Memory-enhanced attention knowledge tracking method and system for skill switching phenomenon

Abstract

The application discloses a memory-enhanced attention knowledge tracking method and system for skill switching phenomenon, and the method comprises the following steps: embedding an exercise label e t , a skill label s t and a response label r t of a time step t to obtain an exercise embedding x t and a knowledge embedding y t ; performing attention knowledge tracking on the exercise embedding x t and the knowledge embedding y t to obtain a context knowledge state, and simultaneously adopting a skill-aware memory-enhanced recurrent neural network (SMRNN) to extract memory knowledge according to the input exercise embedding x t ; integrating the context knowledge state and the memory knowledge to obtain a final knowledge state h t of the time step t; inputting a splicing vector of the final knowledge state h t of the time step t and the exercise embedding x t into a prediction network to obtain a prediction result; and the application can save, extract and update long-term memory knowledge of learners during learning, so that better learner response prediction performance can be achieved when the skill switching phenomenon occurs.

Description

Technical Field

[0001] This invention belongs to the field of online learning technology in online learning platforms, and specifically relates to a memory enhancement, attention, and knowledge tracking method and system for skill switching phenomena. Background Technology

[0002] With the widespread adoption of online learning platforms, learners can acquire knowledge through self-study without leaving home. To recommend suitable learning content, online learning platforms need to accurately understand learners' knowledge status at different times, i.e., their mastery of skills (knowledge concepts or knowledge components). Modeling learners' knowledge status is typically based on Knowledge Tracking (KT) technology. Knowledge Tracking is a crucial task in online learning; its goal is to model each learner's knowledge status (KS) by predicting their responses based on their interaction history with the platform. In a learning platform, learners can learn related skills by completing specific exercises (e.g., if "addition" is a skill, then "1+1" is one of its exercises). The platform tracks learners' knowledge status regarding the learned skills using a KT model. Finally, the platform uses a "when-to-stop" strategy (Joseph Rollinson and Emma Brunskill, "From predictive models to instructional policies," International Educational Data Mining Society, 2015) to determine whether learners have mastered these skills.

[0003] Deep learning possesses powerful non-linear implementation and feature extraction capabilities, making it well-suited for modeling complex learning processes, especially when large amounts of learning interaction data are available. In recent years, numerous research works on Deep Learning-based Knowledge Tracing (DLKT) models have been proposed and have achieved considerable success. As the first application of deep learning in knowledge tracing, DKT utilizes recurrent neural networks (e.g., LSTM) to model the learner's knowledge state during practice to predict unknown responses. However, due to the nature of the network, recurrent neural networks inevitably experience forgetting over long time sequences.

[0004] Inspired by computer architecture, external memory units have been shown to enhance a network's ability to capture long-term dependencies. Memory-augmented neural networks (MANNs) extend traditional recurrent neural networks by setting up a single memory matrix or two static memory matrices. DKVMN (Jiani Zhang, Xingjian Shi, Irwin King, and Dit-YanYeung, “Dynamic key-value memory networks for knowledge tracing,” in Proceedings of the 26th international conference on World Wide Web, 2017, pp.765–774.) further extends memory-augmented neural networks using a static key-memory matrix and a dynamic value-memory matrix. These two matrices are used not only to store and update the learner's dynamic knowledge state during the learning process but also to mine potential correlations between skills. While integrating external memory can help RNN models capture longer-term dependencies, the performance of the most recent practice still has a significant impact on modeling the learner's current knowledge state. Especially when the most recent practice was a guess or a mistake, the result can negatively affect subsequent response predictions.

[0005] Attention mechanisms are a more flexible deep learning technique than recurrent neural networks and memory-enhanced neural networks, and they excel in natural language processing tasks. Multi-head attention mechanisms are often used to model learners' contextual knowledge and predict their unknown responses. SAKT (Shalini Pandey and George Karypis, “A self-attentive model for knowledge tracing,” arXiv preprint arXiv:1907.06837,2019.) identifies skills related to the target skill from learners' past activities, thus predicting their mastery of the next skill using information about their past performance on that skill. AKT (AritraGhosh, Neil Heffernan, and Andrew S. Lan, “Context-aware attentive knowledge tracing,” in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2020, pp. 2330–2339.) also employs multi-head self-attention layers for knowledge tracing, where attention weights decay exponentially based on context-aware relative distance measurements. Furthermore, inspired by the Rasch model, AKT embeds `practice-skill` and `practice-skill-response`, avoiding over-parameterization and overfitting issues. RKT (Shalini Pandey and Jaideep Srivastava, “Rkt: Relation-aware self-attention for knowledge tracing,” in Proceedings of the 29th ACM International Conference on Information & Knowledge Management, 2020, pp. 1205–1214.) extends SAKT by introducing a relation-aware self-attention layer that includes contextual information.SAINT (Youngduck Choi, Youngnam Lee, Junghyun Cho, Jineon Baek, Byungsoo Kim, Yeongmin Cha, Dongmin Shin, Chan Bae, and Jaewe Heo, “Towards an appropriate query, key, and value computation for knowledge tracing,” in Proceedings of the Seventh ACM Conference on Learning@Scale, 2020, pp. 341–344.) is a Transformer-based knowledge tracing model that processes practice information and learner response information separately. Specifically, it inputs the practice stream of learner interactions into the encoder, and the decoder calculates the final output sequence of the model together with the corresponding response sequence obtained from the encoder. SAINT+ (Dongmin Shin, Yugeun Shim, Hangyeol Yu, SeewooLee, Byungsoo Kim, and Youngduck Choi, “Saint+: Integrating temporal features for ednet correctness prediction,” in LAK21: 11th International Learning Analytics and Knowledge Conference, 2021, pp. 490–496.) extends SAINT by incorporating two temporal feature embeddings into the response embedding: elapsed time (i.e., the time taken to complete a task) and lag time (the time interval between adjacent interactions).

[0006] Because learners' performance in recent practice sessions is a better indicator of the stability of their knowledge than their inconsistent performance in the most recent practice session, models based on multi-head attention mechanisms generally outperform those based on recurrent neural networks in predicting learner responses. However, in multi-skill learning, the learned skills often switch irregularly, a phenomenon known as skill switching phenomenon (SSP). When skill switching occurs, if the correlation between the current skill and the learner's recently learned skills is low, especially if the practice related to the current skill was already encountered by the learner earlier, then models based on multi-head attention mechanisms will be unable to accurately predict the response to the current practice. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a memory-enhancing attention knowledge tracking method and system for skill switching phenomena, which addresses the above-mentioned problems in the prior art. This invention can save, extract and update learners' long-term memory knowledge during the learning process, thereby achieving better learner response prediction performance when skill switching occurs.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A memory-enhancing attention-based knowledge tracking method for skill switching phenomena includes:

[0010] S101, time step t Practice tags Skill tags and response tags Embedding to practice embedding and knowledge embedding The response label refers to the actual feedback label submitted by the learner.

[0011] S102, Practice Embedding and knowledge embedding Attention-based knowledge tracking to obtain contextual knowledge state Simultaneously, a Skill-Aware Memory Reinforced Recurrent Neural Network (SMRNN) is employed to embed exercises based on the input. Retrieving knowledge from memory ; Contextual knowledge state Memory of knowledge Integration yields time steps t final state of knowledge ;

[0012] S103, time step t final state of knowledge and practice embedding The concatenated vectors are input into a pre-trained prediction network to obtain the learner's correct response to the practice labels. Skill tags Prediction results .

[0013] Optionally, the practice embedding is obtained in step S101. The function expression is:

[0014] ,

[0015] In the above formula, For activation function, Skill tags The original embedding, To influence the parameters, Indicates practice label The original embedding, D For embedded dimensions, For practice tags, For skill tags, For the response tag, and has:

[0016] ,

[0017] ,

[0018] ,

[0019] In the above formula, and These represent the one-hot encoded vectors for skills and practice, respectively. and They are respectively and The embedding matrix; for Embedded vector, The total number of practice labels, S This indicates the number of skill tags on the online learning platform. D For the embedded dimension.

[0020] Optionally, the knowledge embedding is obtained in step S101. The function expression is:

[0021] ,

[0022] In the above formula, For activation function, Represents the primitive embedding of skill-response pairs. To influence the parameters, The variable vector representing the original embedding of the skill-response pair. D For embedded dimensions, For practice tags, For skill tags, For the response tag, and has:

[0023] ,

[0024] ,

[0025] In the above formula, and These represent the multi-hot encoded vectors for skill-response and practice-response, respectively. and They are respectively and The embedding matrix, the multi-hot encoded vector of skill-response is used to encode skill tags. One-hot encoding and response tags The practice-response multi-hot encoded vector is obtained by concatenating one bit of binary code (0 or 1); the practice-response multi-hot encoded vector is obtained by concatenating the practice labels. One-hot encoding and response tags It is obtained by concatenating one binary code of 0 or 1.

[0026] Optionally, in step S102, a skill-aware memory-enhanced recurrent neural network (SMRNN) is used to embed the input practice. Retrieving knowledge from memory include:

[0027] S201, Embedding based on practice Calculate the practice labels Relevance weight vector with all potential concepts :

[0028] ,

[0029] In the above formula, For activation function, For the key-memory matrix used to store latent concepts, C Indicates the number of potential concepts. D For embedded dimensions, For practicing embedding;

[0030] S202, retrieve the memorized knowledge according to the following formula. :

[0031] ,

[0032] In the above formula, For learners in time step t The value-memory matrix is ​​used to represent the learner's performance at time steps. t Key-memory matrix A matrix showing the level of mastery of each potential concept in the text;

[0033] S203, update the learner's time step according to the following formula. t+ The value of 1 - memory matrix :

[0034] ,

[0035] In the above formula, Update weights for attention. For the retrieved memory knowledge, Represents the vector product. and The vector product is a shape of The matrix, For practice tags The relevance weight vector with all potential concepts. Add to the matrix. For learners in time step t The value of the memory matrix, and the expression for the function to calculate the attention update weights:

[0036] ,

[0037] in, An embedding matrix representing the weights for skill-aware attention updates. S This indicates the number of skill tags on the online learning platform.

[0038] Optionally, the time step is obtained in step S102. t final state of knowledge The function expression is:

[0039] ,

[0040] In the above formula, For contextual knowledge state, To memorize knowledge, Add to the matrix. For practice tags Context-aware attention relevance weights between the learned content and its context. These are learnable relative parameters.

[0041] Optionally, context-aware attention relevance weights The expression for the computation function is:

[0042] ,

[0043] In the above formula, To apply the multi-head self-attention mechanism to the deadline time step Contextual Exercises The representation, For practicing embedding, With a fixed time step, and having , These represent the time steps. Practice embedding in context-aware mode Value embedding, for The self-attention weights for each time step are calculated as follows:

[0044] ,

[0045] In the above formula, For activation function, Indicates at time step Practice embedding in context-aware exercises Query embedding, Indicates at time step Practice embedding in context-aware exercises Key embedding, D For the embedded dimension.

[0046] Optionally, the prediction network used in step S103 is a three-layer response prediction network, and its functional expression is:

[0047] ,

[0048] ,

[0049] ,

[0050] In the above formula, and It is an intermediate variable in a three-layer response prediction network; For the prediction result, { , , }and{ , , } represent the parameters of the transformation matrix and bias vector in the corresponding prediction network, respectively. This is a vector concatenation operation. For practicing embedding.

[0051] Optionally, the three-layer response prediction network and the skill-aware memory-enhanced recurrent neural network (SMRN) are trained end-to-end, and the loss function used during training has the following expression:

[0052] ,

[0053] In the above formula, Represents the loss function. In response, The results are the predictions from the three-layer response prediction network.

[0054] Furthermore, the present invention also provides a memory-enhanced attentional knowledge tracking system for skill switching phenomena, comprising a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute the memory-enhanced attentional knowledge tracking method for skill switching phenomena.

[0055] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program for being programmed or configured by a microprocessor to perform the memory-enhanced attentional knowledge tracking method for skill switching phenomena.

[0056] Compared with the prior art, the present invention has the following main advantages: the method of the present invention includes setting time steps t Practice tags Skill tags and response tags Embedding to practice embedding and knowledge embedding ; Embedding exercises and knowledge embedding Attention-based knowledge tracking to obtain contextual knowledge state Simultaneously, a Skill-Aware Memory Reinforced Recurrent Neural Network (SMRNN) is employed to embed exercises based on the input. Retrieving knowledge from memory ; Contextual knowledge state Memory of knowledge Integration yields time steps t final state of knowledge ; take time steps t final state of knowledge and practice embedding The concatenated vectors are input into the prediction network to obtain the prediction results. This invention can preserve, retrieve, and update learners' long-term memory knowledge during the learning process, thereby achieving better learner response prediction performance when skill switching occurs. Attached Figure Description

[0057] Figure 1 This is an example of a learner interaction sequence given in the explanation of the skill switching phenomenon in the embodiments of the present invention.

[0058] Figure 2 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0059] Figure 3 This is a schematic diagram of the network structure of the method in an embodiment of the present invention.

[0060] Figure 4 This is a structural diagram of practice embedding and knowledge embedding based on TPEN in an embodiment of the present invention.

[0061] Figure 5 This is a schematic diagram of the Skill-Aware Memory Enhanced Recurrent Neural Network (SMRNN) in an embodiment of the present invention.

[0062] Figure 6 This is a schematic diagram of the AUC performance of practice response prediction under the skill switching phenomenon in an embodiment of the present invention.

[0063] Figure 7 This is a schematic diagram of the ACC performance of practice response prediction under the skill switching phenomenon in an embodiment of the present invention. Detailed Implementation

[0064] The so-called skill switching phenomenon refers to the irregular switching of skills learned by learners during the process of learning multiple skills online. Figure 1 A learner interaction sequence of length 15 is given. e 1~ e 15 , e 1~ e 3. e 13 and e 14 To learn the first skill, e 4~ e 8. e 15 To learn a second skill, e 9~ e 12 To learn the third skill, a total of four skill switches occurred in this learner interaction sequence. The exercises corresponding to these four skill switches are as follows: e 4. e 9. e 13 and e 15 Due to practice e 13 and e 15 The corresponding skills have been learned previously, therefore these skills are called repeated switching skills. Skill switching latency (SSL) refers to the time step difference between two interactions with the same skill tag. For repeated switching skills, the interval between the current time step and the time step where the skill last appeared is called the switching skill latency of the repeated switching skill; for non-repeated switching skills, since the skill is appearing in the learning sequence for the first time, the switching latency of these skills is set to a fixed value of 0; for other non-switching skills, since there is no latency between these skills and their last time step, the switching latency of these skills is set to a fixed value of 1. Figure 1 In the middle, practice e 13 and e 15 Skill switching delay SSL e13 and SSL e15They are: SSL e13 =13-3=10 and SSL e15 =15-8=7; due to practice e 1. e 4 and e Skill 9 appears for the first time in the sequence, so its skill switching latency is 0; the skill switching latency for the remaining exercises is 1. Statistical analysis of skill switching in the ASSISTments2017 dataset reveals that skill switching is relatively common in these learners' interaction sequences. In general, the longer the learner's interaction sequence, the more skills are switched (including repeated skill switching), and the greater the skill switching latency.

[0065] like Figure 2 and Figure 3 As shown, the memory-enhanced attention-based knowledge tracking method for skill switching phenomena in this embodiment includes:

[0066] S101, Model Embedding: Incorporating Time Steps t Practice tags Skill tags and response tags Embedding to practice embedding and knowledge embedding The response label refers to the actual feedback label submitted by the learner.

[0067] S102, Knowledge Modeling: Embedding Exercises and knowledge embedding Attention-based knowledge tracking to obtain contextual knowledge state Simultaneously, a Skill-Aware Memory Reinforced Recurrent Neural Network (SMRNN) is employed to embed exercises based on the input. Retrieving knowledge from memory ; Contextual knowledge state Memory of knowledge Integration yields time steps t final state of knowledge ;

[0068] S103, Response Prediction: Time Steps t final state of knowledge and practice embedding The concatenated vectors are input into a pre-trained prediction network to obtain the learner's correct response to the practice labels. Skill tags Prediction results .

[0069] In this embodiment, the complete network consisting of the model embedding, knowledge modeling, and response prediction networks is named the Memory-aware Attentive Network (MAN).

[0070] To address the limitation of existing RME embedding models (Aritra Ghosh, Neil Heffernan, and Andrew S Lan, “Context-aware attentive knowledge tracing,” in Proceedings of the 26th ACMSIGKDD International Conference on Knowledge Discovery & Data Mining, 2020, pp.2330–2339.) in fully extracting features from practice data when the data is not sparse, this embodiment proposes a two-layer proportional embedding network (TPEN) for model embedding. Figure 4 As shown in (a), in step S101 (through a two-layer proportional embedding network TPEN), the practice embedding is obtained. The function expression is:

[0071] ,

[0072] In the above formula, For activation function, Skill tags The original embedding, To influence the parameters, Indicates practice label The original embedding, D For embedded dimensions, For practice tags, For skill tags, For the response tag, and has:

[0073] ,

[0074] ,

[0075] ,

[0076] In the above formula, and These represent the one-hot encoded vectors for skills and practice, respectively. and They are respectively and The embedding matrix; for Embedded vector, The total number of practice labels, S This indicates the number of skill tags on the online learning platform. D For the embedded dimension.

[0077] like Figure 4 As shown in (b), in step S101 (through a two-layer proportional embedding network TPEN), the knowledge embedding is obtained. The function expression is:

[0078] ,

[0079] In the above formula, For activation function, Represents the original embedding of skill-response pairs. To influence the parameters, The variable vector representing the original embedding of the skill-response pair. D For embedded dimensions, For practice tags, For skill tags, For the response tag, and has:

[0080] ,

[0081] ,

[0082] In the above formula, and These represent the multi-hot encoded vectors for skill-response and practice-response, respectively. and They are respectively and The embedding matrix, the multi-hot encoded vector of skill-response is used to encode skill tags. One-hot encoding and response tags The practice-response multi-hot encoded vector is obtained by concatenating one bit of binary code (0 or 1); the practice-response multi-hot encoded vector is obtained by concatenating the practice labels. One-hot encoding and response tags It is obtained by concatenating one bit of binary code (0 or 1). Through the two-layer proportional embedding network TPEN described above, it is possible to fully utilize the practiced representations while avoiding overparameterization and overfitting.

[0083] In this embodiment, step S102 involves embedding the practice. and knowledge embedding Attention-based knowledge tracking to obtain contextual knowledge state This can be achieved using the Context-aware Attentive Knowledge Tracing (AKT) network (Aritra Ghosh, Neil Heffernan, and Andrew S Lan, “Context-aware attentive knowledge tracing,” in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2020, pp. 2330–2339.), or other existing attention-based knowledge tracing models can be used as needed.

[0084] The Contextual Knowledge Modeling Network (AKT) considers not only the correlation between the current practice and the contextual practice, but also the forgetting mechanism during the learning process, in the process of modeling the learner's knowledge state. Compared with other similar models, the AKT is a more reasonable contextual knowledge modeling network. In this embodiment, step S102 involves embedding the practice... and knowledge embedding Attention-based knowledge tracking to obtain contextual knowledge state The contextual knowledge modeling network AKT utilizes three components: the practice encoder EE, the knowledge encoder KE, and the knowledge retrieval unit KR.

[0085] The encoder EE is practiced with time steps 1 to t for the embedding. ~ sequence As input, a context-aware practice embedding sequence is output using a monotonic attention mechanism. And time step t Context-aware practice embedding The expression for the computation function is:

[0086] ,

[0087] In the above formula, For time steps t - i Context-aware practice embedding time steps t Context-aware practice embedding Value embedding, For time steps t - i Self-attention weights For a fixed time step, and with the following:

[0088] ,

[0089] In the above formula, For time steps t Context-aware practice embedding time steps t Context-aware practice embedding Query embedding, For time steps t -i Context-aware practice embedding time steps t Context-aware practice embedding Key embedding.

[0090] Knowledge Encoder KE uses time step 1~ t Embedding Exercises ~ sequence As input, the same monotonic attention mechanism is used to output a context-aware sequence of knowledge embeddings. And at any time step t Context-aware knowledge embedding The expression for the computation function is:

[0091] ,

[0092] In the above formula, For time steps t - i Context-aware practice embedding time steps t Contextual knowledge practice embedding Value embedding, For time steps t - i Self-attention weights For a fixed time step, and with the following:

[0093] ,

[0094] In the above formula, Embedding time steps for context-aware practice of time step t t Contextual knowledge practice embedding Query embedding, Embedding time steps for context-aware practice of time step ti t Contextual knowledge practice embedding Key embedding.

[0095] Knowledge retrieval system KR uses context-aware embedded sequences for practice and knowledge embedding sequence As input, a novel monotonic attention mechanism with exponential decay is used to output the retrieved contextual knowledge state for the current exercise. The basic form of the decay term is an exponential decay curve over time. When past practice is highly similar to the current practice, it may peak at the corresponding time step, as shown below:

[0096] ,

[0097] In the above formula, For time steps t - i Context-aware practice embedding time steps t Contextual knowledge state Value embedding, For time steps t - i Self-attention weights For a fixed time step, and with the following:

[0098] ,

[0099] In the above formula, the middle It is a learnable decay rate parameter. It is a time step and Context-aware temporal distance metric between them Embedding time steps for context-aware practice of time step t t Contextual knowledge state Query embedding, Embedding time steps for context-aware practice of time step ti t Contextual knowledge state The key is embedded, and it has:

[0100] ,

[0101] ,

[0102] In the above formula, Indicates a time step. As an intermediate variable, Indicates time step i +1, Embedding time steps for context-aware practice of time step t t Contextual knowledge state Query embedding, For time steps Context-aware practice embedding time steps t Contextual knowledge state Key embedding.

[0103] To address the inaccuracy in knowledge state modeling caused by skill switching, it's necessary to retrieve relevant information from the learner's existing knowledge memory. Therefore, this embodiment of MAN proposes a novel Skill-aware Memory-augmented Recurrent Neural Network (SMRNN) to store knowledge memories during the learning process. Unlike previous memory-augmented knowledge tracking networks, SMRNN upgrades the knowledge update process by designing a skill-aware attention mechanism. The motivation behind this mechanism is that different skills, due to their varying difficulty, are associated with different learning rates. (See also...) Figure 5 It can be seen that SMRNN in t The time-step operation involves three stages: relevance acquisition, memory retrieval, and memory update. Specifically, SMRNN first relies on practice embeddings... Bond-memory matrix The relevant weights are calculated; then the calculated relevant weights are used to log the value-memory matrix. Reading and writing are performed to predict unknown responses and update the knowledge state, respectively. and These are used to store latent concepts and the learner's level of mastery of each latent concept, respectively. C This represents the number of latent concepts, which can be manually set as a model hyperparameter. The three stages are explained below: Relevance Acquisition: Assuming all skills in an online learning platform share several latent concepts, each skill is likely related to one or more of these latent concepts. For a specific exercise, the SMRNN first calculates the relevance weight vector between that exercise and all latent concepts (…). Then SMRNN will use the calculated The knowledge memory related to the input practice is retrieved and updated separately in the memory retrieval and memory update stages. Memory retrieval: In this stage, the SMRNN uses the relevance weight vector calculated in the relevance acquisition stage. Retrieve knowledge memories of latent concepts related to input practice, and retrieve the retrieved knowledge memories ( This will participate in the knowledge integration process in the next section. Memory Update: In this stage, the SMRNN will update based on the relevant weight vectors. Log-valued memory matrix Personalize and update skills.

[0104] In this embodiment, step S102 uses a skill-aware memory augmented recurrent neural network (SMRNN) to embed the input practice data. Retrieving knowledge from memory include:

[0105] S201, Embedding based on practice Calculate the practice labels Relevance weight vector with all potential concepts :

[0106] ,

[0107] In the above formula, For activation function, The key-memory matrix is ​​used to store latent concepts (defined as model parameters before training and initialized randomly, and updated during training). C Indicates the number of potential concepts. D For embedded dimensions, For practicing embedding;

[0108] S202, retrieve the memorized knowledge according to the following formula. :

[0109] ,

[0110] In the above formula, For learners in time step t The value-memory matrix (defined as a model parameter before training and randomly initialized, and updated during training) is used to represent the learner at time step. t Key-memory matrix A matrix showing the level of mastery of each potential concept in the text;

[0111] S203, update the learner's time step according to the following formula. t+ The value of 1 - memory matrix :

[0112] ,

[0113] In the above formula, Update weights for attention. To acquire context-aware knowledge after the current interaction. Represents the vector product. and The vector product is a vector of shape. The matrix, For practice tags The relevance weight vector with all potential concepts. Add to the matrix. For learners in time step tThe expression for the calculation function of the value minus the memory matrix, and the attention update weights, is as follows:

[0114] ,

[0115] In the above formula, An embedding matrix representing the weights for skill-aware attention updates. S This represents the number of skill tags in the online learning platform; the expression for the calculation function of context-aware acquired knowledge after the current interaction is:

[0116] ,

[0117] In the above formula, For time steps t - i Context-aware practice embedding time steps t Contextual knowledge state Value embedding, For time steps t - i Self-attention weights For a fixed time step, and with the following:

[0118] ,

[0119] In the above formula, the middle It is a learnable decay rate parameter. It is a time step and Context-aware temporal distance metric between them Embedding time steps for context-aware practice of time step t t Contextual knowledge state Query embedding, Embedding time steps for context-aware practice of time step ti t Contextual knowledge state The key is embedded, and it has:

[0120] ,

[0121] ,

[0122] In the above formula, Indicates a time step. As an intermediate variable, Indicates time step i +1, Embedding time steps for context-aware practice of time step t t Contextual knowledge state Query embedding, For time steps Context-aware practice embedding time steps t Contextual knowledge state Key embedding.

[0123] Knowledge modeling also includes knowledge integration to obtain time steps. t final state of knowledge Learners need to model different knowledge states for different specific exercises at each time step to complete the exercise. If the exercise to be completed by the learner at the current time step is similar in type and required skills to the context exercises they have recently interacted with, then the learner only needs to use the context knowledge to complete the current exercise. If the exercise to be completed by the learner is significantly different in type and required skills from the exercises they have recently interacted with, then the learner needs to retrieve relevant knowledge memories from previous interactions during the exercise. Therefore, this section designs a context-aware attention mechanism (CAM), which determines the knowledge memories required to complete the current exercise by calculating the correlation between the current exercise and the set of context exercises. Specifically, in step S102 of this embodiment, the time step is obtained. t final state of knowledge The function expression is:

[0124] ,

[0125] In the above formula, For contextual knowledge state, To memorize knowledge, Add to the matrix. For practice tags Context-aware attention relevance weights between the learned content and its context. These are learnable relative parameters. Among them, the context-aware attention relevance weights... The expression for the computation function is:

[0126] ,

[0127] In the above formula, To apply the multi-head self-attention mechanism to the deadline time step Contextual Exercises The representation, For practicing embedding, With a fixed time step, and having , These represent the time steps. Practice embedding in context-aware mode Value embedding, for The self-attention weights for each time step are calculated as follows:

[0128] ,

[0129] In the above formula, For activation function, Indicates at time step Practice embedding in context-aware mode Query embedding, Indicates at time step Practice embedding in context-aware mode Key embedding, D For the embedded dimension.

[0130] In this embodiment, the prediction network used in step S103 is a three-layer response prediction network, and its function expression is:

[0131] ,

[0132] ,

[0133] ,

[0134] In the above formula, and It is an intermediate variable in a three-layer response prediction network; For the prediction result, { , , }and{ , , } represent the parameters of the transformation matrix and bias vector in the corresponding prediction network, respectively. This is a vector concatenation operation. For practicing embedding.

[0135] In this embodiment, the three-layer response prediction network and the skill-aware memory-enhanced recurrent neural network (SMRN) are trained end-to-end, and the loss function used during training has the following expression:

[0136] ,

[0137] In the above formula, Represents the loss function. In response, The results are the predictions from the three-layer response prediction network.

[0138] To verify the performance of the Memory Enhancement Attention Network (MAN) in this embodiment, extensive experiments were conducted on three public online learning benchmark datasets to validate the effectiveness of the proposed model in four aspects: 1) comparison of the overall predictive performance of MAN with other baseline models; 2) the effectiveness of the three key innovations in MAN (TPEN, SMRNN, and CAM); 3) the predictive performance of MAN in the face of skill switching phenomena; and 4) visualization of the skill-aware attention mechanism in SMRNN. The three public online learning benchmark datasets are shown in Table 1.

[0139] Table 1: Public Online Learning Benchmark Datasets

[0140]

[0141] The Statics2011, ASSISTments2009, and ASSISTments2017 datasets (basic statistics are shown in Table 1) were used to evaluate the performance of the proposed model and the baseline model. The area under the ROC curve (AUC) and prediction accuracy (ACC) were used to evaluate the predictive performance of all involved models. Generally, an AUC or ACC value of 0.5 indicates the performance of random guessing predictions, and a higher AUC or ACC value indicates better predictive performance.

[0142] As a comparison with the Memory-Enhanced Attention Network (MAN) in this embodiment, details of other existing baseline models are as follows: DKT is a simple RNN-based model that uses skill labels and response labels as input features to model the learner's knowledge state as a hidden state vector of the RNN. It is the earliest deep learning-based knowledge tracing model. DKVMN is a memory-enhanced neural network model that uses skill labels and response labels as input features to model the learner's knowledge state as two memory matrices. The key-memory matrix stores knowledge concepts, and the value-memory matrix stores the learner's mastery of related concepts. SAKT is the first knowledge tracing model based on a multi-head self-attention mechanism. It abandons the traditional approach of using RNNs to model learners' historical interactions and instead makes predictions by considering relevant exercises that learners have interacted with in the past. SAKT has been proven to be far more effective than RNN-based knowledge tracing models. EKT is an extension of the Exercise-Enhanced Recurrent Neural Network (EERNN) framework. Compared to EERNN, EKT further introduces information about the knowledge concepts present in each exercise. AKT is a model that relies entirely on a multi-head self-attention mechanism and is the foundational model for the method in this embodiment. It takes skill labels, practice labels, and response labels as input features and makes predictions by considering relevant practices that learners have interacted with in the past. SAINT+ is an extension of SAINT (the first Transformer-based knowledge tracing model), uniquely in that it introduces practice-response information separately, and embeds two temporal features, practice time and lag time, into the embedding of the response information. JKT is a deep knowledge tracing model based on a joint graph convolutional network. It adopts a novel reasoning-generative knowledge tracing framework, taking skill labels, practice labels, and response labels as input features, and attempts to learn the latent embeddings of practice and skills using graph neural networks.

[0143] The evaluation criteria for the Memory Enhancement Attention Network (MAN) in this embodiment and other existing baseline models are as follows:

[0144] I. Overall performance evaluation.

[0145] To validate the model proposed in this embodiment, this sub-experiment compared the overall performance of all models in predicting learner responses. Tables 2 and 3 list the overall results for AUC and ACC, with the best model performance highlighted in bold and the second-best model performance underlined.

[0146] Table 2: AUC performance of all knowledge tracing methods in predicting learner responses across all datasets.

[0147]

[0148] Table 3: ACC performance of all knowledge tracing methods in predicting learner responses across all datasets

[0149]

[0150] As can be seen from these two tables, the Memory-Enhanced Attention Network (MAN) in this embodiment outperforms (and sometimes significantly outperforms) all other baseline models on all datasets. On the Statics2011 dataset, MAN improves AUC and ACC by 11.36% and 10.46% respectively compared to the closest baseline model, and by 11.36% and 13.88% respectively compared to the base model AKT in this embodiment. On the ASSISTments2009 dataset, MAN improves AUC and ACC by 2.88% and 2.85% respectively compared to the closest baseline model (also the base model AKT). On the ASSISTments20117 dataset, MAN improves AUC and ACC by 23.48% and 21.57% respectively compared to the closest baseline model, and by 23.48% and 24.79% respectively compared to the base model AKT. In conclusion, these results demonstrate that the Memory-Enhanced Attention Network (MAN) in this embodiment has better performance in learner response prediction.

[0151] II. Ablation Research.

[0152] To gain a deeper understanding of the model proposed in this embodiment, this sub-experiment conducted an ablation study on the contributions of TPEN, SMRNN, and CAM proposed by MAN to the overall model performance. This ablation experiment primarily compared the performance metrics of MAN with the following model variants across all datasets: MAN-TPEN is a variant of MAN where the TPEN embedding network in MAN is replaced by RME; MAN-SMRNN is a variant of MAN where the SMRNN memory knowledge modeling network in MAN is replaced by DKVMN; MAN-CAM is a variant of MAN where the context-aware attention mechanism in MAN is replaced by a simple concatenation operation; and AKT+DKVMN is a concatenated network of AKT and DKVMN, where the model outputs of AKT and DKVMN are concatenated before the prediction layer and then input into the AKT prediction network to obtain the prediction result. The final results are shown in Table 4.

[0153] Table 4: Ablation Study Results of MAN on All Datasets

[0154]

[0155] Table 4 reports the model performance of MAN and its variants in learner response prediction, with the best model performance highlighted in bold and the second best underlined. From the table, we can observe that: 1) all ablation factors in MAN have a positive impact on the final performance of MAN; 2) compared to TPEN and SMRNN, CAM contributes the most to improving the final performance of the model; 3) for TPEN and SMRNN, their contributions to improving the final performance of MAN vary depending on the dataset. In summary, the effectiveness of the memory-enhanced attention network MAN in this embodiment stems from the synergistic effect of all key innovations. Different key innovations have varying degrees of positive impact on the overall model performance. CAM, with its dynamic integration of memory knowledge and contextual knowledge, can significantly improve the performance of the MAN model in predicting learner responses.

[0156] III. Performance Prediction under Skill Switching Phenomenon.

[0157] The key to the superior performance of the Memory Enhancement Attention Network (MAN) in this embodiment compared to other models lies in its ability to better model the learner's knowledge state during repeated skill switching. To verify this conclusion, this sub-experiment performed additional response predictions on the ASSISTments2017 dataset for exercises with different skill switching delays (SSLs). The experiment compared the response prediction performance of the proposed model with three other models: SMRNN, AKT, and DKVMN, for exercises with different SSLs. Experimental setup: First, the exercises in the test set were divided according to different SSLs: SSL = 0, SSL = 1, SSL = 2, SSL = 3, SSL = 4, SSL = 5, SSL = 6, SSL = 7, SSL = 8, SSL = 9, SSL = 1 ... SSL SSL SSL SSL SSL and SSL Then, experiments were conducted to test the model's performance on different practice groups. According to Definition 1, SSL=0 means that the skill at the corresponding time step appears for the first time in the interaction sequence, SSL=1 means that the skill at the corresponding time step has not been switched, SSL>1 means that the skill at the corresponding time step has been switched, and the larger the value, the more time steps have passed since the last appearance of the same skill. Figure 6 and Figure 7 The AUC and ACC results for practice response prediction under skill switching are reported separately. Figure 6 and Figure 7It can be observed that, overall, the Memory-Enhanced Attention Network (MAN) in this embodiment performs best, followed by SMRNN, and DKVMN performs worst. Specifically, for SMRNN and AKT, when SSL is in the range of 1 to 16, they produce similar results, but as SSL increases, the performance of SMRNN gradually improves; when SSL exceeds approximately 32, SMRNN significantly outperforms AKT; for SMRNN and DKVMN, except for the new skill (SSL=0), SMRNN significantly outperforms DKVMN, especially when SSL is greater than 32, the performance gap between them widens further; for all values ​​of SSL, MAN significantly outperforms other baseline models. Even more remarkably, when facing the new skill (SSL=0), MAN's performance still significantly outperforms other models. In summary, all the above evidence suggests that: 1) the Memory-Enhanced Attention Network (MAN) in this embodiment is insensitive to skill switching, while DKVMN, AKT, and SMRNN are more sensitive to skill switching phenomena; 2) compared to DKVMN, SMRNN can capture longer-term dependencies in the learning sequence; 3) the Memory-Enhanced Attention Network (MAN) in this embodiment has a good ability to predict learner performance by making full use of memory knowledge and contextual knowledge; 4) the Memory-Enhanced Attention Network (MAN) in this embodiment demonstrates superiority in modeling learner knowledge states in cold-start scenarios.

[0158] In summary, this embodiment, starting from the perspective of learner interaction sequences, first discusses the shortcomings of current mainstream DLKT models in modeling learner knowledge states. Then, it proposes a memory-enhanced attention network. This model not only possesses the memory-enhanced neural network's ability to store memorized knowledge and the attention-based model's ability to model contextual knowledge, but also automatically integrates relevant knowledge for response prediction based on the correlation between the current practice and the contextual practice. The proposed network consists of five parts: model embedding, contextual knowledge modeling, memorized knowledge modeling, knowledge integration, and response prediction. Considering the limited information capacity of one-dimensional scalar embeddings, this embodiment proposes a novel two-layer proportional embedding network, TPEN, which can fully utilize the representation of practice while avoiding over-parameterization and overfitting. Taking into account the differences in learnability of different skills (typically with different learning rates), this embodiment proposes a novel skill-aware memory-enhanced recurrent neural network, SMRNN, to store, retrieve, and update the learner's long-term memory knowledge during the learning process. To mitigate the adverse effects of skill switching, this embodiment designs a context-aware attention mechanism (CAM), which integrates knowledge based on the correlation between the current practice and the contextual practice, automatically weighting contextual and memory knowledge during response prediction. Finally, the effectiveness of the proposed network is validated through extensive experiments.

[0159] Furthermore, the present invention also provides a memory-enhanced attentional knowledge tracking system for skill switching phenomena, comprising a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute the aforementioned memory-enhanced attentional knowledge tracking method for skill switching phenomena.

[0160] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program, which is used to be programmed or configured by a microprocessor to perform the aforementioned memory-enhanced attention-based knowledge tracking method for skill switching phenomena.

[0161] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0162] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A memory-enhanced attention knowledge tracking method for skill switching phenomenon, characterized by, The method comprises the following steps: S101, obtain a time step t of practice labels , skill labels and response labels , wherein the response labels refer to the true feedback labels of the answers submitted by the learners and knowledge embeddings ​ S102, obtaining an exercise embedding and knowledge embedding performing attention knowledge tracking to obtain a context knowledge state ; meanwhile, using a skill-aware memory recurrent neural network SMRNN to extract memory knowledge according to the input exercise embedding ; integrating the context knowledge state , the memory knowledge to obtain a final knowledge state t of a time step ;​ S103, time step t final state of knowledge and practice embedding The concatenated vectors are input into a pre-trained prediction network to obtain the learner's correct response to the practice labels. Skill tags Prediction results ; The function expression for the practice embedding obtained in step S101 is: f(x) = x + 1 ， in the above formula, is an activation function, is a raw embedding of a skill tag , is an influence parameter, is a raw embedding of an exercise tag , D is an embedding dimension, is an exercise tag, is a skill tag, is a response tag, and has: ， ， ， In the above formula, and respectively represent the skill and practice one-hot encoding vectors; and respectively are and embedding matrices of is an embedding vector of , is the total number of practice labels, S represents the number of skill labels in the online learning platform, D is the embedding dimension; The function expression of the knowledge embedding obtained in step S101 is: f (x) = x + 1 ， in the above formula, is an activation function, denotes a raw embedding of a skill-response pair, is an influence parameter, denotes a variable vector of raw embeddings of skill-response pairs, D is an embedding dimension, is an exercise label, is a skill label, is a response label, and has: ， ， In the above formula, and respectively represent the multi-hot encoding vectors of skill-response and exercise-response; and respectively are and embedding matrices, the multi-hot encoding vector of skill-response is obtained by splicing the one-hot encoding of skill label and the one-bit binary encoding 0 or 1 of response label ; the multi-hot encoding vector of exercise-response is obtained by splicing the one-hot encoding of exercise label and the one-bit binary encoding 0 or 1 of response label ; In step S102, the skill-aware memory-augmented recurrent neural network SMRNN is used to generate the skill embedding according to the inputted practice Extracting memory knowledge Comprising: S201, embedding according to the practice Compute practice tags Relevance weight vector to all potential concepts : ， In the above formula, is an activation function, is a key-memory matrix for storing latent concepts, C denotes the number of latent concepts, D is an embedding dimension, is a practice embedding; S202, retrieve the memory knowledge according to the following formula : ， In the above formula, a value-memory matrix for characterizing the learner's mastery of each potential concept in the key-memory matrix t at time step t a value-memory matrix for characterizing the learner's mastery of each potential concept in the key-memory matrix at time step S203, update the learner at time step t according to the following formula t+ 1 value - memory matrix : ， In the above equation, is the attention update weight, is the retrieved memory knowledge, denotes the vector product, and the vector product of and is a matrix of shape , is the practice label is the relevance weight vector of all potential concepts, is the matrix addition, is the value-memory matrix of the learner at time step t , and has the computational function expression of the attention update weight: ， wherein, an embedding matrix representing skill-aware attention update weights, S denotes the number of skill tags in the online learning platform.

2. The skill-oriented switching phenomenon-aware memory-augmented attention knowledge tracking method of claim 1, wherein, The time step obtained in step S102 t The final knowledge state The function expression is: ， In the above formula, is a contextual knowledge state, is a memory knowledge, is a matrix addition, is an exercise tag is a contextual awareness attention relevance weight between its contextual learning content, is a learnable relativity parameter.

3. The skill-oriented switching phenomenon-aware memory-augmented attention knowledge tracking method of claim 2, wherein, Context-aware attention relevance weight The computational function expression for the context-aware attention relevance weight is ， In the above formula, is the representation of the context exercise up to time step , is the exercise embedding, is the fixed time step, and , , , , , is the self-attention weight for time step , computed as follows: ， In the above formula, is an activation function, denotes the query embedding at time step context-aware practice embedding is performed the query embedding, denotes the key embedding at time step context-aware practice embedding is performed the key embedding, D is the embedding dimension.

4. The skill-oriented switching phenomenon-aware memory-augmented attention knowledge tracing method of claim 1, wherein, The prediction network used in step S103 is a three-layer response prediction network, and the function expression thereof is as follows: ， ， ， In the above formula, and are intermediate variables of the three-layer response prediction network; is the prediction result, , , } and , , } represent the parameters of the transformation matrix and the bias vector in the corresponding prediction network respectively, is a vector splicing operation, is a practice embedding.

5. The skill-switching phenomenon-oriented memory-enhanced attention knowledge tracking method according to claim 4, characterized by, The three-layer response prediction network and the skill-aware memory-augmented recurrent neural network SMRN are trained in an end-to-end manner, and the function expression of the loss function used in the training is as follows: ， In the above formula, represents a loss function, is a response, is the prediction result of the three-layer response prediction network.

6. A memory-enhanced attention knowledge tracking system oriented to skill switching phenomenon, comprising a microprocessor and a memory connected to each other, characterized in that, The microprocessor is programmed or configured to perform the skill-switching phenomenon-oriented memory-augmented attention knowledge tracking method according to any one of claims 1-5.

7. A computer-readable storage medium having stored therein a computer program, characterized in that, The computer program is used for programming or configuring the microprocessor to perform the skill-switching phenomenon-oriented memory-augmented attention knowledge tracking method according to any one of claims 1-5.

Images