Multi-Task Attention Knowledge Tracing Method and System for Online Learning Platform

Abstract

The present invention discloses a multi-task attention knowledge tracking method and system for an online learning platform. The present invention includes extracting the unbalanced perception attention weights of the current time step based on the historical interaction sequence according to the exercise label \(e_t\), the skill label \(s_t\), and encoding to obtain the exercise embedding \(x_t\); encoding to obtain the knowledge embedding \(y_t\) based on the response label \(r_t\) and the true learning action \(a_t\); performing attention knowledge tracking on \(x_t\) and \(y_t\) to obtain the knowledge state \(h_t\); and performing multi-task prediction based on \(x_t\) and \(h_t\) to obtain the corresponding prediction results. The present invention solves the problem of extremely unbalanced exercise records encountered in the model training process by using an unbalanced perception attention mechanism to assign personalized weights to exercises with different record counts, and realizes the soft fusion of skill personalization through two stages, having excellent knowledge tracking performance.

Description

Technical Field

[0001] The present invention relates to the technical field of online learning (e-learning) platforms, and particularly to a multi-task attention knowledge tracking method and system for an online learning platform. Background Art

[0002] With the popularization of e-learning, learners can acquire knowledge through self-study without leaving home. To develop personalized learning plans for learners, online learning platforms need to accurately understand and master the knowledge levels of learners at various stages. Knowledge tracking (KT) is an important task in online learning, and its goal is to model the knowledge state (KS) of learners by analyzing their learning history data, that is, the degree of mastery of learners over skills (or concepts / knowledge components). On an online learning platform, learners can learn relevant skills by completing specific exercises (for example, if "addition" is a skill, then "1 + 1" is its exercise), and the platform tracks the knowledge state KS of learners regarding the learned skills based on a knowledge tracking model. Finally, the platform determines whether learners have mastered these skills through a when-to-stop strategy.

[0003] In recent years, Deep Learning-based Knowledge Tracing (DLKT) methods have demonstrated superior performance compared to traditional models, such as Bayesian Knowledge Tracing, Latent Factor Models, and Item Response Theory, due to their stronger representation capabilities. For example, Deep Knowledge Tracing (DKT), Dynamic Key-Value Memory Networks (DKVMN), and Self-Attentive Knowledge Tracing (SAKT). To enhance the ability of DLKT to predict future learner responses, we have explored multiple learning-related factors (such as performance, forgetting, practice text, etc.) and integrated them into specific models, such as the Practice-Aware Knowledge Tracing framework, Attentive Knowledge Tracing (AKT), and Relation-aware self-attention for Knowledge Tracing (RKT). Due to the sparsity issue of practice data, early DLKT models used skills instead of practice as model inputs to avoid overparameterization of the models. To further improve the performance of DLKT, multiple learning-related factors (such as practice, performance, forgetting, and practice text, etc.) have been explored and integrated into specific models. However, existing knowledge tracing models typically only utilize the one-size-fits-all (correct or incorrect) response feedback after submitting practice answers, making it impossible for these models to more accurately capture the knowledge growth after each interaction.

[0004] Learning records are a passive reflection of a learner's knowledge level. In contrast, learner feedback provides an active understanding of the learner's own knowledge state, which in turn provides a direct and truthful indicator of the learner's learning situation. However, there are few knowledge tracing models that utilize training data related to learner feedback, despite its important role in correcting knowledge tracing results. Wang et al. pointed out that feedback plays a positive role in learning, as it can facilitate transfer and memory in practice-based learning. Intuitively, precise feedback can help the model better capture the knowledge growth of learners after each learning interaction and more accurately distinguish the difficulty levels of practices, which are the keys to the excellent performance of knowledge tracing models.

[0005] In an interactive online learning platform, learners attempt some exercises without even looking at other exercises. Each learner can attempt an exercise multiple times and can learn between consecutive attempts. When learners are unable to solve an exercise after multiple attempts, they usually seek help. A common learning aid in many interactive online learning platforms is the option to obtain hints during the interaction. ] . By abstracting and simplifying learning activities, this paper notes two types of behavioral feedback that can more intuitively reflect the difficulty of exercises: the number of hint uses and the number of attempts. Figure 1 Combined with the analysis of online learning data, a visual description of these two types of feedback is given, where subfigure (a) is the simplified learning interaction process, and subfigure (b) shows part of the exercise records extracted from real data. According to Figure 1 in subfigure (a), the general online learning interaction process can be summarized into the following three situations: Situation 1 (①): In one interaction process, regardless of whether the answer response to the exercise is correct or not, only one attempt is made and then the interaction stops. This situation can be verified from the records of exercise labels: 733 and 752 in the right subfigure of Figure 1 . In this case, regardless of whether the learner's answer response to the exercise is correct or incorrect, it only indicates the approximate matching degree between their knowledge state and the difficulty of the exercise. Situation 2 (①…①): In one interaction process, the first attempt is wrong, and then multiple attempts are made until it is correct, or it is always wrong and then finally abandoned. This situation can be verified from the records of exercise labels: 821 and 826 in the right subfigure of Figure 1 . In this case, the more times the learner attempts, it may mean that the exercise is more difficult for them. Situation 3 (①②①…②①): In one interaction process, during the multiple attempts in Situation 2, one or more hint use actions are accompanied (the whole hint usually consists of one or more steps). This situation can be verified from the records of exercise labels: 824, 822, and 825 in the right subfigure of Figure 1 . In this case, the more times the learner attempts and uses hints, it may mean that the exercise is more difficult for them.

[0006] In summary, the attempt behavioral feedback and the hint use behavioral feedback can capture the relationship between the learner's knowledge state and the difficulty of the exercise better than the response feedback. Figure 2Shows the distribution of the average correct rate (ACR) of exercises in the ASSISTments2017 dataset with respect to the number of hint uses and the number of attempts (black line chart), where the number of exercises (#exercise) corresponding to each ACR value is referenced by the red line chart. It can be observed from the figure that as the number of attempts and the number of hint uses increase, the ACR metric of the exercises gradually decreases (ignoring the noise caused by small samples), which undoubtedly provides statistical support for the motivation of this article. Although Chaudhry et al. attempted to jointly learn hint use (whether hint is used or not, regardless of the number) prediction (related tasks) and response prediction (main task) through multi-task learning (MTL), the performance improvement of their model in learner response prediction is limited. There are three reasons: 1) The feedback of related tasks is not fine enough; 2) The feature fusion between the main task and related tasks is relatively simple and direct; 3) The problem of unbalanced exercise samples caused by the sparsity of exercise data is also one of the objective challenges restricting the excellent performance of knowledge tracing models. How to apply knowledge tracing models to online learning platforms to more accurately capture the knowledge growth after each interaction has become a key technical problem to be solved urgently. Summary of the Invention

[0007] The technical problem to be solved by the present invention: Aiming at the above problems of the prior art, a multi-task attention knowledge tracing method and system for an online learning platform are provided. The present invention can realize multi-task attention knowledge tracing on the online learning platform, solve the problem of extremely unbalanced exercise records encountered in the model training process by assigning personalized weights to exercises with different record counts through an imbalance-aware attention mechanism, and is based on the soft fusion of skill personalization realized in two stages of multi-task feature fusion and multi-model feature fusion, and has excellent knowledge tracing performance.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0009] A multi-task attention knowledge tracing method for an online learning platform, including:

[0010] S101, extracting the imbalance-aware attention weight at the current time step t according to the historical interaction sequence

[0011] S102, encoding to obtain an exercise embedding x based on the exercise label e t , skill label s t at the current time step t and the imbalance-aware attention weight ; based on the imbalance-aware attention weight t and the response label r at the current time step t​t , real learning action a t Encoding to get knowledge embedding y t ;

[0012] S103, practice embedding x t , knowledge embedding t Perform attention knowledge tracking to obtain the knowledge state h t ;

[0013] S104, embed x according to the training at time step t t , knowledge state h t Perform multi-task prediction to obtain corresponding prediction results, wherein the multi-task prediction includes a main task and at least one related task, wherein the main task is used to predict the learner's correct response practice label e t The predicted probability The related task is used to predict the learner's training label e t The number of uses or attempts when submitting an answer.

[0014] Optionally, step S101 includes:

[0015] S201, from the historical interaction sequence Count the number of prior samples #e for each training label e;

[0016] S202, discretize #e using the log2(#e+1) function;

[0017] S203, setting the number of prior samples after discretization processing as the imbalance factor of the training label e;

[0018] S204: Use one-hot encoding to encode the imbalance factor of each training label e to obtain the imbalance factor vector o # ;

[0019] S205, use the maximum imbalance factor value among all training labels e to adjust the dimension D of the imbalance-aware embedding vector I iam Initialize and calculate the imbalance-aware attention weight of the current time step t according to the following formula

[0020]

[0021] In the above formula, the superscript T represents the transpose operation of the imbalance-aware embedding vector I, and the number of the imbalance-aware embedding vector I will be automatically learned through gradient backpropagation during the training process.

[0022] Optionally, the training embedding x is obtained by encoding in step S102 t The function expression is:

[0023]

[0024] In the above formula, Softmax represents the Softmax activation function, is the original embedding of the skill label s t , is the imbalance-aware attention weight at the current time step t, is the practice label e t 's scalar difficulty parameter, is the vector composed of all practices under the skill label s t , is 's corresponding bias vector, D represents the dimension of the embedding, and there is:

[0025]

[0026]

[0027]

[0028] In the above formula, and respectively represent the one-hot encoding vectors of skills and practices; and are respectively and 's embedding matrices; is 's embedding vector, E represents the total number of practices, S represents the number of skill labels in the online learning platform, and D is the embedding dimension.

[0029] Optionally, the function expression of the knowledge embedding y t encoded in step S102 is:

[0030]

[0031] In the above formula, Softmax represents the Softmax activation function, is the dual-task embedding of skill-response-action, is the imbalance-aware attention weight at the current time step t, is the practice label e t 's scalar difficulty parameter, is the variable vector of skill-response-action, is 's corresponding bias vector, D represents the dimension of the embedding, and there is:

[0032]

[0033]

[0034]

[0035]

[0036] In the above formula, is the original embedding of the skill-response of the main task, is the original embedding of the skill-action of the related task, is the skill-response variable embedding vector of the main task, is the skill-action variable embedding vector of the related task, and respectively represent the skill perception attention weights of the main task and the related task, o s is the one-hot vector of the skill label s, and are the weight vectors for obtaining the feature fusion of the main task and the related task, respectively, and there are:

[0037]

[0038]

[0039]

[0040]

[0041] In the above formula, and respectively represent the multi-hot encoding vectors of skill-response and skill-action; and are respectively and embedding matrices; and are respectively and embedding matrices, S represents the number of skill labels in the online learning platform, D is the embedding dimension, the multi-hot encoding vector of skill-response is obtained by concatenating the skill one-hot encoding and the one-bit binary encoding 0 or 1 of the response; the multi-hot encoding vector of skill-action is obtained by concatenating the skill one-hot encoding and the binary encoding of the action, L is the number of bits of the binary encoding of the action, which is the number of bits of the binary encoding corresponding to the largest action label. For those with fewer bits than L, 0 is padded in front of the binary encoding corresponding to the action label.

[0042] Optionally, in step S104, according to the practice embedding x at time step t t and the knowledge state h tWhen performing multi-task prediction to obtain corresponding prediction results, the multi-task prediction performed includes a main task and a related task, and the main task predicts and obtains the correct response practice label e of the learner through a trained three-layer response prediction network t prediction probability wherein, and the functional expression of the three-layer response prediction network is:

[0043]

[0044]

[0045]

[0046] In the above formula, ReLU and Linear are activation functions, and are intermediate variables of the three-layer response prediction network, and are respectively the transformation matrix and bias vector parameters of each layer in the three-layer response prediction network, || represents the concatenation operation of vectors; the related task predicts and obtains the number of times or attempts used by the learner when submitting answers for the practice label e t and the functional expression of the three-layer action prediction network is:

[0047]

[0048]

[0049]

[0050] In the above formula, ReLU and Linear are activation functions, and are intermediate variables of the three-layer action prediction network, and are respectively the transformation matrix and bias vector parameters of each layer in the three-layer response action network, || represents the concatenation operation of vectors.

[0051] Optionally, the three-layer response prediction network and the three-layer action prediction network together constitute a dual-task attention knowledge tracking model BAKT, and the loss function adopted during the training of the dual-task attention knowledge tracking model BAKT is:

[0052]

[0053] In the above formula, is the loss function, is the cross-entropy loss of the main task, is the cross-entropy loss of the related task, σ rThe uncertainty parameter for the main task, σ a is the uncertainty parameter for the related tasks.

[0054] Optionally, when performing multitask prediction based on the practice embedding x t at time step t, the knowledge state h t to obtain the corresponding prediction results, the multitask prediction performed includes a main task and two related tasks. The two related tasks are respectively used to obtain the number of times or attempts made by the learner when submitting an answer for the practice label e t , and the main task predicts the prediction probability of the learner correctly responding to the practice label e t through a trained three-layer response prediction network. And the functional expression of the three-layer response prediction network is:

[0055]

[0056]

[0057]

[0058] In the above formula, ReLU and Linear are activation functions, and are intermediate variables of the three-layer response prediction network, and are respectively the transformation matrix and bias vector parameters of each layer in the three-layer response prediction network. || represents the vector concatenation operation. and are respectively the knowledge states of the two related tasks, and are respectively the weights of the knowledge states of the two related tasks; the related tasks predict the number of times or attempts made by the learner for the practice label e t when submitting an answer through a trained three-layer action prediction network. And the functional expression of the three-layer action prediction network is:

[0059]

[0060]

[0061]

[0062] In the above formula, ReLU and Linear are activation functions, and are intermediate variables of the three-layer action prediction network, and They are the transformation matrix and bias vector parameters of each layer in the three-layer response action network respectively, and || represents the concatenation operation of vectors. and are the knowledge states of two related tasks respectively. and are the weights of the knowledge states of two related tasks respectively, and the knowledge states of the two related tasks and are generated as follows: based on the practice label e t , skill label s t and the imbalance-aware attention weight to encode the practice embedding x; based on the imbalance-aware attention weight and the response label r t , true learning action a t at the current time step t, use the knowledge encoders corresponding to the two related tasks to encode y 1 and y 2 respectively to obtain the knowledge embeddings and Perform attention knowledge tracking on the practice embedding x with the knowledge embeddings and respectively, so as to obtain the knowledge encoders corresponding to the two related tasks for y 1 and y 2 respectively.

[0063] Optionally, the three-task attention knowledge tracking model TAKT formed by the three-layer response prediction network and the three-layer action prediction network, and the loss function used by the three-task attention knowledge tracking model TAKT during training is:

[0064]

[0065] In the above formula, is the loss function, is the cross-entropy loss of the main task, is the cross-entropy loss of the first related task, is the cross-entropy loss of the second related task, σ r is the uncertainty parameter of the main task, is the uncertainty parameter of the first related task, is the uncertainty parameter of the second related task.

[0066] In addition, the present invention also provides a multi-task attention knowledge tracking system for an online learning platform, including a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the multi-task attention knowledge tracking method for the online learning platform.

[0067] In addition, the present invention also provides a computer-readable storage medium storing a computer program for programming or configuring a microprocessor to execute the multi-task attention knowledge tracking method for an online learning platform.

[0068] Compared with the prior art, the present invention mainly has the following advantages:

[0069] 1. The present invention can perform multi-task prediction based on the exercise embedding x t and knowledge state h t at time step t to obtain corresponding prediction results. The multi-task prediction includes a main task and at least one related task. For the first time, two related tasks, i.e., hint prediction and attempt prediction, are combined with the original knowledge tracking task, and multi-task attention knowledge tracking is proposed.

[0070] 2. The present invention includes extracting the imbalance-aware attention weight at the current time step t according to the historical interaction sequence and encoding to obtain the exercise embedding x based on the exercise label e t , skill label s t and the imbalance-aware attention weight at the current time step t; encoding to obtain the knowledge embedding y t based on the imbalance-aware attention weight , the response label r of the current time step t t and the true learning action a t . In this way, a novel imbalance-aware attention mechanism is realized, and the extreme exercise record imbalance problem encountered in the model training process is solved by automatically assigning personalized weights to exercises with different record counts. t

[0071] 3. The present invention realizes a skill-aware attention mechanism based on skill-related attention weights, and realizes soft fusion of skill personalization in two stages of multi-task feature fusion and multi-model feature fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 Describes the attempt and hint behaviors during the exercise in the prior art.

[0073] Figure 2 Describes the attempt and hint behaviors during the exercise in the prior art.

[0074] Figure 3 Is a schematic diagram of the basic process of the method in Embodiment 1 of the present invention.

[0075] Figure 4 Is a structural diagram of the dual-task attention knowledge tracking (BAKT) model in Embodiment 1 of the present invention.

[0076] Figure 5 This is a visualization diagram of the formation process of the unbalanced perception attention mechanism in Embodiment 1 of the present invention.

[0077] Figure 6 This is the distribution of the average hint usage times and average attempt times statistically obtained from the ASSISTments2009 dataset in Embodiment 1 of the present invention in terms of skills.

[0078] Figure 7 This is the distribution of the average hint usage times and average attempt times statistically obtained from the ASSISTments2017 dataset in Embodiment 1 of the present invention in terms of skills.

[0079] Figure 8 This is the structural diagram of the three-task attention knowledge tracking (TAKT) model in Embodiment 2 of the present invention.

[0080] Figure 9 This is the schematic diagram of the feedback label merging of related tasks in the embodiments of the present invention. Detailed implementation manners

[0081] The overall idea of the multi-task attention knowledge tracking method (abbreviated as the MAKT method) for an online learning platform of the present invention is as follows: First, two fine-grained attention mechanisms are introduced to re-encode the exercise embedding and knowledge embedding at the input layer of the model; then, an attention knowledge tracking network is used to model the current knowledge state of the learner; finally, different task prediction networks are designed to predict the feedback labels of the main task and related tasks. In order to more accurately predict the responses of the learner, two specific prediction models are proposed in the MAKT method in this paper, namely the dual-task attention knowledge tracking (BAKT) model and the three-task attention knowledge tracking (TAKT) model, and two independent embodiments are used for detailed description.

[0082] Embodiment 1:

[0083] As Figure 3 and Figure 4 shown, the multi-task attention knowledge tracking method for an online learning platform in this embodiment includes:

[0084] S101. Extract the unbalanced perception attention weight at the current time step t according to the historical interaction sequence

[0085] S102. Encode to obtain the exercise embedding x based on the exercise label e t t 、skill label s t <000,00053> and the unbalanced perception attention weightt ; based on unbalanced perception attention weights and the response label r at the current time step t t , the true learning action a t to encode the knowledge embedding y t ;

[0086] S103, perform attention knowledge tracking on the practice embedding x t , the knowledge embedding y t to obtain the knowledge state h t ;

[0087] S104, according to the practice embedding x at time step t t , the knowledge state h t perform multi-task prediction to obtain the corresponding prediction results, where the multi-task prediction includes a main task and at least one related task, and the main task is used to predict the prediction probability of the learner's correct response practice label e t The related task is used to predict the number of times or attempts used by the learner when submitting an answer for the practice label e The related task is used to predict the number of times or attempts used by the learner for the practice label e t when submitting an answer.

[0088] In this embodiment, the historical interaction sequence can be expressed as:

[0089]

[0090] where ι1 to ι t are t quadruples of the historical interaction sequence, and any i-th quadruple can be expressed as:

[0091]

[0092] This quadruple ι i contains: the practice label e at time step i i , the optional skill label s i (Some online learning platforms may not involve skill labels), the main task response feedback r i (correct or incorrect) and the set of learning actions of the related task where are the 1st to Nth learning actions at time step i respectively, and N is the number of related tasks.

[0093] The dual task in the Dual-task Attention Knowledge Tracking (BAKT) model refers to the joint learning of a learning action prediction from a related task (hint usage prediction or attempt number prediction) and a response prediction. As Figure 4As shown, the entire dual-task attention knowledge tracing model consists of three main modules: a dual-task embedding module (the black dashed area), an attention knowledge tracing module, and a dual-task prediction module. Among them, the dual-task embedding, as a key module for realizing dual-task joint learning, also plays an important role in the TAKT model. The dual-task embedding module involves three specific components, including an Imbalance-aware Exercise Encoder (IEE), a Bi-task Feature Encoder (BFE), and an Imbalance-aware Knowledge Encoder (IKE). The Imbalance-aware Exercise Encoder IEE is used to generate the exercise embeddings of BAKT; the Bi-task Feature Encoder BFE is used to jointly encode the original knowledge embeddings from two different tasks; the Imbalance-aware Knowledge Encoder IKE is used to encode the output from BFE to generate the knowledge embeddings of BAKT.

[0094] In BAKT, the Imbalance-aware Exercise Encoder IEE and the Imbalance-aware Knowledge Encoder IKE are respectively used to generate exercise embeddings and dual-task knowledge embeddings where D represents the dimension of these embeddings. In fact, these two embeddings are extensions of the Rasch Model-based Embeddings (RME). Although RME achieves an appropriate balance between modeling individual exercise differences and avoiding over-parameterization of the model, it ignores the problem of exercise sample imbalance caused by the sparsity of exercise data. Since the current DLKT models use the same weights to fit all exercise labels during the training process, the further release of the model's prediction potential is restricted.

[0095] In most real-world online learning platforms, the sample sizes of different exercises are extremely unbalanced because most exercises only have interaction records with a small number of learners. The statistics of the exercise sample distribution on the three datasets of Statics2011, ASSISTments2009, and ASSISTments2017 support this claim. For the exercise labels with large samples, the training efficiency is low due to the inability to provide useful learning signals, but they may dominate the entire training and lead to a degenerate model; for the exercise labels with small samples, they can be effectively trained because they can provide useful learning signals, but they may cause the model to be over-parameterized. To address this issue, this paper implements an imbalance-aware attention mechanism (IAM) in step S101, aiming to automatically reduce the label weights of small-sample exercises and large-sample exercises during model training. Specifically, step S101 of this embodiment includes:

[0096] S201, count the prior sample number #e of each exercise label e from the historical interaction sequence ;

[0097] S202, discretize #e using the log2(#e + 1) function;

[0098] S203, set the discretized prior sample number as the imbalance factor of the exercise label e;

[0099] S204, encode the imbalance factor of each exercise label e using one-hot encoding to obtain the imbalance factor vector o # ;

[0100] S205, initialize the dimension D iam of the imbalance-aware embedding vector I with the maximum imbalance factor value among all exercise labels e, and calculate the imbalance-aware attention weight at the current time step t according to the following formula

[0101]

[0102] In the above formula, the superscript T represents the transpose operation of the imbalance-aware embedding vector I, and the number of the imbalance-aware embedding vector I will be automatically learned through gradient backpropagation during training. The visualization formation process of the above process is as Figure 5 shown, where is an imbalance-aware embedding vector (IEV), which is used to generate the imbalance-aware exercise attention weight according to the prior sample number of the exercise.

[0103] Given the imbalance-aware attention weight at the current time step t The unbalanced perception exercise encoder IEE can be used to encode the exercise label e at the current time step t t , the skill label s t and the unbalanced perception attention weight to obtain the exercise embedding x t ; The unbalanced perception knowledge encoder IKE is used to encode the knowledge embedding y based on the unbalanced perception attention weight and the response label r at the current time step t t , the true learning action a t ; In this embodiment, the functional expression of the exercise embedding x encoded in step S102 t is: t

[0104]

[0105] In the above formula, Softmax represents the Softmax activation function, is the original embedding of the skill label s t , is the unbalanced perception attention weight at the current time step t, is the scalar difficulty parameter of the exercise label e t , is the vector composed of all exercises under the skill label s t , is the corresponding bias vector, D represents the dimension of the embedding, and there are:

[0106]

[0107]

[0108]

[0109] In the above formula, and respectively represent the one-hot encoding vectors of skills and exercises; and are respectively and the embedding matrices; is the embedding vector of, E represents the total number of exercises, S represents the number of skill labels in the online learning platform, and D is the embedding dimension. In this embodiment, the functional expression of the knowledge embedding y encoded in step S102 t is:

[0110]

[0111] ​In the above formula, Softmax represents the Softmax activation function, is the dual-task embedding of skill-response-action, is the imbalance-aware attention weight at the current time step t, is the practice label e t 's scalar difficulty parameter, is the variable vector of skill-response-action, is the corresponding bias vector, D represents the dimension of the embedding, and there are:

[0112]

[0113]

[0114]

[0115]

[0116] In the above formula, is the original embedding of skill-response of the main task, is the original embedding of skill-action of the related task, is the variable embedding vector of skill-response of the main task, is the variable embedding vector of skill-action of the related task, and respectively represent the skill-aware attention weights of the main task and the related task, o s is the one-hot vector of the skill label s, and are the weight vectors for obtaining the feature fusion of the main task and the related task, respectively, and there are:

[0117]

[0118]

[0119] <\

[0120]

[0121] In the above formula, and respectively represent the multi-hot encoding vectors of skill-response and skill-action; and are respectively and 's embedding matrices; and are respectively and The embedding matrix, where S represents the number of skill tags in the online learning platform, D is the embedding dimension. The multi-hot encoded vector of skill-response is obtained by concatenating the one-hot encoding of the skill and the one-bit binary encoding 0 or 1 of the response. The multi-hot encoded vector of skill-action is obtained by concatenating the one-hot encoding of the skill and the binary encoding of the action. L is the number of bits of the binary encoding of the action, which is the number of bits of the binary encoding corresponding to the largest action tag. For those with fewer bits than L, 0s are padded in front of the binary encoding corresponding to the action tag.

[0122] In BAKT, the role of the dual-task feature encoding component (BFE) is to encode the information from the main task and the related task based on a skill-aware attention mechanism (which will be explained in detail below). The encoder takes the original embedding of skill-response (from the main task) and the original embedding of skill-action (from the related task) as inputs, and the output is the dual-task knowledge embedding of skill-response-action Taking the skill-response variable embedding vector (from the main task) and the skill-action variable embedding vector (from the related task) as inputs, the output is the dual-task variable embedding vector of skill-response-action One of the simplest ways to fuse the original embedding (or vector) of the main task and the original embedding (or vector) of the related task is to directly add them element-wise. Another common advanced method is to perform a linear weighted sum of the embeddings (or vectors) of each individual task element-wise. This means that for any specific skill, each task has a fixed proportion of contribution to the final dual-task embedding. However, for the two related tasks that will be jointly learned in this paper, this does not seem to be a reasonable solution. To verify this claim, this paper statistically analyzed the average hint usage count (AHC) and the average attempt count (AAC) of different skills in the ASSISTments2009 and ASSISTments2017 datasets, and the results are as Figure 6 and Figure 7 shown. It can be observed from the figure that for different skills, the average hint usage count and the average attempt count are different; the ratio of AHC to AAC is also different. This result shows that it is unreasonable for all skills to share the same dual-task fusion weights. The entire fusion process should reflect the personalized differences of skills, that is, use skill-related weights to weight the features of different tasks. Therefore, this paper designs a skill-aware attention mechanism: To obtain the weights for multi-task feature fusion related to skills, SAM defines two skill-aware embedding vectors and Used to obtain the weights for feature fusion of the main task and related tasks respectively. The parameters of the vector will be automatically learned during training, and S is the number of different skill labels. By respectively performing dot product operations on these two embedding vectors with the "one-hot" vector of skill label s the skill-aware attention weights of the main task and related tasks can be obtained. Given the skill label s t at the current time step, first calculate two skill-aware attention weights and respectively from the main task and related tasks. Then use these two weights to encode the skill-response-action dual-task embedding and variable vector.

[0123] In this embodiment, step S103 performs attention knowledge tracing on the practice embedding x t and knowledge embedding y t to obtain the knowledge state h t which is implemented using the context-aware knowledge modeling network AKT (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.). Additionally, other existing attention knowledge tracing models can also be used as needed.

[0124] During the process of modeling the learner's knowledge state by the context-aware knowledge modeling network AKT, it not only considers the correlation between the context practice and the current practice, but also considers the forgetting mechanism in the learning process. Compared with other similar models, the context-aware knowledge modeling network AKT is a more reasonable context-aware knowledge modeling network. In this embodiment, in step S102, attention knowledge tracing is performed on the practice embedding x t and knowledge embedding y t to obtain the context knowledge state which utilizes that the context-aware knowledge modeling network AKT involves three components: practice encoder EE, knowledge encoder KE, and knowledge retriever KR.

[0125] The practice encoder EE takes the sequence {1,…, t} of practice embeddings x1~x t from time steps 1 to t as input and uses a monotonic attention mechanism to output a sequence of context-aware practice embeddings and the context-aware practice embedding at time step t The computational function expression is:

[0126]

[0127] In the above formula, is the context-aware exercise embedding of time step t-i into time step t, the value embedding of α (t-i), is the self-attention weight at time step t-i, τ is a fixed time step length, and there is:

[0128]

[0129] In the above formula, is the context-aware exercise embedding of time step t into time step t, the query embedding of is the context-aware exercise embedding of time step t-i into time step t, the key embedding of

[0130] The knowledge encoder KE takes the sequence of exercise embeddings y1 to y t from time steps 1 to t, {1,…, t} as input and uses the same monotonic attention mechanism to output a sequence of context-aware knowledge embeddings and the context-aware knowledge embedding at any time step t The computational function expression is:

[0131]

[0132] In the above formula, is the context-aware knowledge exercise embedding of time step t-i into time step t, the value embedding of α (t-i), is the self-attention weight at time step t-i, τ is a fixed time step length, and there is: <�

[0133]

[0134] In the above formula, is the context-aware knowledge exercise embedding of time step t into time step t, the query embedding of is the context-aware knowledge exercise embedding of time step t-i into time step t, the key embedding of

[0135] The knowledge retriever KR takes the sequence of context-aware exercise embeddings and the knowledge embedding sequence As inputs, use a new monotonic attention mechanism with an exponential decay term to output the retrieved context knowledge state for the current exercise The basic form of the decay term is an exponential decay curve over time, which may peak at the corresponding time step when past exercises are highly similar to the current exercise. The specific form is as follows:

[0136]

[0137] In the above formula, is the context knowledge state of the context-aware exercise at time step t - i embedded in time step t Value embedding of, α (t-i), is the self-attention weight at time step t - i, τ is a fixed time step, and there is:

[0138]

[0139] In the above formula, θ > 0 is a learnable decay rate parameter, d(t, i) is the context-aware time distance metric between time steps t and i, is the context knowledge state of the context-aware exercise at time step t embedded in time step t Query embedding of, is the context knowledge state of the context-aware exercise at time step t - i embedded in time step t Key embedding of, and there is:

[0140]

[0141]

[0142] In the above formula, i represents the time step, γ t,τ′ is an intermediate variable, τ′ represents time step i + 1, is the context knowledge state of the context-aware exercise at time step t embedded in time step t Query embedding of, is the context knowledge state of the context-aware exercise at time step τ′ embedded in time step t Key embedding of.

[0143] In this embodiment, the dual-task prediction component uses the knowledge state (h t ) retrieved from the attention knowledge tracking component and the current exercise embedding (x t ) as inputs to output the learner's response and action prediction for the current exercise. Specifically, in step S104, according to the exercise embedding x at time step tt Knowledge state h t When performing multi-task prediction to obtain corresponding prediction results, the multi-task prediction performed includes a main task and a related task, and the main task predicts and obtains the correct response practice label e of the learner through a trained three-layer response prediction network t Prediction probability And the functional expression of the three-layer response prediction network is as follows:

[0144]

[0145]

[0146]

[0147] In the above formula, ReLU and Linear are activation functions, And Are intermediate variables of the three-layer response prediction network, And Are respectively the transformation matrix and bias vector parameters of each layer in the three-layer response prediction network, || represents the concatenation operation of vectors; for the prediction of related tasks, considering that the two related tasks involved in this article are multi-label prediction tasks, BAKT replaces the ReLU function in the main task prediction network with the Softmax function to output the action prediction for the related tasks Specifically, the related task predicts and obtains the number of times or attempts used by the learner for the practice label e t When submitting an answer, and the functional expression of the three-layer action prediction network is as follows:

[0148]

[0149]

[0150]

[0151] In the above formula, ReLU and Linear are activation functions, And Are intermediate variables of the three-layer action prediction network, And Are respectively the transformation matrix and bias vector parameters of each layer in the three-layer response action network, || represents the concatenation operation of vectors.

[0152] Multi-task learning (MTL) involves the optimization problem for multiple target models, which is common in many deep learning problems. Due to the large difference in the loss scales of different tasks, the overall loss of the model may be dominated by a single task. To avoid the problem that the losses of other tasks cannot properly affect the learning process of the network sharing layer, MAKT adopts a method of using homoscedastic uncertainty to combine multi-task loss functions, which was first proposed in the field of computer vision, to simultaneously learn multiple targets, namely UWLoss, that is, using homoscedastic uncertainty to weight the losses of different tasks. In this embodiment, a dual-task attention knowledge tracking model BAKT (one each) composed of a three-layer response prediction network and a three-layer action prediction network is used. The loss function adopted by the dual-task attention knowledge tracking model BAKT during training is:

[0153]

[0154] In the above formula, is the loss function, is the cross-entropy loss of the main task, is the cross-entropy loss of the related task, and σ r is the uncertainty parameter of the main task, and σ a is the uncertainty parameter of the related task.

[0155] In addition, this embodiment also provides a multi-task attention knowledge tracking system for an online learning platform, including a microprocessor and a memory connected to each other. The microprocessor is programmed or configured to execute the foregoing multi-task attention knowledge tracking method for an online learning platform.

[0156] In addition, this embodiment also provides a computer-readable storage medium. A computer program is stored in the computer-readable storage medium, and the computer program is used to be programmed or configured by a microprocessor to execute the foregoing multi-task attention knowledge tracking method for an online learning platform.

[0157] Embodiment 2:

[0158] This embodiment is basically the same as Embodiment 1, and the main difference is the implementation method of step S104. In this embodiment, a three-task attention knowledge tracking model TAKT is composed of a three-layer response prediction network and a three-layer action prediction network. In order to enable the main task to jointly learn with two related tasks simultaneously, this embodiment proposes a three-task knowledge tracking model TAKT based on a skill-aware attention mechanism, dual-task embedding (from BAKT), and an encoder and a retriever (from the attention knowledge tracking component). The original motivation for jointly learning three tasks was to encode the features of all three tasks in the dual-task feature encoder of the BAKT model, but the experimental results showed that this scheme was not ideal, and in some cases, the performance was even worse than that of the dual-task model. In order to obtain better learner response prediction performance, TAKT first forms a dual-task model based on the main task and one related task respectively; then uses the skill-aware attention mechanism to integrate the knowledge states from the two dual-task models, Figure 8 shows the structure of the three-task attention knowledge tracking model, where two dual-task embedding components from BAKT achieve joint learning by sharing a series of parameters, including embedding parameters, imbalance-aware embedding vectors, and imbalance-aware practice embedding parameters. First, TAKT uses dual-task embedding to encode the original embeddings of skills, skill-responses, skill-actions 1 / 2 , practice, and imbalance-aware factors in the main task and one of the related tasks respectively, generating two practice embeddings (x 1 and x 2 ) and two dual-task knowledge embeddings (y 1 and y 2 ), where due to parameter sharing, x 1 =x 2 =x. Then, TAKT uses the practice encoder to encode x, generating a context-aware practice embedding uses the knowledge encoder to encode y 1 and y 2 respectively, generating context-aware knowledge embeddings ( and ). Then, use the knowledge retriever in AKT to encode and to produce the knowledge states of the main task and the first related task encode and to generate the knowledge states of the main task and the second related task Finally, apply the skill-aware attention mechanism SAM to and Perform weighting, and input the weighted knowledge state together with the practice embedding x into the corresponding prediction network to predict the response of the main task and two related task rows. The prediction networks of the main task and the related tasks are similar in structure to those in the BAKT model.

[0159] In this embodiment, in step S104, according to the practice embedding x at time step t t , knowledge state h t When performing multi-task prediction to obtain the corresponding prediction results, the multi-task prediction includes the main task and two related tasks. The two related tasks are respectively used to obtain the number of times or attempts of the learner when submitting answers for the practice label e t , and the main task predicts the prediction probability of the learner correctly responding to the practice label e t through a trained three-layer response prediction network , and the functional expression of the three-layer response prediction network is:

[0160]

[0161]

[0162]

[0163] In the above formula, ReLU and Linear are activation functions, and are intermediate variables of the three-layer response prediction network, and are respectively the transformation matrix and bias vector parameters of each layer in the three-layer response prediction network. || represents the concatenation operation of vectors, and are respectively the knowledge states of the two related tasks, and are respectively the weights of the knowledge states of the two related tasks; the related tasks predict the number of times or attempts of the learner when submitting answers for the practice label e t through a trained three-layer action prediction network, and the functional expression of the three-layer action prediction network is:

[0164]

[0165]

[0166]

[0167] In the above formula, ReLU and Linear are activation functions, and are intermediate variables of the three-layer action prediction network, and are the transformation matrix and bias vector parameters of each layer in the three - layer response action network respectively, and || represents the concatenation operation of vectors. and are the knowledge states of two related tasks respectively. and are the weights of the knowledge states of two related tasks respectively. As Figure 8 shown, the generation of the knowledge states and of two related tasks includes: encoding the practice label e t at the current time step t, the skill label s t and the imbalance - aware attention weight to obtain the practice embedding x; based on the imbalance - aware attention weight and the response label r t at the current time step t, the true learning action a t , respectively using the knowledge encoders corresponding to two related tasks to encode y 1 and y 2 to obtain the knowledge embeddings and Performing attention - based knowledge tracking on the practice embedding x with the knowledge embeddings and respectively, so as to obtain the knowledge encoders corresponding to two related tasks for y 1 and y 2 respectively.

[0168] In this embodiment, the loss function adopted by the three - task attention - based knowledge tracking model TAKT during training is:

[0169]

[0170] In the above formula, is the loss function, is the cross - entropy loss of the main task, is the cross - entropy loss of the first related task, is the cross - entropy loss of the second related task, σ r is the uncertainty parameter of the main task, is the uncertainty parameter of the first related task, is the uncertainty parameter of the second related task.

[0171] This embodiment has conducted extensive experiments on three public online learning benchmark datasets (the three datasets of Statics2011, ASSISTments2009, and ASSISTments2017) with the aim of verifying the effectiveness of the proposed BAKT and TAKT models in learner response prediction.

[0172] Three datasets, Statics2011, ASSISTments2009, and ASSISTments2017 (basic statistical information is shown in Table 1), were used to evaluate the performance of the proposed model and the baseline models. In the Statics2011 dataset, the number of attempts made by the learner in each interaction is recorded in each learning record; in the ASSISTments2009 and ASSISTments2017 datasets, not only the number of attempts made by the learner in each interaction but also their use of hints is recorded in each learning record. Two metrics, the area under the ROC curve (AUC) and the prediction accuracy (ACC), were used to evaluate the prediction performance of all the models involved. Generally, an AUC or ACC value of 0.5 represents the performance prediction result of random guessing, and the larger the AUC or ACC value, the better the prediction performance.

[0173] Table 1: Dataset statistical information.

[0174]

[0175] Different from the responses (correct or incorrect) of the main task, the feedback of the related tasks varies greatly. To avoid over-parameterization, the feedback from the related tasks: the number of attempts (#attempt) and the number of hint uses (#hint) need to be discretized and compressed using the function log2(* + 1), and then the compressed values are rounded to form the final related task prediction labels l attempt and l hint . Figure 9 Shows the preprocessing scheme for the related task feedback in all datasets, where the number of exercises (#exercise) corresponding to each average correct rate (ACR) value refers to the red dashed line. As Figure 9 shown in subfigure (a) of attempt , for the Statics2011 dataset, the value range of l attempt is from 1 to 7, and those l attempt greater than 7 are merged into one label, i.e., l Figure 9 = 7, because their ACR values tend to 0 when ignoring small sample noise. As attempt shown in subfigures (b) and (c) of hint , for the ASSISTments2009 dataset, the range of l attempt is from 0 to 2, l hint is 0 or 1, and those l Figure 9 greater than 2 are merged, and those l attemptranges from 1 to 7, l hint ranges from 0 to 6 and does not involve any merging operations.

[0176] To verify the effectiveness of the MAKT method, this paper compares the dual-task attention knowledge tracing (BAKT) model and the triple-task attention knowledge tracing (TAKT) model proposed in this embodiment with existing state-of-the-art DLKT methods. The existing state-of-the-art DLKT methods are as follows: DKT is a simple RNN-based model that takes skill labels and response labels as input features and models the learner's knowledge state as the hidden state vector of the RNN. It is the earliest deep learning-based knowledge tracing model. DKVMN is a memory-augmented neural network-based model that takes skill labels and response labels as input features and models the learner's knowledge state as two memory matrices, where the key-memory matrix stores knowledge concepts and the value-memory matrix stores the learner's mastery of related concepts. Colearn is an improved version of DKVMN for multi-task learning that integrates the task of predicting the number of hint uses on the basis of response prediction. It is the only multi-task knowledge tracing model currently. SAKT is the first knowledge tracing model based on the multi-head self-attention mechanism. It abandons the traditional method of using RNN to model the learner's historical interactions and instead makes predictions by considering relevant exercises that the learner has interacted with in the past. SAKT has been verified to be far more effective than RNN-based knowledge tracing models. AKT is a model that fully relies on the multi-head self-attention mechanism and is also the basic model of this paper. It takes skill labels, exercise labels, and response labels as input features and makes predictions by considering relevant exercises that the learner has interacted with in the past. JKT is a deep knowledge tracing based on a joint graph convolutional network. It adopts a new inference generation knowledge tracing framework that takes skill labels, exercise labels, and response labels as input features and attempts to learn the latent embeddings of exercises and skills using a graph neural network. BPS+AKT is an extended version of AKT with a scalable embedding framework BPS. It improves the RME method in AKT by simultaneously modeling skills, exercises, performance factors, and side factors. Bi-CLKT is a knowledge tracing model based on dual-graph contrastive learning that obtains discriminative representations of exercises using node-level contrastive learning and obtains discriminative representations of skills using graph-level contrastive learning. The relevant evaluations are as follows:

[0177] I. Overall performance evaluation.

[0178] Since this paper aims to use MTL to improve the performance of knowledge tracing methods in predicting learners' responses, the AUC metrics of BAKT_Hint, BAKT_Attempt, and TAKT are compared with those of all baseline models on all three datasets. The mean values of five-fold cross-validation are reported in Table 2, where the best model is in bold, the best baseline model is underlined, and "-" indicates the corresponding statistical item is missing in Statics2011 due to the absence of hint feedback.

[0179] Table 2 Overall performance evaluation results.

[0180]

[0181] As can be seen from Table 2, TAKT outperforms all other DLKT methods on all datasets; BAKT_Hint and BAKT_Attempt perform better than all baseline models on the Statics2011 and ASSISTments2017 datasets, and only BPS+AKT performs slightly better than BAKT_Hint and BAKT_Attempt on the ASSISTments2009 dataset. Compared with the closest baseline model, TAKT increases the AUC value by 4.61% and 7.66% on the ASSISTments2009 and ASSISTments2017 datasets respectively, by 13.93% and 19.72% compared with the basic model AKT, and by 16.13% and 25.87% compared with the MTL-based knowledge tracing model Colearn. BAKT_Hint increases the AUC value of AKT by 7.28% and 18.79% on the ASSISTments2009 and ASSISTments2017 datasets respectively. More importantly, BAKT_Hint increases the AUC value by 9.36% and 24.88% compared with Colearn on the ASSISTments2009 and ASSISTments2017 datasets with hint feedback. Although BAKT_Attempt does not outperform the baseline model BPS+AKT on the ASSISTments2009 dataset, compared with the closest baseline model, it increases the AUC value by 6.59% and 7.59% on the Statics2011 and ASSISTments2017 datasets respectively. BAKT_Attempt also increases the AUC value of the basic model AKT by 11.51%, 6.78%, and 19.57% on the three datasets of Statics2011, ASSISTments2009, and ASSISTments2017 respectively. All in all, the above results indicate that BAKT and TAKT are advanced in predicting learners' responses.

[0182] II. Ablation Study.

[0183] To gain insights into how the Imbalance-Aware Attention Mechanism (IAM) and Skill-Aware Attention Mechanism (SAM) proposed by MAKT affect the final results of specific MAKT models, further ablation experiments were conducted in this section. The AUC and ACC metrics of TAKT, BAKT_Hint, and BAKT_Attempt were compared with their following variants respectively: -I means ablation of the Imbalance-Aware Attention Mechanism on the basis of a specific MAKT model. -S means ablation of the Skill-Aware Attention Mechanism on the basis of a specific MAKT model. -IS means ablation of both IAM and SAM on the basis of a specific MAKT model. Since the dataset Statics2011 does not involve the feedback on learners' hint usage, ablation experiments on TAKT and BAKT_Hint models were only conducted on the two datasets ASSISTments2009 and ASSISTments2017 in this section, but ablation experiments on BAKT_Attempt model were conducted on all datasets. Tables 3 to 5 show the ablation experiment results of TAKT, BAKT_Hint, and BAKT_Attempt respectively.

[0184] Table 3: Ablation Experiment Results of TAKT.

[0185]

[0186] Table 4: Ablation Experiment Results of BAKT_Hint

[0187]

[0188] As can be observed from Tables 3 and 4: TAKT and BAKT_Hint perform better than all their ablation variants, and the variants that ablate any one component {TAKT-I / S, BAKT_Hint-I / S} perform better than the variants that ablate two components {TAKT-IS, BAKT_Hint-IS} on ASSISTments2009 and ASSISTments2017. As can be observed from Table 5: BAKT_Attempt performs better than all its ablation variants, and the variant that ablates any one component BAKT_Attempt-I / S performs better than the variant that ablates two components BAKT_Attempt-IS on all datasets. In addition, for different datasets and different models, the performance contributions of IAM and SAM vary. Generally speaking, the above results can illustrate that: 1) The imbalance-aware attention mechanism can help the BAKT and TAKT models improve the performance of learner response prediction; 2) The skill-aware attention mechanism can help the BAKT and TAKT models improve the performance of learner response prediction; 3) The simultaneous action of these two fine-grained attention mechanisms can also help the BAKT and TAKT models improve the performance of learner response prediction.

[0189] In summary, the method of this embodiment jointly learns the learner response prediction (main task) with two more refined related tasks (hint usage behavior prediction and attempt behavior prediction) by introducing two fine-grained attention mechanisms (imbalance-aware attention mechanism and skill-aware attention mechanism), aiming to improve the prediction performance on the main task. More specifically, the proposed multi-task attention knowledge tracing method makes two major improvements to the existing DLKT method: 1) Utilize the imbalance-aware attention mechanism to overcome the adverse effects caused by the problem of unbalanced practice samples; 2) Utilize the skill-aware attention mechanism to achieve the skill-personalized fusion of multi-task features and multi-model features, avoid the rigid fusion of multi-task features, and ensure that different skills can obtain adaptive proportions of features from different tasks. The experimental results on three real-world benchmark online learning datasets show that: The method of this embodiment is generally superior to the existing state-of-the-art DLKT methods in terms of learner response prediction performance.

[0190] In addition, this embodiment also provides a multi-task attention knowledge tracing system for an online learning platform, including a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the foregoing multi-task attention knowledge tracing method for an online learning platform.

[0191] In addition, this embodiment further provides a computer-readable storage medium, in which a computer program is stored, and the computer program is used to be programmed or configured by a microprocessor to execute the foregoing multi-task attention knowledge tracking method for an online learning platform.

[0192] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented 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. The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks. These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured product including an instruction device, and the instruction device implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0193] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and refinements made without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A multi-task attention knowledge tracking method for an online learning platform, characterized in that including: S101, Extract the imbalance-aware attention weights of the current time step according to the historical interaction sequence ; t Extract the imbalance-aware attention weights of the current time step ; S102, based on the current time step t of the practice label , skill label and the unbalanced perception attention weight to encode and obtain the practice embedding ; Based on unbalanced perception attention weights and the current time step t response label , true learning action Encode to obtain knowledge embedding ; S103, perform exercise embedding , knowledge embedding , and perform attention-based knowledge tracking to obtain the knowledge state ; S104, according to the time step t of the practice embedding and the knowledge state to perform multi-task prediction to obtain corresponding prediction results. The multi-task prediction includes a main task and at least one related task. The main task is used to predict the prediction probability of the learner correctly responding to the practice label , and the related task is used to predict the number of times or attempts when the learner submits an answer for the practice label ; ​ The function expression for encoding to obtain the practice embedding in step S102 is as follows: ， In the above formula, Softmax represents the Softmax activation function, is the original embedding of the skill label , is the current time step t of the imbalance-aware attention weight, is the practice label of the scalar difficulty parameter, is the skill label of all the practices under which the vector formed, is the corresponding bias vector, D represents the dimension of the embedding, and there is: ， ， ， In the above formula, and respectively represent the one-hot encoded vectors of skills and exercises; and are respectively and embedding matrices; is embedding vector, E represents the total number of exercises, S represents the number of skill tags in the online learning platform, and D is the embedding dimension; The knowledge embedding obtained by encoding in step S102 has the following function expression: ， In the above formula, Softmax represents the Softmax activation function, is the dual-task embedding of skill-response-action, is the current time step t of the imbalance-aware attention weight, is the practice label of the scalar difficulty parameter, is the variable vector of skill-response-action, is the corresponding bias vector, D represents the dimension of the embedding, and there is: ， ， ， ， In the above formula, is the original embedding of the skill-response of the main task, is the original embedding of the skill-action of the related task, is the skill-response variable embedding vector of the main task, is the skill-action variable embedding vector of the related task, and respectively represent the skill perception attention weights of the main task and the related task, is the skill label s is the one-hot vector of and are the weight vectors for obtaining the feature fusion of the main task and the related task, respectively, and there is: ， ， ， ， In the above formula, and respectively represent the multi-hot encoding vectors of skill-response and skill-action; and are respectively and embedding matrices; and are respectively and embedding matrices, S represents the number of skill labels in the online learning platform, D is the embedding dimension, the multi-hot encoding vector of skill-response is obtained by concatenating the one-hot encoding of the skill and the one-bit binary encoding 0 or 1 of the response; the multi-hot encoding vector of skill-action is obtained by concatenating the one-hot encoding of the skill and the binary encoding of the action, L is the number of bits of the binary encoding of the action, which is the number of bits of the binary encoding corresponding to the largest action label. For those with the number of bits less than L, 0 is padded in front of the binary encoding corresponding to the action label.

2. The multi-task attention knowledge tracing method for an online learning platform according to claim 1, wherein Step S101 includes: S201, count the prior sample number of each exercise label from the historical interaction sequences e ;# e ; S202, use function to discretize #e; S203, set the number of prior samples after discretization as the imbalance factor of the practice label e ; S204, use one-hot encoding to encode the imbalance factor of each exercise label e to obtain an imbalance factor vector ; S205, using all exercise tags e Initialize the dimension of the imbalance-aware embedding vector I with the largest imbalance factor value among all D iam and calculate the imbalance-aware attention weights at the current time step according to the following formula t as follows : ， In the above formula, the superscript T represents the transpose operation of the imbalance-aware embedding vector I, and the number of the imbalance-aware embedding vector I will be automatically learned through gradient backpropagation during the training process.

3. The multi-task attention knowledge tracing method for an online learning platform according to claim 1, wherein In step S104, according to the time step t of the practice embedding , the knowledge state When performing multi-task prediction to obtain the corresponding prediction result, the multi-task prediction performed includes a main task and a related task, and the main task predicts the correct response practice label of the learner through a trained three-layer response prediction network of the prediction probability of, and the functional expression of the three-layer response prediction network is: ， ， ， In the above formula, and are activation functions, and are intermediate variables of the three-layer response prediction network, { , , } and { , , } are the transformation matrix and bias vector parameters of each layer in the three-layer response prediction network respectively, represents the concatenation operation of vectors; the relevant tasks are predicted by a trained three-layer action prediction network to obtain the number of times or attempts used by the learner when submitting answers for the exercise label , and the functional expression of the three-layer action prediction network is: ， ， ， In the above formula, and are activation functions, and are intermediate variables of the three-layer action prediction network. { , , } and { , , } are the transformation matrix and bias vector parameters of each layer in the three-layer response action network respectively. represents the concatenation operation of vectors.

4. The multi-task attention knowledge tracing method for an online learning platform according to claim 3, wherein The dual-task attention knowledge tracing model BAKT composed of the three-layer response prediction network and the three-layer action prediction network, and the loss function adopted by the dual-task attention knowledge tracing model BAKT during training is: ， In the above formula, is the loss function, is the cross-entropy loss of the main task, is the cross-entropy loss of the related task, is the uncertainty parameter of the main task, is the uncertainty parameter of the related task.

5. The multi-task attention knowledge tracing method for an online learning platform according to claim 1, wherein, In step S104, according to the time step t of the practice embedding , the knowledge state When performing multi-task prediction to obtain corresponding prediction results, the multi-task prediction performed includes a main task and two related tasks. The two related tasks are respectively used to obtain the number of times or attempts used by the learner when submitting answers for the practice label . And the main task predicts the prediction probability of the learner correctly responding to the practice label through a trained three-layer response prediction network. And the functional expression of the three-layer response prediction network is: ， ， ， In the above formula, and are activation functions, and are intermediate variables of the three-layer response prediction network, { , , } and { , , } are the transformation matrix and bias vector parameters of each layer in the three-layer response prediction network respectively, represents the concatenation operation of vectors, and are the knowledge states of two related tasks respectively, and are the weights of the knowledge states of two related tasks respectively; the related tasks are predicted by a trained three-layer action prediction network to obtain the number of times or attempts used by the learner when submitting an answer for the practice label , and the functional expression of the three-layer action prediction network is: ， ， ， In the above formula, and is the activation function, and is the intermediate variable of the three-layer action prediction network, { , , }and{ , , } are the transformation matrix and bias vector parameters of each layer in the three-layer response action network, Represents the vector concatenation operation, and are the knowledge states of two related tasks, and are the weights of the knowledge states of the two related tasks, and the knowledge states of the two related tasks and The generation includes: based on the current time step t Practice Tags , Skill Tags and imbalance-aware attention weights Coding gets practice embedded ; Based on imbalance-aware attention weight and the current time step t Responsive Tags , real learning movements Two knowledge encoders corresponding to two related tasks are used and Encode to get knowledge embedding and ; For practice embedding Knowledge Embedding and Perform attention knowledge tracking to obtain the knowledge encoder pairs corresponding to the two related tasks and .

6. The multi-task attention knowledge tracking method for an online learning platform according to claim 5, characterized in that, The triple-task attention knowledge tracing model TAKT composed of the three-layer response prediction network and the three-layer action prediction network, and the loss function adopted by the triple-task attention knowledge tracing model TAKT during training is: ， In the above formula, is the loss function, is the cross-entropy loss of the main task, is the cross-entropy loss of the first related task, is the cross-entropy loss of the second related task, is the uncertainty parameter of the main task, is the uncertainty parameter of the first related task, is the uncertainty parameter of the second related task.

7. A multi-task attention knowledge tracing system for an online learning platform, comprising a microprocessor and a memory connected to each other, characterized in that, The microprocessor is programmed or configured to execute the multi-task attention knowledge tracing method for the online learning platform according to any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, characterized in that, The computer program is used to be programmed or configured by the microprocessor to execute the multi-task attention knowledge tracing method for the online learning platform according to any one of claims 1 to 6.

Images