Fuzzy neural network-based interpretable knowledge tracking method

Abstract

The invention discloses an interpretable knowledge tracking method based on a fuzzy neural network. The method comprises the steps of collecting input of input information; performing fuzzification processing on continuous scores of the learner on the knowledge concepts; based on an attention mechanism, dynamically screening out information of historical related exercises beneficial to cognitive state updating; determining a fuzzy cognition state according to the fuzzy cognition state of the learner at the previous moment and the current exercise answering performance, and realizing updating of the fuzzy cognition state, local fuzzy reasoning and global fuzzy cognition state summarization through three types of preset fuzzy rules; and splicing the updated and determined global fuzzy cognitive state with the to-be-predicted exercise features, and outputting an answer prediction result through a preset function. According to the method, the interpretability of fuzzy logic and the characterization capability of deep learning are fully fused, the change process of the cognitive state of the learner can be effectively described, the model is made to have interpretability, and powerful support is provided for practical application of the knowledge tracking method.

Description

Technical Field

[0001] This invention relates to the technical fields of knowledge tracing technology and related fields, specifically to an interpretable knowledge tracing method based on fuzzy neural networks. Background Technology

[0002] Knowledge tracing, as a core technology of intelligent education systems, aims to dynamically model and predict learners' mastery of knowledge concepts by analyzing their historical answer records, thereby enabling personalized learning path recommendations and instructional interventions. In recent years, researchers have improved prediction accuracy by incorporating more question information and more accurately modeling changes in learners' cognitive states to better reflect their learning progress. However, these methods often focus only on objective questions while neglecting subjective ones. Therefore, traditional knowledge tracing models are typically limited to objective questions and difficult to apply to subjective ones. Objective questions, however, present challenges in determining whether a learner's answer reflects a correct guess when they haven't mastered the concept or a mistake when they have, introducing a degree of randomness. Compared to objective questions, subjective questions more easily assess learners' mastery of knowledge concepts and more accurately model their cognitive states. However, the openness of their answers and the ambiguity of their scoring criteria make them difficult for existing models to handle effectively.

[0003] Although existing deep learning-based knowledge tracing methods have made significant progress in predictive performance, the inherent black-box nature of deep learning models means that while they can accurately predict learners' future performance, they cannot explain the logic behind the predictions. This lack of transparency within the model makes it difficult for educators to understand and trust the model's decision-making process, thus reducing the model's practical value.

[0004] In summary, to address the dual shortcomings of existing knowledge tracing methods—limited to objective questions in application and lacking interpretability in model mechanisms—an interpretable knowledge tracing method based on fuzzy neural networks is proposed. Summary of the Invention

[0005] The main objective of this invention is to provide an interpretable knowledge tracing method based on fuzzy neural networks, which addresses the dual shortcomings of existing knowledge tracing methods, namely, their limitation to objective questions in application and their lack of interpretability in model mechanisms. This method introduces fuzzy neural networks to handle the uncertainty of subjective questions and uses fuzzy rules to reason about cognitive states, thereby more accurately modeling learners' cognitive states and further improving the accuracy and interpretability of knowledge tracing.

[0006] Based on the first main aspect of the present invention, an interpretable knowledge tracing method based on a fuzzy neural network is provided, comprising the following steps:

[0007] The input information is collected and entered into the computer system; the input information includes the learner's historical practice sequence, current practice, and practice to be predicted; the historical practice sequence includes each historical practice and its corresponding knowledge concept, as well as the continuous scores of the knowledge concept; the current practice includes its corresponding response performance;

[0008] The computer system executes a computer program to perform fuzzification processing on the learner's continuous scores on knowledge concepts, mapping them into multiple fuzzy scores, and describing the membership degree of the fuzzy scores using a Gaussian membership function;

[0009] Based on the attention mechanism, each historical practice in the historical practice sequence is assigned a relevance weight to the current practice, and historical practice information that is conducive to updating the cognitive state is dynamically selected according to the relevance weight.

[0010] The learner's current fuzzy cognitive state is determined based on their previous fuzzy cognitive state and their performance in the current exercise. The fuzzy cognitive state is updated, local fuzzy reasoning is performed, and the global fuzzy cognitive state is summarized through three types of preset fuzzy rules.

[0011] The computer system executes a computer program to concatenate the updated and determined global fuzzy cognitive state with the exercise to be predicted, and obtains the learner's predicted response to the exercise through a linear function or a sigmoid function.

[0012] As a further preferred embodiment, in the aforementioned method, the step of describing the membership degree of fuzzy fractions using a Gaussian membership function is performed using the following formula:

[0013] Specifically, the formula is:

[0014]

[0015] Where j represents the total number of fuzzy score sets. This represents the continuous score for knowledge concept k at time t. express The membership degree belongs to fuzzy fractions. and This represents the mean and standard deviation of the fuzzy scores.

[0016] As a further preferred embodiment, in the aforementioned method, the step of assigning a relevance weight to each historical practice in the historical practice sequence based on the attention mechanism includes:

[0017] When the learner answers the current exercise, information related to the current exercise is extracted from the historical exercise sequence, and its relevance is calculated through the attention mechanism;

[0018] Specifically, the current exercise is represented as a query vector, and each exercise in the historical exercise sequence is represented as a key vector and a value vector, respectively;

[0019] Calculate the similarity between the query vector and all key vectors, and normalize the similarity into weights using the Softmax function;

[0020] Finally, the weights are used to sum all the value vectors to obtain a weighted aggregate representation of the historical practice information.

[0021] As a further preferred embodiment, in the aforementioned method, the step of determining its current fuzzy cognitive state includes:

[0022] The number of combinations of the fuzzy cognitive state set and the fuzzy score set of the previous moment and the current practice response corresponds to the number of fuzzy rules; the output of the fuzzy rule node is shown in the following formula:

[0023] Specifically:

[0024]

[0025] Where i represents the total number of fuzzy cognitive state sets, and j represents the total number of fuzzy score sets.

[0026] The calculation of the fuzzy cognitive state is shown in the following formula:

[0027]

[0028] in, For adjustable weights, .

[0029] As a further preferred option, in the aforementioned method, the execution steps for updating the fuzzy cognitive state using the first type of fuzzy rules are as follows:

[0030] When the knowledge concept corresponding to the exercise at a certain moment is the same as the knowledge concept currently being focused on, the update condition of the knowledge concept state is triggered.

[0031] Specifically, the update condition is to update the fuzzy cognitive state of the previous moment using the calculation formula of the fuzzy cognitive state;

[0032] If the cognitive concept being practiced is different from the cognitive concept being focused on, then the state of fuzzy cognition remains unchanged from the previous moment.

[0033] The execution steps for implementing local fuzzy inference using the second type of fuzzy rules are as follows:

[0034] The activation intensity of the fuzzy rule nodes is determined by the probability that the learner's fuzzy cognitive state of the knowledge concept in the previous moment belongs to the corresponding fuzzy set, and the probability that the learner's answer performance in the current moment belongs to the corresponding fuzzy set.

[0035] Based on the activation intensity, the interaction between the previous moment's ambiguous cognitive state and the current moment's response performance is transmitted to all possible levels of mastery through a weighted method.

[0036] The execution steps for summarizing the global fuzzy cognitive state using the third type of fuzzy rule are as follows:

[0037] Integrate all the fuzzy rule nodes corresponding to the activated second type of fuzzy rules to obtain the learner's fuzzy cognitive state at the current moment;

[0038] The degree of influence of the fuzzy cognitive state can be obtained by a certain fuzzy rule node and the influence weight of the fuzzy rule node on the fuzzy cognitive state. The probability of the fuzzy cognitive state is obtained by summing the contributions of all fuzzy rule nodes.

[0039] The summation result unifies the local reasoning results into the overall cognitive state, completing the update from local conditions to a global fuzzy cognitive state.

[0040] As a further preferred embodiment, in the aforementioned method, the steps for obtaining the learner's predicted response to the prediction exercise are as follows:

[0041]

[0042] in, This represents the connection between the learner's global cognitive state and practice characteristics. Represents the weight matrix and bias;

[0043] Specifically, when the prediction practice It's a subjective question. For the linear function; when the exercise to be predicted It's an objective question. This refers to the Sigmoid function.

[0044] Based on a second key aspect of the present invention, an interpretable knowledge tracing system based on a fuzzy neural network is provided, comprising:

[0045] The information input module is used to input the learned learners' historical practice sequences, current practice, and practice to be predicted as raw data into the system;

[0046] The fuzzification module is used to map learners' continuous scores on knowledge concepts into a fuzzy distribution with multiple intervals;

[0047] The related practice module is used to filter relevant information from historical practice through an attention mechanism;

[0048] The fuzzy reasoning module is used to update the learner's fuzzy cognitive state based on preset fuzzy rule units;

[0049] The prediction module combines the global fuzzy cognitive state with the features of the exercise to be predicted and outputs the prediction result of the exercise to be predicted.

[0050] As a further preferred embodiment, in the aforementioned system, the fuzzy inference module includes:

[0051] The first type of fuzzy rule unit is used to control the conditions for updating fuzzy cognitive states;

[0052] The second type of fuzzy rule unit is used to establish a local reasoning relationship between fuzzy cognitive state and response performance;

[0053] The third type of fuzzy rule unit is used to summarize the fuzzy cognitive state from local to global.

[0054] Based on a third key aspect of the present invention, an electronic device is provided, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0055] The memory stores a computer program, which, when executed by the processor, causes the processor to perform the aforementioned interpretable knowledge tracing method based on fuzzy neural networks.

[0056] Based on a fourth key aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon that, when executed, implements the aforementioned interpretable knowledge tracing method based on a fuzzy neural network.

[0057] Compared with existing technologies, this invention provides an interpretable knowledge tracing method based on fuzzy neural networks. First, this invention combines fuzzy logic with neural networks, overcoming the limitations of traditional knowledge models in handling subjective questions by designing fuzzy processing and fuzzy reasoning. Specifically, fuzzy processing can uniformly map learners' responses to both subjective and objective questions to a fuzzy distribution, thereby preserving some positive and rich evaluation information from subjective questions. Simultaneously, dynamic updates of the fuzzy cognitive state based on predefined fuzzy rules make the model's decision-making process a transparent and traceable reasoning process, enhancing the model's inherent interpretability.

[0058] Secondly, this invention utilizes an attention mechanism to dynamically filter the most relevant information from the learner's historical practice sequence. This not only enables the model to more accurately capture key factors influencing the learner's current ambiguous cognitive state and improve prediction accuracy, but also makes the model's explanation of the ambiguous cognitive state more targeted and persuasive.

[0059] Finally, this invention expands the application scope of knowledge tracing technology by unifying the evaluation methods for subject and object questions, making it applicable to more complex and realistic educational scenarios. The method of this invention can handle scoring uncertainty while capturing structural and sequential information, and possesses interpretable reasoning capabilities. Experimental results show that this method achieves a good balance between prediction accuracy and interpretive transparency, effectively alleviating the trust problem caused by the lack of interpretability in deep knowledge tracing models. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.

[0061] Figure 1 The following is a flowchart illustrating the execution of an interpretable knowledge tracing method based on a fuzzy neural network in one embodiment of the present invention.

[0062] Figure 2 The diagram illustrates the response at prediction time step 7 of an interpretable knowledge tracing method based on a fuzzy neural network in one embodiment of the present invention.

[0063] Figure 3 The diagram illustrates the response at prediction time step 8 of an interpretable knowledge tracing method based on a fuzzy neural network in one embodiment of the present invention.

[0064] Figure 4 The diagram illustrates the response at prediction time step 9 of an interpretable knowledge tracing method based on a fuzzy neural network in one embodiment of the present invention.

[0065] Figure 5 The diagram illustrates the response at prediction time step 10 of an interpretable knowledge tracing method based on a fuzzy neural network according to an embodiment of the present invention. Detailed Implementation

[0066] The preferred embodiments of the present invention will be described in detail below to provide a clearer understanding of the purpose, features, and advantages of the invention. It should be understood that the following embodiments are not intended to limit the scope of the invention, but are merely illustrative of the essential spirit of the technical solution of the invention.

[0067] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known techniques associated with the invention may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0068] The following is a description of the specific meanings of technical terms, English abbreviations, and formula parameters that may be used in this invention:

[0069] Knowledge tracing (KT) refers to the technology of dynamically modeling learners' mastery of various knowledge concepts by analyzing their historical response sequences and predicting their future response performance. It is a core module of intelligent education systems.

[0070] Fuzzy Neural Network (FNN): A hybrid model that combines fuzzy logic and neural networks. It can use fuzzy logic to process uncertain information and leverage the learning ability of neural networks to optimize model parameters.

[0071] Membership degree: It represents the probability that a learner's response or cognitive state belongs to a certain fuzzy level, and is the core indicator of fuzzy logic modeling.

[0072] Attention mechanism: a technique widely used in deep learning. Its core idea is to enable the model to automatically focus on the parts of the input data that are most relevant to the current task and give these parts higher weights, thereby improving the model's performance and interpretability.

[0073] FNN-KT: Knowledge tracing based on fuzzy neural networks.

[0074] DKT (Deep Knowledge Tracing) uses an LSTM recurrent neural network to encode learners' historical response sequences, mapping cognitive states to latent vectors to achieve dynamic prediction of knowledge mastery. However, it suffers from the problem of insufficient interpretability due to its black-box nature.

[0075] AKT: Attention-based Knowledge Tracing introduces an attention mechanism to replace recurrent neural networks. By calculating the attention weights between the current practice and historical practice, it filters highly correlated historical information to assist in cognitive state modeling, improving the ability to capture long sequence dependencies. However, the reasoning process still lacks transparency.

[0076] GKT: Graph-based Knowledge Tracing models the relationships between knowledge concepts as graph constructions, uses graph neural networks to learn the representation of knowledge concepts, captures the impact of interactions between knowledge on cognitive states, and optimizes prediction performance in knowledge association scenarios.

[0077] DKVMN: Dynamic Key-Value Memory Networks for Knowledge Tracing. It adopts a key-value memory network structure, where keys correspond to knowledge concepts and values ​​correspond to the knowledge mastery status. The memory content is updated through dynamic read and write operations, enabling independent modeling of the cognitive status of different knowledge concepts.

[0078] KQN: Knowledge Query Network. It introduces a query mechanism to retrieve relevant information from historical cognitive states based on the knowledge needs of the current exercise, thereby enhancing the model's ability to accurately locate knowledge concepts and improving the relevance of predictions.

[0079] EKTA: Enhanced Knowledge Tracing with Attention optimizes the attention mechanism design based on AKT, further improving the accuracy of capturing key historical information.

[0080] EKTM: Enhanced Knowledge Tracing with Memory optimizes the memory structure of memory network models such as DKVMN, improving the ability to store and dynamically update long-term cognitive states.

[0081] DGEKT: Dynamic Graph Enhanced Knowledge Tracing, which introduces a dynamic graph mechanism on the basis of GKT, updates the graph structure of the relationship between knowledge concepts in real time according to the learner's answer process, and more accurately captures the impact of the dynamic changes in knowledge relationships on cognitive state during the learning process.

[0082] like Figure 1 As shown, in one embodiment, an interpretable knowledge tracing method based on a fuzzy neural network according to the present invention includes the following steps S1-S5:

[0083] S1, Input information is collected and input into the computer system; the input information includes the learner's historical practice sequence, current practice, and practice to be predicted; the historical practice sequence includes each historical practice and its corresponding knowledge concept, as well as the continuous scores of the knowledge concept; the current practice includes its corresponding response performance;

[0084] S2, The computer system executes a computer program to perform fuzzification processing on the learner's continuous scores on knowledge concepts, mapping them into multiple fuzzy scores, and describing the membership degree of the fuzzy scores through a Gaussian membership function;

[0085] S3. Based on the attention mechanism, assign a relevance weight to each historical practice in the historical practice sequence and the current practice, and dynamically filter out relevant historical practice information that is conducive to updating the cognitive state according to the relevance weight.

[0086] S4 determines the learner's current fuzzy cognitive state based on the learner's previous fuzzy cognitive state and the current performance in the exercise, and realizes the updating of fuzzy cognitive state, local fuzzy reasoning, and global fuzzy cognitive state summary through three types of preset fuzzy rules.

[0087] S5, the computer system executes a computer program to concatenate the updated and determined global fuzzy cognitive state with the exercise to be predicted, and obtains the learner's predicted response to the exercise through a linear function or a sigmoid function.

[0088] The method provided by this invention is based on an interpretable cognitive state model of a fuzzy neural network. This model takes the high-order relationship between the question and knowledge concept and the temporal dependency relationship of the answer sequence as input. On this basis, it combines a fuzzy neural network to model the cognitive state and provide interpretability to the model. It also combines the latest cognitive state to predict new questions, so as to become a complete knowledge tracking model.

[0089] The steps S2-S5 above correspond to the four modules in the model, including the fuzzification module, the relevant question selection module, the fuzzy inference module, and the prediction module.

[0090] The fuzzification module maps learners' responses to a fuzzy distribution across multiple intervals, unifying the evaluation methods for both objective and subjective questions.

[0091] The reasoning module uses predefined fuzzy rules to dynamically update learners’ cognitive state changes and reveal their underlying logic to provide interpretability.

[0092] The relevant question selection module dynamically filters out information from historical exercises that are most helpful for updating cognitive states.

[0093] The prediction module outputs prediction results based on cognitive state.

[0094] In one feasible implementation, the fuzzification module is described in detail:

[0095] The continuous score is fuzzified into several fuzzy scores. Each continuous score has a probability of belonging to each fuzzy score, which is called the membership degree. The Gaussian membership function is used to describe the membership degree of the fuzzy scores, as shown in the following formula:

[0096]

[0097] Where j represents the total number of fuzzy score sets. This represents the continuous score for knowledge concept k at time t. express The membership degree belongs to fuzzy fractions , and Representing fuzzy fractions The mean and standard deviation.

[0098] In one feasible implementation, the relevant exercise selection module is described in detail:

[0099] When learners respond to new exercises, they tend to choose more similar exercises to update their cognitive state. However, in the learner's historical practice, exercises related to knowledge concepts are not clustered together but exist throughout the entire historical practice sequence. This distributional characteristic presents the model with two challenges when updating cognitive state:

[0100] First, how to effectively find the exercises most relevant to the current exercise within a complex historical sequence.

[0101] Secondly, how to balance the short-term effects of adjacent practice with the long-term effects of distant related practice. To address these challenges, this invention introduces an attention mechanism.

[0102] Specifically, the attention mechanism can assign a relevance weight to each practice in the historical practice sequence, thereby dynamically filtering out the information from historically relevant practices that are most helpful for updating cognitive state.

[0103] By introducing an attention mechanism, the model can not only highlight the contribution of local neighboring practice, but also flexibly capture the influence of distant but relevant historical practice.

[0104] Subsequently, after completing the relevant historical practice screening based on the attention mechanism, the model updates the learner's cognitive state through the fuzzy inference module to obtain the overall probability distribution of the cognitive state at the next moment.

[0105] In learners' new exercises When answering questions, the relevant practice selection module needs to be selected from the historical practice sequence. Extracting and The most relevant information.

[0106] This invention calculates its correlation through an attention mechanism, as shown in the following formula:

[0107]

[0108] in, , indicating the current exercise The query vector; The key vector matrix representing the history of exercises; This represents the value vector matrix of historical exercises. It represents the dimension of the key vector, used for scaling to avoid excessively large values; Ensure that the weights are normalized.

[0109] Therefore, the model can provide historical practice. Assign a weight The calculation is shown in the following formula:

[0110]

[0111] in, This represents the relationship between the i-th historical practice and the current practice. The correlation contribution is used to obtain the attention output, as shown in the following formula:

[0112]

[0113] This represents a weighted aggregation of historical information, including context information most relevant to the current exercise, and... It is passed as input to the fuzzy inference module to update the learner's cognitive state.

[0114] In one feasible implementation, the fuzzy inference module is described in detail:

[0115] The fuzzy reasoning module is the core module that provides interpretability for the interpretable knowledge tracing method based on fuzzy neural networks provided in this invention.

[0116] It determines the current fuzzy cognitive state based on the previous moment's fuzzy cognitive state and the current performance in answering questions.

[0117] Therefore, the vague cognitive state of the previous moment and the current practice performance are different. The combination corresponds to a total of A fuzzy rule.

[0118] Therefore, fuzzy rule nodes Output As shown in the following formula:

[0119]

[0120] Where i represents the total number of fuzzy cognitive state sets, and j represents the total number of fuzzy score sets.

[0121] The calculation for fuzzy cognitive states is shown in the following formula:

[0122]

[0123] in, For adjustable weights, .

[0124] To characterize the dynamic changes in learners' cognitive states during the response process, this invention designs three types of fuzzy rules (R1~R3): R1 corresponds to cognitive state updates, R2 is local fuzzy reasoning, and R3 is global state summarization.

[0125] In one embodiment, the specific rules for R1 are described as follows:

[0126] The specific rules for R1 are shown in the following formula:

[0127]

[0128]

[0129] This rule indicates that when the knowledge concept corresponding to the exercise at a certain time t is If the state of the knowledge concept is the same as the currently focused knowledge concept k, then the state of that knowledge concept is... The fuzzy reasoning module will use the state of the previous moment to determine the state. Update the program to reflect the learning outcomes of new exercises.

[0130] Conversely, if the knowledge concepts being practiced are... If the concept is different from the knowledge concept k being studied, it means that the knowledge concept has not been examined, and its fuzzy cognitive state remains unchanged from the previous moment. .

[0131] The R1 rule mainly implements the update process of fuzzy cognitive state. It will only be updated after the relevant knowledge concept has been learned through practice; otherwise, it will maintain the state of the previous moment.

[0132] In one embodiment, the specific rules for R2 are described as follows:

[0133] The specific rules for R2 are shown in the following formula:

[0134]

[0135]

[0136]

[0137]

[0138] This rule indicates that in The learner's cognitive state regarding knowledge concept k at any given moment Belongs to fuzzy set Its membership degree is And the learner's performance at time t Belongs to fuzzy set Its membership degree is The activation strength of both rules is determined by , The decision is made jointly, and its calculation is as follows: As shown.

[0139] Together, they determine the impact on the current state of ambiguous cognition, and this impact is transmitted to all possible levels of mastery through a weighted approach. .

[0140] The R2 rule is mainly used to describe the interaction between learners' past fuzzy cognitive states and their current response performance, and to update the distribution of their fuzzy cognitive states in a weighted manner.

[0141] In one embodiment, the specific rules for R3 are described as follows:

[0142] The specific rules for R3 are shown in the following formula:

[0143]

[0144]

[0145]

[0146]

[0147] This rule is used to integrate all activated local fuzzy inference rules R2 to obtain the learner's cognitive state at time t. If the fuzzy rule nodes... Cognitive state The function is ,but The probability is obtained by summing the contributions of all rule nodes, as shown in the following formula:

[0148]

[0149] It unifies the results of local reasoning into the overall cognitive state, thus completing the update from local conditions to the global cognitive state.

[0150] In this embodiment, to prevent too many rules from causing interpretation difficulties, the number of fuzzy score sets and fuzzy cognition sets I and J in the model are determined by expert experience.

[0151] The three rules R1 to R3 complete the cognitive state update process in the knowledge tracking task. R1 controls the conditions for updating the cognitive state, R2 establishes the local reasoning relationship between the cognitive state and the response performance, and R3 completes the summary of the cognitive state from the local to the global level. It can effectively capture the dynamic changes in the learner's knowledge level and provide interpretability for the prediction results.

[0152] The fuzzy reasoning process is completed collaboratively by three rule sets: rule set R1 is responsible for selecting the knowledge concepts to be updated from the fuzzified response results.

[0153] The rule set R2 is responsible for receiving the cognitive state of the knowledge concept to be updated and the current answer result. The number of rules it contains depends on all possible combinations of the fuzzy cognitive state set and the fuzzy score set. Each rule node in the rule set affects the current cognitive state according to its corresponding rule.

[0154] The rule set R3 is responsible for aggregating the outputs of all rule nodes to obtain the probability that the learner belongs to each fuzzy cognitive state.

[0155] In this invention, the network structure of the model is constructed from prior knowledge and fuzzy rule system in the field of education. Each neural node corresponds to a fuzzy reasoning rule (such as the selection of knowledge concepts, cognitive updates, and aggregation). This approach ensures the mapping between the network at the logical level and educational semantics, making the model inherently interpretable.

[0156] In one feasible implementation, the prediction module is described in detail:

[0157] After the fuzzy reasoning module, the learner's global cognitive state at a certain moment is obtained. Therefore, this invention obtains a prediction of the response to a new exercise by splicing the cognitive state with the exercise to be predicted, as shown below:

[0158]

[0159] in, This represents the connection between the learner's global cognitive state and practice characteristics. Represents the weight matrix and bias;

[0160] Specifically, when It's a subjective question. For the linear function; when It's an objective question. This refers to the Sigmoid function.

[0161] In this embodiment, the parameters in the fuzzification module The mean and standard deviation of the corresponding membership function are used to determine the degree to which a learner's answer score belongs among different fuzzy scores. This indicates the teacher's central judgment on learners' scores. A fuzzy boundary reflecting learners' scores.

[0162] In the fuzzy inference module, Represents a rule node The weighting of the influence on fuzzy cognitive states determines the direction and intensity of learners' cognitive state updates under a specific combination of fuzzy scores and cognitive states, reflecting the dynamic changes in learners' knowledge acquisition under different learning scenarios.

[0163] In the prediction module, This indicates the contribution of different cognitive states to the prediction of learning performance.

[0164] Traditional deep knowledge tracing methods often rely on implicit vector state updates, and their hidden layer parameters are difficult to map to specific cognitive behavior processes. In contrast, the proposed method introduces a fuzzy rule system to make the relationship between learners' answers, knowledge concepts and changes in cognitive states explicit, enabling the reasoning process to be explained using rules.

[0165] This invention can characterize both the learning behavior of the input and the predictive performance of the output through a series of rules, thus possessing good interpretability.

[0166] The performance of the present invention is analyzed in the following embodiments:

[0167] Table 1. Performance comparison of the model on three benchmark datasets.

[0168]

[0169] As shown in Table 1, the performance of this invention on the three benchmark datasets Assistments2009, Assistments2012, and Algebra2005 is in the middle range, higher than DKT, KQN, GKT and DGEKT models, but lower than AKT, DKVMN and EKT (EKTA and EKTM are two variants of EKT, EKTA follows a policy similar to an attention mechanism, and EKTM follows a policy similar to a Markov attribute policy).

[0170] The main reason is that the model sacrifices some accuracy in order to balance interpretability and accuracy. It can be seen that FNN-KT still maintains good prediction performance despite sacrificing some accuracy.

[0171] It is worth noting that, on the Assistments2012 dataset, all models except the GKT model suffer some performance loss.

[0172] The main reason is that the Assistments2012 dataset contains a large number of exercises, but only a small portion of the exercises are labeled with knowledge concept relationships. Therefore, using graph-structured data can better represent the relationship between its exercises and knowledge concepts, making it perform well on GKT based on graph-structured data models.

[0173] Meanwhile, the performance degradation of the FNN-KT model of this invention is not significant because this invention also uses GCN to aggregate and learn its practice-knowledge concept graph during the embedding process. However, FNN-KT mainly makes predictions based on changes in cognitive state and is still a model based on sequence data. Therefore, its performance still decreases on this dataset, but the magnitude is small.

[0174] The two variants of EKT demonstrate superior performance on all three datasets. This is primarily because most current models only utilize learners' practice logs and rarely pay attention to the text content of the practice. EKT, through its Exercise-Enhanced Recurrent Neural Network (EERNN) framework, focuses on both the learner's practice logs and the corresponding text content, resulting in excellent performance.

[0175] In summary, by introducing fuzzy neural networks, the FNN-KT model can effectively model changes in learners' cognitive states, exhibits good performance in knowledge tracing tasks, and provides a certain degree of interpretability, demonstrating its strong potential and application value.

[0176] The interpretability of the present invention is analyzed in the following embodiments.

[0177] The interpretability of a model is divided into the intrinsic interpretability of the model's operation process and the ex-post interpretability of the results.

[0178] In this embodiment, the main focus is on interpreting the model from its inherent interpretability through experiments, thereby obtaining the interpretability of the method of the present invention.

[0179] Regarding inherent interpretability, this invention uses rules for interpretation. The fuzzy reasoning module of this invention consists of a large number of rule nodes and follows fuzzy rules. Since the logical decision rules of IF-THEN are inherently interpretable, this embodiment only needs to verify that the definition of fuzzy rules in this invention is reasonable to show that the model has inherent interpretability.

[0180] Therefore, the following experiment in this embodiment demonstrates that the definition of fuzzy rules in this invention is reasonable.

[0181] Select several students from the dataset who are doing the same knowledge concept exercises, record their prediction process, obtain their fuzzy score at the current moment and their fuzzy cognition at the previous moment after fuzzification, and obtain their fuzzy cognition at the current moment based on fuzzy rules. Compare the fuzzy cognition at the current moment with the fuzzy cognition at the previous moment.

[0182] This indicates whether the user's fuzzy cognition has improved or declined compared to the previous moment. It predicts the user's performance score in practicing related knowledge concepts in the next moment and compares it with the score at the current moment to determine whether the score has increased or decreased. It also compares the predicted score with the actual score to observe whether the predicted score trend is consistent with the actual score, so as to ensure that the fuzzy rules defined in this invention are reasonable.

[0183] Therefore, this embodiment selects ten records from the dataset of a learner, all of which are consecutive responses to a specific knowledge concept, and will demonstrate the results of predicting the 7th, 8th, 9th, and 10th records.

[0184] like Figure 2 As shown, after the fuzzification module, it can be observed that the learner's actual score at time step 6 is 0.52, the probability of the fuzzy score belonging to the middle category is the highest (0.46), and the probability of the fuzzy cognition belonging to the middle category is the highest (0.49). Through fuzzy rule reasoning, it is found that the probability of the fuzzy cognition belonging to the middle category at time step 7 is the highest (0.58). This indicates that the learner's fuzzy cognition has improved compared to the previous time step, so the learner's prediction of the answer at time step 7 is correct.

[0185] like Figure 3As shown, it can be observed that the learner's actual score at time step 7 is 1, the probability of the fuzzy score being "medium" is the highest (0.52), and the probability of the fuzzy cognition being "medium" is the highest (0.58). Through fuzzy rule reasoning, it is found that the probability of the fuzzy cognition being "good" at time step 8 is the highest (0.48). This indicates that the learner's fuzzy cognition has improved compared to the previous time step. Therefore, the learner's predicted performance at time step 8 is 0.59, which is higher than the previous subjective question score of 0.52. The predicted score is consistent with the actual score, showing an increase.

[0186] like Figure 4 As shown, it can be observed that the learner's actual score at time step 8 is 0.64. Among the fuzzy scores, the probability of being "good" is the highest (0.51), and the probability of being "good" in fuzzy cognition is the highest (0.48). Through fuzzy rule reasoning, it is found that the probability of being "good" in fuzzy cognition at time step 9 is the highest (0.53). This indicates that the learner's fuzzy cognition has improved compared to the previous moment. Therefore, the learner's prediction of the answer at time step 9 is correct. However, the actual answer shows that the answer to this exercise is wrong. This may be because the exercise is too difficult or there are mistakes in the answer.

[0187] like Figure 5 As shown, it can be observed that the learner's actual score at time step 9 is 0, and the probability of being "medium" among the fuzzy scores is the highest (0.53). The probability of being "good" in fuzzy cognition is the highest (0.53). Through fuzzy rule reasoning, the probability of being "medium" in fuzzy cognition at time step 10 is the highest (0.59), indicating that their fuzzy cognition has declined compared to the previous time. Therefore, their predicted performance at time step 10 is 0.58, which is lower than the previous subjective question score of 0.64. The predicted score also shows a consistent trend with the actual score, both showing a decline.

[0188] The above analysis shows that performance on objective questions is more susceptible to factors such as guessing, mistakes, and the difficulty of practice compared to subjective questions. However, when cognition improves or declines, the performance on subjective questions is as expected; that is, when cognition improves or declines relative to the previous moment, the predicted score also increases or decreases simultaneously, and the trend is consistent with the actual score. This demonstrates that the fuzzy rules defined in this invention are reasonable, as they reveal the process of cognitive state changes, thus explaining the prediction results and providing interpretability for the model.

[0189] The ablation process of the present invention is analyzed in the following embodiments:

[0190] Table 2 Performance comparison of variant methods on three benchmark datasets

[0191]

[0192] As shown in Table 2, FNN-KT outperforms the model with the embedding module and practice selection module removed on all three datasets. After removing the embedding module, the model struggles to capture the higher-order relationships between practice and knowledge concepts and the temporal relationships between practice sequences.

[0193] The performance degradation was most pronounced on the Assistments2012 dataset. The main problem was that most exercises on this dataset were not labeled with knowledge concepts, resulting in a too-scattered and unusable relationship between exercises and knowledge concepts. Graph-structured data is better able to represent the relationship between exercises and knowledge concepts. Therefore, removing the embedding model makes it difficult to capture the relationship between exercises and knowledge concepts, leading to excessive performance degradation. This also shows that GCN can capture the relationship between exercises and knowledge concepts very well, and that this relationship helps improve the performance of the knowledge tracking model.

[0194] The performance also decreased after removing the practice selection module, indicating that practice related to the prediction has a greater effect on updating the cognitive state. When both the embedding module and the practice selection module were removed, the performance was lower than the other two variants, indicating that the embedding module and the practice selection module can be used simultaneously to improve model performance.

[0195] The technical terms, principles, or means related to the technical solutions of the present invention mentioned in the above embodiments, which are not described in detail above, are all well-known technologies or common practices that are known to those skilled in the art.

[0196] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for interpretable knowledge tracing based on fuzzy neural networks, characterized in that, Includes the following steps: The input information is collected and entered into the computer system. The input information includes the learner's historical practice sequence, current practice, and practice to be predicted. The historical practice sequence includes each historical practice and its corresponding knowledge concept, as well as the continuous scores of the knowledge concept. The current practice includes its corresponding response performance. The computer system executes a computer program to perform fuzzification processing on the learner's continuous scores on knowledge concepts, mapping them into multiple fuzzy scores, and describing the membership degree of the fuzzy scores using a Gaussian membership function; Based on the attention mechanism, each historical practice in the historical practice sequence is assigned a relevance weight to the current practice, and historical practice information that is conducive to updating the cognitive state is dynamically selected according to the relevance weight. The learner's current fuzzy cognitive state is determined based on their previous fuzzy cognitive state and their performance in the current exercise. The fuzzy cognitive state is updated, local fuzzy reasoning is performed, and the global fuzzy cognitive state is summarized through three types of preset fuzzy rules. The computer system executes a computer program to concatenate the updated and determined global fuzzy cognitive state with the exercise to be predicted, and obtains the learner's predicted response to the exercise through a linear function or a sigmoid function.

2. The interpretable knowledge tracing method based on fuzzy neural networks according to claim 1, characterized in that, The membership degree execution step, which describes the fuzzy score using the Gaussian membership function, is performed using the following formula: Specifically, the formula is: ； Where j represents the total number of fuzzy score sets. This represents the continuous score for knowledge concept k at time t. express The membership degree belongs to fuzzy fractions. and This represents the mean and standard deviation of the fuzzy scores.

3. The interpretable knowledge tracing method based on fuzzy neural networks according to claim 1, characterized in that, The execution step of assigning a relevance weight to the current practice for each historical practice in the historical practice sequence based on the attention mechanism includes: When the learner answers the current exercise, information related to the current exercise is extracted from the historical exercise sequence, and its relevance is calculated through the attention mechanism; Specifically, the current exercise is represented as a query vector, and each exercise in the historical exercise sequence is represented as a key vector and a value vector, respectively; Calculate the similarity between the query vector and all key vectors, and normalize the similarity into weights using the Softmax function; Finally, the weights are used to sum all the value vectors to obtain a weighted aggregate representation of the historical practice information.

4. The interpretable knowledge tracing method based on fuzzy neural networks according to claim 1, characterized in that, The steps for determining its current ambiguous cognitive state include: The number of combinations of the fuzzy cognitive state set and the fuzzy score set of the previous moment and the current practice response corresponds to the number of fuzzy rules; the output of the fuzzy rule node is shown in the following formula: Specifically: ； Where i represents the total number of fuzzy cognitive state sets, and j represents the total number of fuzzy score sets; The calculation of the fuzzy cognitive state is shown in the following formula: ； in, For adjustable weights, .

5. The interpretable knowledge tracing method based on fuzzy neural networks according to claim 1 or 4, characterized in that, The execution steps for updating the fuzzy cognitive state using the first type of fuzzy rule are as follows: When the knowledge concept corresponding to the exercise at a certain moment is the same as the knowledge concept currently being focused on, the update condition of the knowledge concept state is triggered. Specifically, the update condition is to update the fuzzy cognitive state of the previous moment using the calculation formula of the fuzzy cognitive state; If the cognitive concept being practiced is different from the cognitive concept being focused on, then the state of fuzzy cognition remains unchanged from the previous moment. The execution steps for implementing local fuzzy inference using the second type of fuzzy rules are as follows: The activation intensity of the fuzzy rule nodes is determined by the probability that the learner's fuzzy cognitive state of the knowledge concept in the previous moment belongs to the corresponding fuzzy set, and the probability that the learner's answer performance in the current moment belongs to the corresponding fuzzy set. Based on the activation intensity, the interaction between the previous moment's ambiguous cognitive state and the current moment's response performance is transmitted to all possible levels of mastery through a weighted method. The execution steps for summarizing the global fuzzy cognitive state using the third type of fuzzy rule are as follows: Integrate all the fuzzy rule nodes corresponding to the activated second type of fuzzy rules to obtain the learner's fuzzy cognitive state at the current moment; The degree of influence of the fuzzy cognitive state can be obtained by a certain fuzzy rule node and the influence weight of the fuzzy rule node on the fuzzy cognitive state. The probability of the fuzzy cognitive state is obtained by summing the contributions of all fuzzy rule nodes. The summation result unifies the local reasoning results into the overall cognitive state, completing the update from local conditions to a global fuzzy cognitive state.

6. The interpretable knowledge tracing method based on fuzzy neural networks according to claim 1, characterized in that, The steps for obtaining learners' predicted responses to the prediction exercises are shown in the following formula: ； in, This represents the connection between the learner's global cognitive state and practice characteristics. Represents the weight matrix and bias; Specifically, when the prediction practice It's a subjective question. For the linear function; when the exercise to be predicted It's an objective question. This refers to the Sigmoid function.

7. An interpretable knowledge tracing system based on a fuzzy neural network, characterized in that, include: The information input module is used to input the learned learners' historical practice sequences, current practice, and practice to be predicted as raw data into the system; The fuzzification module is used to map learners' continuous scores on knowledge concepts into a fuzzy distribution with multiple intervals; The related practice module is used to filter relevant information from historical practice through an attention mechanism; The fuzzy reasoning module is used to update the learner's fuzzy cognitive state based on preset fuzzy rule units; The prediction module combines the global fuzzy cognitive state with the features of the exercise to be predicted and outputs the prediction result of the exercise to be predicted.

8. The interpretable knowledge tracing system based on a fuzzy neural network according to claim 7, characterized in that, The fuzzy reasoning module includes: The first type of fuzzy rule unit is used to control the conditions for updating fuzzy cognitive states; The second type of fuzzy rule unit is used to establish a local reasoning relationship between fuzzy cognitive state and response performance; The third type of fuzzy rule unit is used to summarize the fuzzy cognitive state from local to global.

9. An electronic device, comprising: The processor, communication interface, memory, and communication bus are characterized in that the processor, communication interface, and memory communicate with each other via the communication bus. The memory stores a computer program that, when executed by the processor, causes the processor to perform the interpretable knowledge tracing method based on a fuzzy neural network as described in any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed, it implements the interpretable knowledge tracing method based on fuzzy neural networks as described in any one of claims 1 to 6.

Images