the sample pair interaction representations to learn the preferences of different learners, and construct a learner comparison loss function;

[0048] a test question comparison module configured to obtain a text label representation and a test question representation by comparing label information of a test question and a learner who does not answer the test question according to the learner enhanced embedding representation, and construct a test question comparison loss function according to the text label representation and the test question representation;

[0049] a loss function construction module configured to construct a learner's interaction matrix of test questions that the learner answers correctly loss function and a learner's interaction matrix of test questions that the learner does not answer correctly loss function according to the learner enhanced embedding representation and the test question enhanced embedding representation;

[0050] a model training module configured to perform model training according to the learner comparison loss function, the test question comparison loss function, the learner's interaction matrix of test questions that the learner answers correctly loss function, and the learner's interaction matrix of test questions that the learner does not answer correctly loss function to obtain a target recommendation model.

[0051] In a fourth aspect, an embodiment of the present application provides an electronic device, which comprises

[0052] a memory,

[0053] a processor, and

[0054] A computer program stored on the memory and executable on the processor, the processor implementing the training method of the test question recommendation model or the test question recommendation method when executing the computer program.

[0055] In a fifth aspect, an embodiment of the present application provides a computer readable storage medium, having a computer program stored thereon, wherein the program, when executed by a processor, implements the training method of the test question recommendation model or the test question recommendation method.

[0056] Compared with the prior art, the training method of the test question recommendation model, the test question recommendation method and the system provided by the embodiment of the present application have the following beneficial effects:

[0057] The present application enhances each node in the obtained multi-relation interaction graph by using a preset multi-relation graph convolution, so as to capture the collaborative information between the learner and the neighbor and the test question, thereby enhancing the embedding of the learner and the test question. By constructing a loss function from two angles of learner contrast learning and test question contrast learning, the feature representation of the test question and the learner is assisted to be optimized, the difference in the preferred test question of different learners is increased, the different interaction preferences between the same learner and the test question are reduced, and the test question and the label are optimized in the aspect of annotation and interpretation, thereby fundamentally enhancing the ability and accuracy of personalized test question recommendation. In addition, the model is trained in combination with the learner full-correct test question loss function and the learner poor test question loss function, the data sparsity and the answer history dependency are comprehensively reduced, and the overall performance of the personalized test question recommendation is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0058] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0059] Figure 1 A flowchart of a training method of a test question recommendation model provided by an embodiment of the present application is shown in the figure;

[0060] Figure 2 A multi-relation interaction graph provided by an embodiment of the present application is shown in the figure;

[0061] Figure 3 A structure diagram of a recommendation model provided by an embodiment of the present application is shown in the figure;

[0062] Figure 4 A flowchart of a test question recommendation method provided by an embodiment of the present application is shown in the figure;

[0063] Figure 5A block diagram of a test question recommendation model training system is provided for embodiments of the present application.

[0064] Figure 6 An internal structure schematic diagram of an electronic device is provided for embodiments of the present application. DETAILED DESCRIPTION

[0065] The specific embodiments of the present application will be further described with reference to the drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application.

[0066] It is apparent that the accompanying drawings described below are only some examples or embodiments of the present application, and for those skilled in the art, the present application can be applied to other similar scenarios without creative labor on the basis of these drawings. In addition, it can be understood that although the efforts made in this development process can be complex and lengthy, for those skilled in the art related to the content disclosed in the present application, some design, manufacture or production changes based on the technical content disclosed in the present application are only routine technical means, and should not be understood as insufficient disclosure of the present application.

[0067] In the present application, the phrase "embodiments" means that the specific features, structures or characteristics described in conjunction with the embodiments can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily mean the same embodiment, nor is it independent or alternative to other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described in the present application can be combined with other embodiments without conflict.

[0068] Unless otherwise defined, technical terms and scientific terms used in the present disclosure shall have the same meaning as commonly understood by one of ordinary skill in the art to which this present disclosure belongs. The terms "a", "an", "the", and the like, do not denote the quantity of stated elements, but rather denote the presence of at least one of those elements. The terms "comprise", "include", "have" and any variations thereof are intended to cover a non-exclusive inclusion. For example, a process, method, article, or apparatus that comprises a list of steps or units is not necessarily limited to the listed steps or units, but can include additional steps or units not expressly listed or inherent to such process, method, article, or apparatus. The terms "connect", "couple", and the like, are not limited to direct or physical connections, but can include indirect connections or connections through intervening elements. The term "multiple" means two or more. The term "and / or" describes associated objects, and can mean three conditions: alone, simultaneously, or alone. The character " / " generally means an "or" relationship between the associated objects. The terms "first", "second", "third", and the like, merely distinguish similar objects, and do not represent a specific order.

[0069] The traditional teaching mode uses indiscriminate teaching and blind question training to increase the burden of learners. With the high attention to smart education and intelligent auxiliary teaching, a number of online education platforms have emerged at home and abroad. However, the massive resource data brought by these platforms often leads to the problem of "information overload and selection confusion", especially how to accurately recommend appropriate test questions to learners has become a hot concern.

[0070] At present, the recommendation system for "recommending items or projects for users" is applied to "recommending test questions for learners". In the aspect of personalized test question recommendation, many scholars have carried out series of researches on the three core works of the recommendation model, i.e. learner modeling, test question modeling and recommendation algorithm, and have achieved results. However, the existing models have problems in recommendation accuracy, learning state change adaptability, data sparsity and cold start in actual application. The main reason is that the perception of the individual cognitive level of the static and dynamic characteristics of the learners and test questions, the interactive preference of the learners to the test questions and the knowledge mastery degree is not sufficient, and the representation is not comprehensive, rich and flexible enough.

[0071] In terms of user interest preference-based recommendation, Top-N recommendation provides an effective personalized service method for users with different preferences. However, the current most advanced item recommendation model, including the personalized recommendation model based on the cognitive diagnosis method (CDM), has a significant difference in the recommendation performance for learners with different cognitive levels, which seriously affects the service quality of personalized recommendation.

[0072] Existing research often simply assumes that all learners and items have static and unchanging characteristics, and uses a fixed mode (such as a latent vector) to model their embedding, while ignoring the dynamic characteristics of learners and items over time. In addition, many recommendation algorithms rely heavily on answer history due to data sparsity and cold start problems, and lack understanding and representation of the propagation and role of various interaction relationships between learners and their neighbors and items and their collaborative information.

[0073] Learner modeling aims to accurately diagnose and represent learner information by mining explicit and latent features such as basic information, demand goals, learning preferences, behavior characteristics, and knowledge mastery levels, so that they have as rich personalized semantics as possible. Current modeling methods are mainly based on knowledge tracking (KT) and cognitive diagnosis (CD). Item modeling focuses on semantic representation and knowledge analysis of items, involving classical test theory (CTT) and multi-modal deep semantic representation.

[0074] Personalized item recommendation algorithms recommend appropriate items to learners based on the comprehensive consideration of matching between learner personalized characteristics and item characteristics. Current main recommendation algorithms include collaborative filtering (CF), cognitive diagnosis method (CDM), and deep learning (DL). Among them, CDM is widely used due to its dynamic cognitive level perception ability, including IRT, DINA, and neural network cognitive diagnosis model.

[0075] However, existing personalized item recommendation models still face challenges in comprehensive, rich, and flexible representation of static and dynamic characteristics of learners and items, interaction preferences of learners for items, and individual cognitive levels such as knowledge mastery. In particular, the real-time perception and representation ability of dynamic changing learner and item characteristics is insufficient, which limits the effectiveness and adaptability of the model.

[0076] Based on the above situation, the embodiment of the present application provides a training method of an item recommendation model, an item recommendation method and a system.

[0077] In a first aspect, the embodiment of the present application provides a training method of an item recommendation model, Figure 1 The flowchart of the training method of the item recommendation model provided by the embodiment of the present application is shown in Figure 1 The method comprises the following steps:

[0078] Step S101, obtaining a multi-relation interaction graph; the multi-relation interaction graph is constructed by taking a test question and a learner as a node and taking a learner to do a poorly done test question interaction matrix and a learner to do a full correct test question interaction matrix as an edge; the learner to do a poorly done test question interaction matrix represents at least one knowledge point corresponding to a test question done by a learner; and the learner to do a full correct test question interaction matrix represents all knowledge points corresponding to a test question done by a learner.

[0079] In the embodiment, in order to reduce the data sparsity and the answer history dependence of the model, the multiple interaction behaviors between the learners and the test questions need to be considered to comprehensively understand the preferences, needs and the like of the learners in subsequent model training, and thus the accuracy of personalized recommendation is improved. The present application first defines a group of graph data, taking a test question node set V i and a learner node set V u , taking different interaction relationships between the learners and the test questions as edges, and constructing two types of heterogeneous graphs, and then constructing a multi-relation interaction graph according to the two types of heterogeneous graphs. Please refer to Figures 2-3 , Figure 2 the multi-relation interaction graph provided by the embodiment of the present application, Figure 3 the structure schematic diagram of the recommendation model provided by the embodiment of the present application, and the specificities are as follows:

[0080] a full correct interaction graph G p =(V,R p ) including the learner node set V u , the test question node set V i and the learner to do a full correct test question interaction matrix R p . As shown in Figure 2 , each test question i done by a learner u is connected to only one learner u.

[0081] a poorly done interaction graph G b =(V,R b ) including the learner node set V u , the test question node set V i and the learner to do a poorly done test question interaction matrix R b . As shown in Figure 2 , each test question i done by a learner u is connected to at least one learner u. Wherein, there is only one edge between the same learner and the test question in the poorly done interaction graph G b , ensuring the uniqueness of the interaction relationship.

[0082] a multi-relation interaction graph G total =(G p , G b ) combining the full correct interaction graph G p and the poorly done interaction graph G b . As shown inFigure 2 A comprehensive view is provided to analyze the interaction behavior between the learner and the test question.

[0083] In step S102, each node in the multi-relation interaction graph is enhanced by using a preset multi-relation graph convolution to obtain an enhanced embedding representation corresponding to the node; wherein the enhanced embedding representation includes a learner enhanced embedding representation and a test question enhanced embedding representation.

[0084] Specifically, the multi-relation graph convolution MRGC (Multi-Relation Graph Convolution) is an extension of the graph convolution network (GCN), and the core idea of MRGC is to capture and integrate the information of different relations in the multi-relation interaction graph to improve the learning ability of the node representation. Specifically, MRGC learns the embedding representation of each node and the interaction relation in the graph, and these embeddings capture the collaborative information between the learner and its neighbors and the test question. High-level features of graph data are gradually extracted through multi-layer convolution operations. Each layer of convolution processes the features of the node and its neighbors to learn more abstract feature representations. The neighbor node information from different relation types is integrated into each node, so that the generated enhanced embedding representation is more comprehensive, rich and flexible, fully utilizes the multiple interaction relations between the learner and the test question and the test question context features, and reduces the data sparsity and historical answering dependency.

[0085] In some embodiments, before each node in the multi-relation interaction graph is enhanced by using a preset multi-relation graph convolution to obtain an enhanced embedding representation corresponding to the node, the training method further includes: constructing a multi-relation adjacency matrix based on the learner-did-not-do-well test question interaction matrix and the learner-did-do-well test question interaction matrix; based on the multi-relation adjacency matrix, each node in the multi-relation interaction graph is enhanced by using a preset multi-relation graph convolution to obtain an enhanced embedding representation corresponding to the node.

[0086] Specifically, in the multi-relation graph convolution, the neighbor information is iteratively combined to update the representation of the node by using GCN. In this network structure, the neighbors of the node and their high-order connection relations are determined by the multi-relation adjacency matrix Madj. Specifically, the multi-relation adjacency matrix is:

[0087]

[0088] wherein R total represents the total interaction matrix between the learner and the test question, including the learner-did-do-well test question interaction matrix R p and the learner-did-not-do-well test question interaction matrix R b , (R total ) T is R totalThe transpose of the total interaction matrix R

[0089] Then, the multi-relation graph convolution is used to iteratively combine the neighbor information to supplement the representation of the (learner or test item) node. Based on the multi-relation adjacency matrix, the interaction relationship between the learners and the test items can be appropriately enhanced by the multi-layer convolution operation to obtain an embedding representation, thereby enhancing the representation ability and learning effect of the model.

[0090] In some embodiments, the multi-relation adjacency matrix is constructed based on the learner-wrong-test item interaction matrix and the learner-all-right-test item interaction matrix, including: constructing a sparse interaction matrix according to the learner-wrong-test item interaction matrix and the learner-all-right-test item interaction matrix; and converting the sparse interaction matrix to obtain the multi-relation adjacency matrix.

[0091] Specifically, the learner-all-right-test item interaction matrix R p and the learner-wrong-test item interaction matrix R b are used to construct a total interaction matrix R total =R p +αR b , where α is a hyperparameter, used to avoid the increase in the embedding size of the graph convolution operation to weight R b , and the value range is 0-1. However, since the learners usually interact with only part of the test items, the total interaction matrix R total is a sparse matrix. Here, R total needs to be converted into edges in the graph, and the bidirectional relationship between the learners and the test items is also considered. The multi-relation adjacency matrix obtained after conversion is:

[0092]

[0093] The structure of the multi-relation adjacency matrix retains all possible interaction modes between the learners and the test items. In the subsequent graph convolution operation, the structural information in the multi-relation adjacency matrix, such as the connection relationship and the relationship type between the nodes, can be used for learning and inference of the node representation.

[0094] In some embodiments, based on a multi-relationship adjacency matrix, each node in a multi-relationship interaction graph is enhanced using a preset multi-relationship graph convolution to obtain an enhanced embedding representation corresponding to the node, including: obtaining an input vector of the node; according to the input vector, using a preset nonlinear co-embedding function to extract the initial feature embedding vector of the node; in the convolution layer of any layer of the multi-relationship graph convolution, aggregating the information of the neighbor nodes of any node in the multi-relationship interaction graph through the multi-relationship adjacency matrix to obtain an updated embedding vector of any node; traversing each convolution layer, for each node, determining the average value of the updated embedding vectors corresponding to the nodes in all convolution layers to obtain an enhanced embedding representation corresponding to the node.

[0095] Specifically, to improve the representation of learner and question nodes, GCN iteratively aggregates information about the node's neighbors. In each iteration, or aggregation, the information about neighboring nodes is weighted and aggregated into the feature representation of the node (learner or question). The multi-relational adjacency matrix Madj defines the connections between learner and question nodes, including direct and high-order connections. At each layer of the graph convolution, a node (learner or question) aggregates information about its neighboring nodes using the multi-relational adjacency matrix Madj to update its embedding representation.

[0096] For example, see Figure 3 , first obtain the input vector x of the learner and the test question u and x i , x u It can be just a one-hot encoding of the learner ID, or it can flexibly add or subtract some learner’s personalized characteristics (such as learning style, knowledge base, interest preference and personal status, etc.) according to actual needs. i It mainly consists of the test question ID (one-hot encoding) and the test question's characteristic attributes. The test question's characteristic attributes include knowledge points, difficulty level, previous answer records, and multimodal data (text, pictures, videos, etc.).

[0097] Then, using two nonlinear co-embedding functions f U and f I , for the input vector x from the learner u And the input vector x of the test question i Extract the initial feature embedding vector z u and z i .

[0098]

[0099] Among them, θ U and θ I are the parameters of the GCN nonlinear fully connected layer.

[0100] Then, for the node of the test item i and the node of the learner u, after the kth iteration, the embedding of the test item is updated by aggregating the neighbor information, and e i (k) The updated embedding vector e i (k+1) is obtained by transformation. u (k+1) The updated embedding vector e u (k+1) is obtained by transformation. u (0) = z u , e i (0) = z i . Finally, for each node, the average of the updated embedding vectors in all convolution layers is calculated to obtain the enhanced embedding representation Z u and Z i .

[0101]

[0102] Step S103, constructing different sample pair interaction representations according to the learner enhanced embedding representation and the test item enhanced embedding representation, and learning the preferences between different learners through the sample pair interaction representation to construct a learner comparison loss function.

[0103] Specifically, the enhanced embedding representations of the learners and the test items are used to construct sample pairs. Each sample pair includes the interaction representation between a learner and a test item, and the preferences and differences between different learners are learned through the sample pair interaction representation to construct a learner comparison loss function. The goal of the learner comparison loss function is to push the enhanced embeddings of similar learners closer in the embedding space of the learners, and to push the enhanced embeddings of dissimilar learners away. Cosine similarity can be used to measure the similarity between the enhanced embedding representations of different sample pairs. For each learner-test item pair, the loss function can be defined as maximizing the similarity of similar learners and minimizing the similarity of dissimilar learners.

[0104] In some embodiments, different sample pair interaction relationship representations are constructed based on the learner enhanced embedding representation and the test question enhanced embedding representation, and the preferences between different learners are learned by comparing the sample pair interaction relationship representations, and a learner contrast loss function is constructed, including: learning the learner enhanced embedding representation and the test question enhanced embedding representation through the MLP layer to generate different sample pair interaction relationship representations; wherein the sample pair interaction relationship representations include positive sample pair interaction relationship representations, fine-grained negative sample pair interaction relationship representations and coarse-grained negative sample pair interaction relationship representations; the positive sample pair interaction relationship representations are used to represent the answer situation of the same learner on the same test question, including the answer situation of the test question that is not answered well and the answer situation of the test question that is not answered well. The interactive behavior of answering all the questions correctly; the fine-grained negative sample pair interactive relationship representation is used to characterize the answers of different learners on the same test question, including the interactive behavior of answering the test questions poorly; the coarse-grained negative sample pair interactive relationship representation is used to characterize the answers of different learners on different test questions, including the interactive behavior of answering the test questions poorly; according to the similarity between the positive sample pair interactive relationship representation and the fine-grained negative sample pair interactive relationship representation, a fine-grained loss function is constructed; according to the similarity between the positive sample pair interactive relationship representation and the coarse-grained negative sample pair relationship representation, a coarse-grained loss function is constructed; the fine-grained loss function and the coarse-grained loss function are weighted to obtain the learner contrast loss function.

[0105] In this implementation, see Figure 3 , a multi-layer perceptron (MLP) model is used to capture different interactive relationship representations, as follows:

[0106]

[0107]

[0108]

[0109]

[0110] in, 、 、 、 、 Learners 、 、 The corresponding learner-enhanced embedding representation and test question i j 、i k The corresponding test question is enhanced with embedded representation.

[0111] The interaction relationship between positive samples is represented as , refers to the interactive behavior of the same learner in answering the same test questions, including the interactive behavior of answering the test questions poorly and answering all the test questions correctly, see Figure 3The learner (user) perspective compares the POS in the learning module; the fine-grained negative sample pair interaction relationship is represented as , refers to the interactive behavior of different learners in answering the same test questions, including those who do not answer the test questions well, see Figure 3 The learner (user) perspective contrasts the neg_fine in the learning module; the coarse-grained negative sample pair interaction relationship is represented as , refers to the responses of different learners on different test questions, including the interactive behaviors of those who do not do well on the test questions, see Figure 3 The learner (user) perspective contrasts neg_coarse in the learning module.

[0112] Because fine-grained negative pairs represent interaction relationship representations, indicating that different learners have relatively different preferences for the same test question, a contrastive learning loss function based on positive and fine-grained negative pairs is effective when data is sparse. This fine-grained contrastive learning based on positive and fine-grained negative pairs is also the basis for learning "learners' unique behavioral preferences." To this end, the following fine-grained loss function is constructed to increase the differences in preferences for a single relationship (whether or not they all get the same question right) between different learners, while reducing the differences in the same learner's preference representations for multiple relationships (whether or not they all get the same question right):

[0113]

[0114]

[0115] Among them, f(ξ1,ξ2,ξ3) is the pairwise distance function, ξ1,ξ2,ξ3 are different feature vectors, and σ(▪) is the s-type activation function, which is used to convert the distance into a probability form so that the output of the f function is the probability of similarity. In addition, 、 They represent the set of “do not do well” and “do all right” interaction behaviors, i.e., the corresponding samples. Therefore, when learning “a fixed learner-test pair” Fine-grained representation of multi-relation interaction preferences In contrastive learning, in addition to sampling the same learner For the same test question i j Representation of the interactive relationship between "all correct" and "poor" positive sample pairs In addition, another different learner was randomly sampled For the same test question i j "Not good" at representing interaction relationships with fine-grained negative samples .

[0116] Since coarse-grained negative samples on interaction representation are from different learners The “doing badly” behaviors on different test items (i j , i k ) belong to completely different interaction representations and are quite different, so in order to better distinguish the interaction preferences of different learners, that is, to perform coarse-grained contrastive learning according to the positive sample on the interaction representation and the coarse-grained negative sample on the interaction representation, the following coarse-grained loss function is constructed:

[0117]

[0118] This further increases the preference difference of different learners in “doing badly” on different test items.

[0119] Finally, in order to integrate the effects of fine-grained and coarse-grained contrastive learning, the fine-grained loss function and the coarse-grained loss function are weighted to obtain the learner contrastive loss function:

[0120]

[0121] where β is a hyperparameter.

[0122] Integrating fine-grained and coarse-grained contrastive learning helps the model better understand the personalized preferences of learners, thereby improving the performance and accuracy of personalized recommendations.

[0123] Step S104, according to the learner enhanced embedding representation, the text label representation and the test item representation are obtained by contrastively learning the label information of the test item and the learner who does badly on the test item, and a test item contrastive loss function is constructed according to the text label representation and the test item representation.

[0124] Specifically, by contrastive learning, the label information of the test item (such as knowledge points, difficulty, etc.) is converted into a text label representation, and the representation of the learner who does badly on the test item (such as answering incorrectly or not completing the answer) is obtained. These representations reflect the deficiencies or needs of the learner on a specific test item. Combined with the text label representation and the test item representation, a test item contrastive loss function is constructed, and by minimizing the test item contrastive loss function, the model learns to align the test item representation with the correct label representation, while distinguishing the representations of different test items, ensuring that when recommending test items, the specific content and learning goals of the test items are considered, thereby improving the relevance and accuracy of the recommendations.

[0125] In some embodiments, the learner enhanced embedding representation is enhanced according to the learner, the text label representation and the question representation are obtained by comparing the label information of the question and the learner who does not do the question well through contrastive learning, and a question contrastive loss function is constructed according to the text label representation and the question representation, comprising: extracting the label information of the question through a text encoder to obtain a text label representation; selecting the learner enhanced embedding representation according to a preset selection rule to obtain the learner embedding representation of the learner who does not do the question well; converting the learner embedding representation of the learner who does not do the question well into a question representation through a question encoder; constructing a text label loss function and a question loss function according to the similarity between the text label representation and the question representation; and averaging the text label loss function and the question loss function to obtain the question contrastive loss function.

[0126] Specifically, in order to ensure that the label is clearer and more meaningful in the context, the label l (i) of each question, such as "trigonometric function", needs to be reconstructed into a more descriptive text, obtaining "trigonometric function question". Here, a tokenizer is used to process the text label l (i) and the context description c. The tokenizer is used to decompose the text into semantic units, which may include word, subword or character level representations, to facilitate model understanding and processing. The attribute matrix obtained after conversion is:

[0127]

[0128] wherein, . Next, a text encoder (such as a Transformer model) is used to encode these text descriptions to obtain the text label representation of each label l (i) :

[0129]

[0130] wherein, , N represents the total number of nodes, and D represents the dimension. The learners are sorted according to the answer order of the questions completed by the learners, and for each question i j , the last K learners who failed to correctly answer the question are selected from the learner enhanced embedding representation, and the behavior of these learners can better reflect the difficulty of the current question or the general difficulty of the learners. A group of learner sequences [Z u1 , Z u2 ,..., Z uK ] who do not do well on this question are obtained, which capture the personal feature information of the learners on the question, and if the number of learners is less than K, a specific padding vector unK is used to fill the sequence to ensure that the input sequence length is K. The sequence [Z u1 , Z u2,...,Z uK ] is used as the input of the question encoder to obtain the question representation:

[0131]

[0132] in, I can't do the test well j The embedding of the kth learner (k=1, 2, ...K) is concat, and the embeddings of all K learners who did not do well in the test questions are connected together as the input of the test question encoder, and the test question representation is obtained through Transformer transformation. .

[0133] Then, the relationship between the text label representation and the question representation is calculated. Similarity, construct text label loss function And the test loss function :

[0134]

[0135]

[0136] Among them, L represents the set of all text labels, and the similarity function Measured by dot product, τ is a learnable temperature variable used to scale the logits of the model output layer, which can be any real number in (0, +∞).

[0137] Finally, there is a symmetric loss between the text encoder and the question encoder, so it is necessary to average the text label loss function and the question loss function to obtain the question contrast loss function:

[0138]

[0139] It should also be noted here that the present embodiment uses a soft classifier model (such as a convolutional neural network, a recurrent neural network, etc.) to predict the probability of test questions belonging to different categories. For each test question, the model will output the probability of belonging to each category, and then select the category with the highest probability as the predicted category. For example, in an English test question, there may be different categories such as grammar, vocabulary, and logic, and the model may output probability values ​​such as 0.1, 0.2, and 0.7. These probability values ​​are converted into probability values ​​between 0 and 1 by the softmax function. The predicted category is the category that the model considers to be the most likely, which is used to compare with the true category, and adjust the model parameters according to the comparison results to reduce the error. The model uses the loss function The difference between the predicted probability distribution and the true label is measured to encourage the model to output a high probability corresponding to the true label, which can enhance the contrast learning of the question label text features and the learners who do not do well on the questions.

[0140] In step S105, the learner full-correct question loss function and the learner not-correct question loss function are constructed based on the learner enhanced embedding representation and the question enhanced embedding representation.

[0141] In this embodiment, in order to make the model better understand the needs and preferences of the learners in different situations, it is also necessary to separately construct the corresponding functions for the interaction of the learner full-correct question and the learner not-correct question.

[0142] In some embodiments, the learner full-correct question loss function and the learner not-correct question loss function are constructed based on the learner enhanced embedding representation and the question enhanced embedding representation, including: determining the learner not-correct question score function and the learner full-correct question score function by a fully connected neural network based on the learner enhanced embedding representation and the question enhanced embedding representation; constructing an initial learner not-correct question loss function and an initial learner full-correct question loss function using a binary cross-entropy loss function for the learner not-correct question score function and the learner full-correct question score function, respectively; using mean square error to optimize the preset relative difficulty coefficient to obtain a relative difficulty coefficient loss function for the learner not-correct question and a relative difficulty coefficient loss function for the learner full-correct question; wherein the relative difficulty coefficient represents the mastery of the knowledge points in the questions by the learners; obtaining the learner full-correct question loss function based on the sum of the initial learner full-correct question loss function and the relative difficulty coefficient loss function for the learner full-correct question; obtaining the learner not-correct question loss function based on the sum of the initial learner not-correct question loss function and the relative difficulty coefficient loss function for the learner not-correct question.

[0143] In this embodiment, in order to consider more complex and possibly non-linear interaction relationships between the learners and the questions, a fully connected layer is used here to predict the scores between the learners and the questions. This method not only considers linear relationships, but also captures more detailed and complex interactions. Specifically, the learner enhanced embedding representation and the question enhanced embedding representation are mapped to a score function within a probability range by a fully connected layer to perform score prediction:

[0144]

[0145] wherein, is the learner not-correct question score, is the learner full-correct question score, g b is the learner not-correct question score function, g pThe scoring function is used to score all correct answers. The scoring function uses an activation function such as the sigmoid function. The scores of learner u are ranked according to their weaknesses in test i, and then test i is recommended to learner u for practice based on the ranking results, so as to improve the learner's ability in a more targeted manner. The higher the score, the lower the learner's mastery of the test question. The test questions are sorted from high to low according to the scores and recommended to the corresponding learner u for test question practice.

[0146] In order to further optimize the representation of the interactive relationship between learners and test questions, the binary cross entropy loss function is used to construct the loss function:

[0147]

[0148]

[0149] in, and They represent the interactive sets of questions that the learners cannot answer well and questions that the learners answer correctly, that is, the corresponding samples. and Represents the predicted interaction between the number of questions that the learner did not do well and the number of questions that the learner did correctly, which may include latent or missing data.

[0150] In this embodiment, in order to more carefully capture the details of the learner's mastery of the test questions, it is necessary to define the relative difficulty coefficient d u,i , used to quantify the learner's mastery of a specific test question. u,i It is calculated by comparing the number of incorrect answers to the total number of knowledge points on the test. If a learner gets a large number of incorrect answers, the relative difficulty coefficient is higher, indicating that the learner needs more practice.

[0151] For interactions where learners answer all questions correctly and interactions where learners do not answer questions correctly, the corresponding relative difficulty coefficients are defined as follows:

[0152]

[0153] Among them, n i is the total number of knowledge points tested by question i, The number of knowledge points for which the learner made mistakes.

[0154] Then, for the defined relative difficulty coefficient, the mean square error (MSE) loss function is used to optimize the predicted relative difficulty coefficient. Relative difficulty coefficient to actual The differences between:

[0155]

[0156] wherein the loss function is only for the existing interactions is optimized, p and b represent the interactions of the learner doing all the questions correctly and the learner doing the questions incorrectly, respectively, the predicted value calculated according to the self-defined score function The self-defined score function may adopt an activation function output of a full connection layer. is the relative difficulty coefficient loss function of the learner doing all the questions correctly, is the relative difficulty coefficient loss function of the learner doing the questions incorrectly.

[0157] Finally, the learner doing all the questions correctly loss function and the learner doing the questions incorrectly loss function are obtained according to the following formula:

[0158]

[0159] Step S106, model training is performed according to the learner comparison loss function, the question comparison loss function, the learner doing all the questions correctly loss function and the learner doing the questions incorrectly loss function, to obtain a target recommendation model.

[0160] In the embodiment, the total target loss function of the model is defined as:

[0161]

[0162] wherein γ is a hyperparameter. In the model training process, the total target loss function is minimized, the training of the model is completed, and a target recommendation model is obtained.

[0163] In a second aspect, an exam question recommendation method is provided in the embodiments of the present application, Figure 4 a flowchart of the exam question recommendation method provided in the embodiments of the present application is shown in FIG. 2, and the method comprises the following steps: Figure 4

[0164] Step S201, a target multi-relation interaction graph is obtained; wherein the target multi-relation interaction graph is constructed by taking a target exam question and a to-be-recommended learner as nodes, and taking a to-be-recommended learner doing the target exam question incorrectly interaction matrix and a to-be-recommended learner doing all the target exam questions correctly interaction matrix as edges.

[0165] Step S202, the target multi-relation interaction graph is input into the target recommendation model obtained by the training method of the exam question recommendation model to perform prediction, to obtain a target recommended learner list corresponding to the target exam question. ​

[0166] Step S203, recommend the target test question to the learners in the target recommended learner list.

[0167] In the embodiment, the obtained target multi-relation interaction graph is input into the trained target recommendation model for prediction to generate a target recommended learner list. This list is generated based on the model prediction of the learners' answering needs for the target test question, and is ranked and recommended to the Top-N learners according to the scores, to provide more accurate and personalized test question recommendations, and to ensure that suitable learners can receive test questions suitable for their needs and abilities.

[0168] In order to illustrate the technical effects achieved by the embodiments of the present application, comparative algorithm experiments, ablation studies and hyperparameter tuning of analogous scenarios were carried out on two published platform datasets, aiming to verify the performance and effects of the target recommendation model provided by the present application. In the experiment, Adam was used as the optimizer to initialize the model parameters, and each model was trained from scratch, including the comparative algorithm, while adjusting the hyperparameter settings of each model.

[0169] (1) In the comparative algorithm experiment stage, the target recommendation model of the present application was compared with three benchmark models (Random, Popular and pureMF) and three state-of-the-art related models (LightGC, MultiRec and Multi-VAE). The results showed that the target recommendation model of the present application improved the performance by 34.8% to 65.6% and 2.2% to 29.1% on the two datasets, respectively, which was superior to almost all benchmark models and state-of-the-art related models, especially on relatively sparse datasets.

[0170] (2) For the ablation study, the multi-relation graph convolution module (corresponding to MRGC in Figure 3 ), the learner perspective contrast learning module (corresponding to User-CL in Figure 3 ) and the test question perspective contrast learning module (corresponding to Item-CL in Figure 3The ILCL in the GCN) designed four different models: MR (remove the GCN core module), MRGC, MRGCw / ILCL and MRGCw / User-CL. The research results show that the multi-relation interaction behavior is successfully modeled by MRGC, and User-CL and ILCL significantly improve the performance of the model on the two data sets. Further analysis shows that MRGCw / ILCL has slightly lower performance than MRGC and may cause overfitting, while MRGCw / User-CL significantly improves the performance. The combination of ILCL and User-CL can significantly enhance the representation learning effect, solve the overfitting problem, and thus improve the overall performance of the target recommendation model of the application, and the representation learning performance on the two data sets is improved by 46.8% and 12.1%, respectively.

[0171] (3) In terms of hyperparameter tuning, D=32 is set, and three hyperparameters of the target recommendation model of the application are systematically tuned. The experimental results show that when α=0.01, the target recommendation model performs significantly on the two data sets, showing that the interaction does not affect the propagation of the collaborative signal between the learners and the test questions, although it is smaller than that of the full correct. In addition, on one data set, the performance of β=1.0 and 0.5 is similar, indicating that fine-grained and coarse-grained contrast learning effects are similar, while on the other data set, the performance is best when β=1.0 and 0.5, respectively. On the sparse data set, γ=30 performs well, while on the other data set, γ=10 performs relatively well. Overall, through the tuning of the hyperparameters, excellent performance results are obtained.

[0172] In summary, the target recommendation model of the application has shown significant performance advantages and effects in the recommendation system, especially in handling multi-relation interactions and sparse data sets.

[0173] It should be noted that the steps shown in the above process or the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from here.

[0174] In a third aspect, the embodiments of the application provide a test question recommendation model training system, which is used to implement the above embodiments and preferred embodiments, and has been described. As used below, the terms "module", "unit", "sub-unit" and the like can be a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, hardware or a combination of software and hardware is also possible and is conceived.

[0175] Figure 5A block diagram of a test question recommendation model training system provided by an embodiment of the present application is shown in Figure 5 The system comprises:

[0176] An interaction graph acquisition module 301 is configured to acquire a multi-relation interaction graph, wherein the multi-relation interaction graph is constructed by taking test questions and learners as nodes and taking a learner's incorrect test question interaction matrix and a learner's all-correct test question interaction matrix as edges; the learner's incorrect test question interaction matrix represents at least one knowledge point corresponding to a test question that a learner answers incorrectly, and the learner's all-correct test question interaction matrix represents all knowledge points corresponding to a test question that a learner answers correctly.

[0177] A node enhancement module 302 is configured to enhance each node in the multi-relation interaction graph by using a preset multi-relation graph convolution to obtain an enhanced embedding representation corresponding to the node; wherein the enhanced embedding representation comprises a learner enhanced embedding representation and a test question enhanced embedding representation.

[0178] A learner comparison module 303 is configured to construct different sample pair interaction relationship representations according to the learner enhanced embedding representation and the test question enhanced embedding representation, to compare the sample pair interaction relationship representations to learn the preferences of different learners, and to construct a learner comparison loss function.

[0179] A test question comparison module 304 is configured to obtain a text label representation and a test question representation by comparing the label information of a test question and a learner who answers the test question incorrectly according to the learner enhanced embedding representation, and to construct a test question comparison loss function according to the text label representation and the test question representation.

[0180] A loss function construction module 305 is configured to construct a learner all-correct test question loss function and a learner incorrect test question loss function according to the learner enhanced embedding representation and the test question enhanced embedding representation.

[0181] A model training module 306 is configured to perform model training according to the learner comparison loss function, the test question comparison loss function, the learner all-correct test question loss function, and the learner incorrect test question loss function to obtain a target recommendation model.

[0182] It should be noted that each of the above modules can be a functional module or a program module, and can be implemented by software or hardware. For modules implemented by hardware, each of the above modules can be located in the same processor; or each of the above modules can also be located in different processors in any combination.

[0183] In a fourth aspect, an electronic device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the training method of the test question recommendation model or the test question recommendation method when executing the computer program.

[0184] Optionally, the electronic device can further include a transmission device and an input and output device, wherein the transmission device is connected with the processor, and the input and output device is connected with the processor.

[0185] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation manners, and the embodiment will not be described here.

[0186] In a fifth aspect, in combination with the training method of the test question recommendation model or the test question recommendation method in the above embodiments, an embodiment of the present application can provide a storage medium to implement. The storage medium has a computer program stored therein; the computer program is executed by a processor to implement the training method of any test question recommendation model in the above embodiments or the test question recommendation method.

[0187] In an embodiment, a computer device is provided, which can be a terminal. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. The processor of the computer device is configured 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 operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement the training method of the test question recommendation model or the test question recommendation method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0188] In an embodiment, Figure 6 The internal structure diagram of the electronic device provided by the embodiment of the present application is shown in Figure 6 The internal structure diagram of the electronic device provided by the embodiment of the present application is shown in Figure 6As shown. The electronic device includes a processor, a network interface, an internal memory, and a non-volatile memory connected via an internal bus, wherein the non-volatile memory stores an operating system, a computer program, and a database. The processor is used to provide computing and control capabilities, the network interface is used to communicate with an external terminal via a network connection, the internal memory is used to provide an environment for the operation of the operating system and the computer program, and when the computer program is executed by the processor, it implements the training method of the test question recommendation model or the test question recommendation method described above, and the database is used to store data.

[0189] Those skilled in the art will understand that Figure 6 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the electronic device to which the solution of the present invention is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0190] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0191] Those skilled in the art should understand that the various technical features of the above-described embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the various technical features in the above-described 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.

[0192] The above merely describes the preferred embodiments of the present application, and it should be pointed out that, for those skilled in the art, several improvements and replacements can be made without departing from the technical principles of the present application, and these improvements and replacements should also be considered as the protection scope of the present application.

Claims

1. A training method for a test question recommendation model, characterized in that: The training method comprises: Obtain a multi-relationship interaction graph; the multi-relationship interaction graph is constructed using test questions and learners as nodes, and an interaction matrix for test questions that the learner did not answer correctly and an interaction matrix for test questions that the learner answered correctly as edges; the interaction matrix for test questions that the learner did not answer correctly represents at least one knowledge point corresponding to the test question that the learner answered incorrectly, and the interaction matrix for test questions that the learner answered correctly represents all knowledge points corresponding to the test questions that the learner answered correctly; Using a preset multi-relation graph convolution to enhance each of the nodes in the multi-relation interaction graph, to obtain an enhanced embedding representation corresponding to the node; wherein the enhanced embedding representation includes a learner enhanced embedding representation and a test question enhanced embedding representation; constructing different sample pair interaction relationship representations based on the learner enhanced embedding representation and the test question enhanced embedding representation, comparing the sample pair interaction relationship representations to learn preferences between different learners, and constructing a learner comparison loss function; According to the learner enhanced embedding representation, a text label representation and a test question representation are obtained by comparing the label information of the learned test questions with the learner who did not do well in the test questions, and a test question contrast loss function is constructed based on the text label representation and the test question representation; Constructing a loss function for all correct questions and a loss function for poorly answered questions based on the learner-enhanced embedding representation and the question-enhanced embedding representation; Model training is performed based on the learner comparison loss function, the test question comparison loss function, the learner's correct test question loss function, and the learner's poor test question loss function to obtain a target recommendation model.

2. The training method according to claim 1, characterized in that Before enhancing each of the nodes in the multi-relationship interaction graph using a preset multi-relationship graph convolution to obtain an enhanced embedding representation corresponding to the node, the training method further includes: Constructing a multi-relation adjacency matrix based on the interaction matrix of test questions that the learner did not answer well and the interaction matrix of test questions that the learner answered correctly; Based on the multi-relation adjacency matrix, each of the nodes in the multi-relation interaction graph is enhanced using a preset multi-relation graph convolution to obtain an enhanced embedding representation corresponding to the node.

3. The training method according to claim 2, characterized in that The constructing of a multi-relation adjacency matrix based on the interaction matrix of test questions that the learner did not answer well and the interaction matrix of test questions that the learner answered correctly includes: constructing a sparse interaction matrix based on the interaction matrix of the test questions that the learner did not answer well and the interaction matrix of the test questions that the learner answered correctly; The sparse interaction matrix is ​​transformed to obtain the multi-relation adjacency matrix.

4. The training method according to claim 2, characterized in that The method of enhancing each node in the multi-relationship interaction graph based on the multi-relationship adjacency matrix using a preset multi-relationship graph convolution to obtain an enhanced embedding representation corresponding to the node includes: Obtaining an input vector of the node; Extracting an initial feature embedding vector of the node using a preset nonlinear co-embedding function according to the input vector; In a convolution layer of any layer of the multi-relation graph convolution, information of neighbor nodes of any node in the multi-relation interaction graph is aggregated through the multi-relation adjacency matrix to obtain an updated embedding vector of the any node; Traverse each convolutional layer, and for each node, determine the average value of the updated embedding vectors corresponding to the node in all the convolutional layers to obtain the enhanced embedding representation corresponding to the node.

5. The training method according to claim 1, wherein: The step of constructing different sample pair interaction relationship representations based on the learner-enhanced embedding representation and the question-enhanced embedding representation, comparing the sample pair interaction relationship representations to learn preferences between different learners, and constructing a learner comparison loss function includes: The learner enhanced embedding representation and the test question enhanced embedding representation are learned through a multi-layer perceptron model to generate different sample pair interaction relationship representations; wherein the sample pair interaction relationship representations include positive sample pair interaction relationship representations, fine-grained negative sample pair interaction relationship representations, and coarse-grained negative sample pair interaction relationship representations; the positive sample pair interaction relationship representations are used to represent the same learner's answer situation on the same test question, including the interactive behavior of answering the test question poorly and answering all the test questions correctly; the fine-grained negative sample pair interaction relationship representations are used to represent the answer situation of different learners on the same test question, including the interactive behavior of answering the test question poorly; the coarse-grained negative sample pair interaction relationship representations are used to represent the answer situation of different learners on different test questions, including the interactive behavior of answering the test question poorly; Constructing a fine-grained loss function according to the similarity between the positive sample pair interaction relationship representation and the fine-grained negative sample pair interaction relationship representation; Constructing a coarse-grained loss function based on the similarity between the positive sample pair interaction relationship representation and the coarse-grained negative sample pair relationship representation; The fine-grained loss function and the coarse-grained loss function are weighted to obtain the learner contrast loss function.

6. The training method according to claim 1, characterized in that The method enhances the embedding representation of the learner, obtains a text label representation and a test question representation by comparing the label information of the learning test question with the learner who did not do well in the test question, and constructs a test question contrast loss function based on the text label representation and the test question representation, including: Extracting label information of the test question through a text encoder to obtain a text label representation; Selecting the learner enhanced embedded representation according to a preset selection rule to obtain the learner embedded representation who did not do well in the test; The embedding representation of the learner who failed to answer the test question is converted into a test question representation by a test question encoder; Constructing a text label loss function and a test question loss function according to the similarity between the text label representation and the test question representation; The text label loss function and the test question loss function are averaged to obtain the test question comparison loss function.

7. The training method according to claim 1, characterized in that The step of constructing a loss function for all correct questions and a loss function for poorly answered questions based on the learner's enhanced embedding representation and the question's enhanced embedding representation includes: Determining, by a fully connected neural network, a score function for test questions that the learner does not answer well and a score function for test questions that the learner answers correctly based on the learner's enhanced embedded representation and the test question's enhanced embedded representation; Using the binary cross entropy loss function to construct the initial learner's poorly done test question loss function and the initial learner's fully correct test question loss function for the learner's poorly done test question score function and the learner's fully correct test question score function respectively; The preset relative difficulty coefficient is optimized using mean square error to obtain a relative difficulty coefficient loss function for questions that the learner cannot answer well and a relative difficulty coefficient loss function for questions that the learner answers correctly. The relative difficulty coefficient represents the learner's mastery of the knowledge points in the question. The loss function of the learner getting all the questions right is obtained by summing the loss function of the initial learner getting all the questions right and the loss function of the relative difficulty coefficient of the questions the learner gets all the questions right; the loss function of the learner getting the questions wrong is obtained by summing the loss function of the initial learner getting the questions wrong and the loss function of the relative difficulty coefficient of the questions the learner gets the questions wrong.

8. A test question recommendation method, characterized in that: The recommended methods include: Obtain a target multi-relationship interaction graph; wherein the target multi-relationship interaction graph is constructed with target test questions and learners to be recommended as nodes, and with interaction matrices of target test questions that the learners to be recommended did not answer correctly and interaction matrices of target test questions that the learners to be recommended answered correctly as edges; Inputting the target multi-relationship interaction graph into the target recommendation model obtained by the training method of the test question recommendation model according to any one of claims 1 to 7 for prediction, and obtaining a target recommended learner list corresponding to the target test question; The target test question is recommended to the learners in the target recommended learner list.

9. A training system for a test question recommendation model, characterized in that: The training system comprises: An interaction graph acquisition module is configured to acquire a multi-relationship interaction graph, wherein the multi-relationship interaction graph is constructed using test questions and learners as nodes, and an interaction matrix for test questions that the learner failed to answer correctly and an interaction matrix for test questions that the learner answered correctly as edges. The interaction matrix for test questions that the learner failed to answer correctly represents at least one knowledge point corresponding to the test question that the learner answered incorrectly, and the interaction matrix for test questions that the learner answered correctly represents all knowledge points corresponding to the test questions that the learner answered correctly. A node enhancement module is configured to enhance each node in the multi-relationship interaction graph using a preset multi-relationship graph convolution to obtain an enhanced embedding representation corresponding to the node; wherein the enhanced embedding representation includes a learner enhanced embedding representation and a test question enhanced embedding representation; A learner comparison module is configured to construct different sample-pair interaction relationship representations based on the learner-enhanced embedding representation and the question-enhanced embedding representation, compare the sample-pair interaction relationship representations to learn the preferences of different learners, and construct a learner comparison loss function; A test question comparison module is used to enhance the embedding representation based on the learner, obtain a text label representation and a test question representation by comparing the label information of the learned test questions with the learner who did not do well on the test questions, and construct a test question comparison loss function based on the text label representation and the test question representation; A loss function construction module is used to construct a loss function for the learner to answer all the questions correctly and a loss function for the learner to answer the questions poorly based on the learner enhanced embedding representation and the question enhanced embedding representation; The model training module is used to perform model training based on the learner comparison loss function, the test question comparison loss function, the learner's correct test question loss function and the learner's poor test question loss function to obtain a target recommendation model.

10. An electronic device, characterized in that: include Memory, processor, and A computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the training method for the test question recommendation model as described in any one of claims 1 to 7 or the test question recommendation method as described in claim 8 is implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the training method of the test question recommendation model according to any one of claims 1 to 7 or the test question recommendation method according to claim 8 is implemented.

Images