A trend inference deep knowledge tracking method and system

Abstract

The application provides a trend inference deep knowledge tracking method and system, comprising: obtaining a historical practice interaction sequence of a target object, the historical practice interaction sequence being used to represent answering conditions of different exercises at different time nodes; inputting the historical practice interaction sequence into a trend inference model to obtain a knowledge change vector of the target object, the knowledge change vector being used to represent a knowledge change process of the target object, and the trend inference model being built by game playing of pseudo interaction information generated based on the historical practice interaction sequence and real interaction information corresponding to the historical practice interaction sequence; and predicting a knowledge state of the target object for a target exercise according to the knowledge change vector, an exercise vector corresponding to the historical practice interaction sequence and interaction time of each exercise. The application introduces time information, can pay attention to a change feature of the knowledge state with time and effectively improves accuracy of knowledge state prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent data processing, and in particular to a trend inference deep knowledge tracking method and system. BACKGROUND

[0002] Knowledge tracking is to model students' knowledge states in different learning activities. It is a widely studied problem in the field of computer-aided education. In this section, we briefly review the following related work. Traditional knowledge tracking methods include Bayesian knowledge tracing (BKT), collaborative filter, etc. BKT is a classic and outstanding knowledge tracking method, which has been widely studied. The BKT model models the latent knowledge state of a student as a set of binary variables, representing whether a certain knowledge component (KC) is mastered. KC can be understood as a knowledge concept, principle, fact or skill in a generalized sense. For a specific question q t , KC can be regarded as an answer to q t, the knowledge that must be mastered. In the single KC model (i.e., one question corresponds to one KC), the question and the KC can be considered equivalent. With the rise of deep learning, the deep knowledge tracing (DKT) model was proposed in 2015, which uses multiple free parameters and multiple concept cross-inputs to produce better results compared to BKT. Following the mainstream method of DKT, there are many other development methods, such as the EKT (exercise Enhanced knowledge Tracing) and NPA (Neural News Recommendation with Personalized Attention) models, which are bidirectional long short-term memory (Bi-LSTM) models that establish an attention mechanism at the top layer. The networks in them have deep LSTM layers that can analyze the sequential nature of interactions. Inspired by memory-augmented neural networks, the dynamic key-value memory network (DKVMN, Dynamic Key-Value Memory Networks for Knowledge Tracing) creates a static matrix and a dynamic matrix to store all concepts and maintain the student's knowledge learning state, respectively. SAKT (self attentive knowledge tracing) uses a self-attention mechanism to assign weights to previously answered exercises to predict a student's performance in a particular exercise. EERNN (Exercise-enhanced sequential modeling for student performance prediction) uses the text content of the exercises and the student's exercise records to predict the student's performance. The existing technology also proposes a student learning modeling method based on Transformer, which modifies the original Transformer architecture and adds consideration of forgetting in the learning process. Models such as RNN, LSTM, and Transformer have shown great promise in time series tasks, but they focus on the order of input information, ignoring the frequency and stability of input nodes.

[0003] However, there are two limitations of the deep learning-based knowledge tracing model. First, these models ignore the fact that the knowledge state is a non-stationary time series, so they cannot capture the complex nature of the student's learning activities over time. Second, the model lacks the ability to remember and reason about different times of knowledge state. Although they can all make good use of the explicit relationship between exercises and knowledge to predict the student's concept mastery at a certain time in the past, when the time is inconsistent with the prediction time, these models often cannot make accurate predictions. For example, the training set and the test set are collected at different times, which may cause the model trained by the training set to have unsatisfactory prediction effect on the test set, because the knowledge state has changed at different times. SUMMARY

[0004] In view of the problems existing in the prior art, the present application provides a trend reasoning deep knowledge tracing method and system, which mainly solves the problem that the existing knowledge state prediction method is difficult to capture the characteristics of learning activities changing over time and has poor prediction accuracy.

[0005] In order to achieve the above-mentioned purpose and other purposes, the technical scheme adopted by the present application is as follows.

[0006] The present application provides a trend reasoning deep knowledge tracing method, comprising:

[0007] obtaining a historical exercise interaction sequence of a target object, the historical exercise interaction sequence being used to represent the answering conditions of different exercises at different time nodes;

[0008] inputting the historical exercise interaction sequence into a trend reasoning model to obtain a knowledge change vector of the target object, the knowledge change vector being used to represent the knowledge change process of the target object, and the trend reasoning model being built by playing the pseudo interaction information generated based on the historical exercise interaction sequence with the real interaction information corresponding to the historical exercise interaction sequence;

[0009] predicting the knowledge state of the target object for a target exercise according to the knowledge change vector, the exercise vector corresponding to the historical exercise interaction sequence, and the interaction time of each exercise.

[0010] In an embodiment of the present application, the step of building the trend reasoning model by playing the pseudo interaction information generated based on the historical exercise interaction sequence with the real interaction information corresponding to the historical exercise interaction sequence comprises:

[0011] providing a trend reasoning network and a discriminant network;

[0012] inputting the historical exercise interaction sequence into the discriminant network to obtain a first discriminant result corresponding to the real interaction information;

[0013] input the historical practice interaction sequence into the trend inference network to obtain a knowledge change curve corresponding to the pseudo interaction information;

[0014] input the knowledge change curve into the discriminative network to obtain a corresponding second discriminative result, perform game playing based on the first discriminative result and the second discriminative result, adjust parameters of the trend inference network and the discriminative network, and obtain the trend inference model.

[0015] In an embodiment of the present application, the trend inference network comprises a first input layer and a plurality of first convolution units connected in series, each of the first convolution units comprises a plurality of convolution layers and at least one multi-scale convolution layer, the input layer is connected with each convolution layer in the first convolution unit, respectively, the multi-scale convolution layer fuses and pools the output features of each convolution layer in the current convolution unit and takes the pooled features as the input of each convolution layer in the next convolution unit, and the output of the last convolution unit is integrated with the features of the input layer after being down-sampled and then up-sampled to obtain the pseudo interaction features.

[0016] In an embodiment of the present application, the discriminative network comprises a second input layer, a plurality of second convolution units connected in series, and a fully connected layer, the structure and connection mode of the second convolution units are the same as those of the first convolution units, and the discriminative result of the real interaction features and the pseudo interaction features is obtained after sequentially passing through the second input layer, the second convolution units, and the fully connected layer.

[0017] In an embodiment of the present application, the game playing based on the first discriminative result and the second discriminative result comprises:

[0018] a first loss function of the discriminative network is constructed by minimizing the first discriminative result and maximizing the second discriminative result;

[0019] a second loss function of the trend inference network is constructed by minimizing the second discriminative result;

[0020] the network parameters of the discriminative network and the trend inference network are updated according to the first loss function and the second loss function.

[0021] In an embodiment of the present application, the historical practice interaction sequence is input into the trend inference model to obtain the knowledge change vector of the target object, comprising:

[0022] the features obtained by the last time of down-sampling of the trend inference network in the trend inference model are taken as the model output to obtain the knowledge change vector of the target object.

[0023] In an embodiment of the present application, the knowledge state of the target object for a target question is predicted according to the knowledge change vector, the question vector corresponding to the historical practice interaction sequence, and the interaction time of each question, and the prediction method comprises the following steps:

[0024] providing an interaction prediction network;

[0025] composing the knowledge change vector, the question vector, and the corresponding interaction time into an input matrix as the input of the interaction prediction network, and giving different weights to the data in different dimensions in the input matrix through the interaction prediction network;

[0026] normalizing the knowledge change vector, the question vector, and the corresponding interaction time to obtain the knowledge state.

[0027] In an embodiment of the present application, the normalization of the knowledge change vector, the question vector, and the corresponding interaction time is represented as:

[0028]

[0029] wherein K is the knowledge state, F is the knowledge change vector, Q is the question vector, and d k is the dimension of the historical feature vector or the question vector, and T is the time sequence corresponding to the interaction time.

[0030] The present application also provides a trend inference deep knowledge tracking system, which comprises:

[0031] a data acquisition module configured to acquire a historical practice interaction sequence of a target object, the historical practice interaction sequence being used to represent the answering conditions of different questions at different time nodes;

[0032] a trend inference module configured to input the historical practice interaction sequence into a trend inference model to obtain a knowledge change vector of the target object, the trend inference model being obtained by playing a game based on a knowledge change curve and interpolation information generated based on the historical practice interaction sequence;

[0033] an interaction prediction module configured to predict a knowledge state of the target object for a target question according to the knowledge change vector, the question vector corresponding to the historical practice interaction sequence, and the interaction time of each question.

[0034] As described above, the present application provides a trend inference deep knowledge tracking method and system, which has the following beneficial effects.

[0035] The application performs trend inference based on historical exercise interaction sequences, enriches time-varying knowledge features, and introduces interaction time for knowledge state prediction, which can effectively improve the accuracy of knowledge state prediction and provide reliable data support for further judgment of target object performance based on the predicted knowledge state. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The figure shows the exercise records of students at different time intervals in an embodiment of the application.

[0037] Figure 2 The figure shows the flowchart of the trend inference deep knowledge tracking method in another embodiment of the application.

[0038] Figure 3 The figure shows the schematic diagram of the trend inference deep knowledge tracking framework in an embodiment of the application.

[0039] Figure 4 The figure shows the network architecture of the trend inference network and the discriminative network in an embodiment of the application.

[0040] Figure 5 The figure shows the architecture of the knowledge state prediction based on the trend inference model in an embodiment of the application.

[0041] Figure 6 The figure shows the module diagram of the trend inference deep knowledge tracking system in another embodiment of the application. DETAILED DESCRIPTION

[0042] The embodiments of the present application will be described in detail below with specific reference to the drawings. Those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the specification. The present application can also be implemented or applied in other different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0043] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and only show the components related to the present application in the diagrams, not the number, shape and size of the components when actually implemented. The actual implementation of each component may be a random change in shape, number and proportion, and the layout pattern of the components may also be more complex.

[0044] Knowledge Tracing (KT) is a powerful tool for realizing artificial intelligence assisted education, and has become a major component of Intelligent Tutoring System (ITS), which is widely used in various online education platforms. KT aims to establish a model of the change of students' knowledge state over time to determine the degree of students' mastery of knowledge. Existing KT models are based on a basic assumption that a series of observed interactive behaviors of students (such as the correctness of students' answers) during the completion of an exercise reflect their potential knowledge state (such as the proficiency of required skills) at the current time. In fact, this ignores the fact that the knowledge state changes in real time and is non-stationary. After a period of time, or even after each interactive behavior, the knowledge state of the student changes in different situations. The size of the change is determined by various factors, such as the content of the exercise, the understanding ability and learning habits of the student, and the previous activity also affects the trend of the subsequent activity. This leads to a model that performs well on the training set but is difficult to apply in practice.

[0045] Please refer to Figure 1 , Figure 1 For the exercise records of a student at different time intervals in an embodiment of the present application, a student has performed 6 exercises e1, e2,... e6, each exercise is at a different time interval, and each exercise completes 5 exercises, each of which tests different skills. In the first three exercises e1, e2, e3, the student's score is on the rise and finally reaches 100 points. However, in the last three exercises e4, e5, e6, the student's performance is quite different even for the same knowledge points, and the last score even drops. The reason is that the time interval of each exercise is different. The time interval of exercises e1, e2, e3 is 30 minutes, and the interval is one week to start exercises e4, e5, e6, each interval is 2 days. It can be predicted that the knowledge state obtained through a certain stage of exercise (such as e1, e2, e3) is not necessarily accurate for predicting other stages (such as e4, e5, e6). Therefore, it is dangerous to ignore the influence of time on the knowledge state.

[0046] Based on the problems existing in the prior art, the present application proposes a TRKT (Trend Reasoning Knowledge Tracing) framework, which attempts to introduce time information when establishing a knowledge tracing model, capture the change curve of the knowledge state, and retain the knowledge state of multiple time periods, solving the influence of time change on knowledge tracing. The framework focuses on exploring the relationship between time and knowledge to improve the prediction accuracy of the model and the interpretability of the method.

[0047] Please refer to Figure 2 , Figure 2This is a flowchart illustrating a trend-reasoning deep knowledge tracing method according to another embodiment of this application. The trend-reasoning deep knowledge tracing method provided in this embodiment includes the following steps:

[0048] Step S200: Obtain the historical practice interaction sequence of the target object. The historical practice interaction sequence is used to represent the answering situation of different exercises at different time points.

[0049] In one embodiment, the target object can be a student or any object that needs to practice exercises; here, a student is used as an example. Students first select a set of exercises to practice (e.g., e1, ..., e4) and leave their answers (e.g., right or wrong). Then, the goal of knowledge tracing is to infer their actual knowledge state of the corresponding concepts (such as equations) to predict their performance on subsequent exercises. Typically, the KT task is formalized as a supervised sequence learning task, given a record of the student's historical learning interactions. t ={i1, i2, ..., i t The implicit knowledge state of students is extracted through a pre-defined model, and its changes over time are tracked. Each interaction in a student's historical practice interaction sequence is typically represented as a tuple x of a question and its answer. i =(q i c i ), here q i c represents the question the student is trying to answer. i This indicates whether the student answered the question correctly, c i ∈0,1. Since it is difficult to directly measure a learner's actual learning state, existing KT models typically employ an alternative solution where the model predicts the probability P(c) of answering the next question correctly. i+1 =correct|x0,…,x i q i+1 This application embodiment changes the form of the sequence input, using q i As a predictor, with c i The predictor variable is P(c) that enables the model to predict the probability P(c) of answering the current question correctly. i correct|q i In addition to the main input and the target, two auxiliary inputs were used. First, the model was input u. i Therefore, it doesn't need to guess which student will answer the question next. Secondly, we will use the timestamp t i The input (measured in hours) is fed into the model. This input is crucial for understanding time-frequency information. Formally, the model attempts to predict P(c i =correct|u i q i , t i ,).

[0050] Step S210, inputting the historical practice interaction sequence into a trend inference model to obtain a knowledge change vector of the target object, the trend inference model being built by playing the generated pseudo interaction information with the real interaction information corresponding to the historical practice interaction sequence.

[0051] In an embodiment, generally, the prior art encodes the sequence data by using a model such as CNN or RNN (including GRU or LSTM), and then uses the hidden state at the last t time as the vector output of the sequence data. Such an encoding method cannot reflect the attention degree to different time points in a sequence data. In the real world, different parts of a sequence have different meanings and importance, for example, the results of the final exam and the regular practice. The result of the final exam usually has more reference significance. The final exam is more representative, not only because of the comprehensiveness of the questions, but also because of the individual characteristics of the students. Since the importance of each practice cannot be known, the TRKT starts from the individual characteristics of the students to find the weight of each interaction in the interaction sequence, and more accurately simulates and predicts the dynamic change of the knowledge state of the students.

[0052] Please refer to Figure 3 , Figure 3 is a schematic diagram of the trend inference deep knowledge tracking framework in an embodiment of the present application. The structure of the TRKT is shown in Figure 3 It consists of three main parts: trend inference network, discriminant network and interaction prediction network. The role of the trend inference network is to generate interaction results that conform to the trend of the student's knowledge change, which is optimized through the discriminant network. The ultimate goal is to expect to map the student's knowledge change curve to the middle layer of the trend inference network. At the same time, the framework gives the highest weight to the current time point from the trend of the knowledge state change by introducing the time parameter, obtains the knowledge state at the corresponding time point, and interacts with the corresponding exercise to predict the interaction result, realizing joint optimization. The key idea of TRKT is to mine deeper data relationships (i.e. the trend of the student's individual characteristics) from existing data and represent them in a reasonable way. By constructing Uvector, we represent the individualized knowledge state of the student as a high-dimensional dense vector, which can better reveal the relationship between different students and exercises. The model plays with the real historical interaction information and the generated interpolation information to fit the change curve of the characteristics, judge the current knowledge state of the student, and can predict the future. Finally, the model fuses multiple information, including the interaction exercises and knowledge points, and the structural information of their mutual relationship. At the same time, the student attributes given by the model can be explained to some extent.

[0053] In an embodiment, the step of building the trend reasoning model by playing the game between the pseudo-interaction information generated based on the historical practice interaction sequence and the real interaction information corresponding to the historical practice interaction sequence comprises:

[0054] In step S211, a trend reasoning network (Trend Reasoning Net) and a discriminator network (Discriminator Net) are provided.

[0055] In an embodiment, by playing the game between the trend reasoning network and the discriminator network, in the environment lacking personalized feature labeling information, the discriminator network is added to the reasoning model to conduct confrontation, which can not only generate a large amount of pseudo-labeling information, but also explore the trend of change of the data itself. In the whole training process, the discriminator network and the trend reasoning network learn from each other and are optimized together. The discriminator network helps the interpolation points generated by the trend reasoning network to be more accurate, and after the ability of the trend reasoning network is improved, it can help the discriminator network to make more accurate judgments.

[0056] In an embodiment, the trend reasoning network comprises a first input layer and a plurality of first convolution units connected in series, each of the first convolution units comprises a plurality of convolution layers and at least one multi-scale convolution layer, the input layer is connected with each convolution layer in the first convolution unit, respectively, the multi-scale convolution layer fuses and pools the output features of each convolution layer in the convolution unit to serve as the input of each convolution layer in the next convolution unit; the output of the last convolution unit and the features of the input layer are integrated after the output of the last convolution unit is down-sampled and then up-sampled through a plurality of first convolution units, to obtain the pseudo-interaction features.

[0057] In an embodiment, the discriminator network comprises a second input layer, a plurality of second convolution units connected in series and a fully connected layer, the structure and connection mode of the second convolution unit are the same as those of the first convolution unit, and the discrimination result of the real interaction features and the pseudo-interaction features is obtained after sequentially passing through the second input layer, the second convolution unit and the fully connected layer.

[0058] Please refer to Figure 4 , Figure 4This is a schematic diagram of the network architecture of the trend inference network and the discriminant network in one embodiment of this application. In the trend inference network, each first convolutional unit includes three convolutional layers. Taking five first convolutional units as an example, the convolutional layers of the first convolutional unit sequentially use 3x3, 5x5, and 7x7 convolutional kernels to convolve the input of the first input layer, respectively. The feature maps obtained from the convolutions are then subjected to feature fusion and pooling operations through multi-scale convolutional layers. The fused features are then input into the convolutional layers of the second convolutional unit. Subsequently, the first, second, and third convolutional units are concatenated for downsampling, mapping the historical practice interaction sequence into a knowledge change curve. The output of the third convolutional unit is received by the fourth and fifth convolutional units for upsampling, interpolating the knowledge change curve. Finally, the first input layer is mapped to the output of the fifth convolutional unit, and the historical practice interaction sequence is fitted with the generated pseudo-interaction information to obtain the knowledge change curve corresponding to the pseudo-interaction information.

[0059] like Figure 4 As shown, the discriminant network consists of a second input layer, three second convolutional units, and a fully connected layer. The structure and connection of the first two second convolutional units are exactly the same as the structure and connection of the corresponding first two first convolutional units in the trend inference network. The third second convolutional unit can be chosen to have the same structure as the second convolutional unit. The fully connected layer can be a two-layer structure, with each layer using a 1x1 convolutional kernel.

[0060] Step S212: Input the historical practice interaction sequence into the discriminant network to obtain the first discriminant result corresponding to the real interaction information.

[0061] In one embodiment, the real historical practice interaction sequence is first input into the discriminant network, which then learns the real interaction information corresponding to the real historical practice interaction sequence.

[0062] Step S213: Input the historical practice interaction sequence into the trend reasoning network to obtain the knowledge change curve corresponding to the pseudo-interaction information.

[0063] In one embodiment, the trend inference network performs personalized feature trend inference for the corresponding student based on the student ID information, and generates pseudo-interaction information at different time points; the generated information is sent to the discriminant network, which uses the discriminant network to identify the difference between the generated pseudo-interaction information and the real information, and optimizes the trend inference network in reverse.

[0064] Step S214, input the knowledge change curve into the discrimination network to obtain a corresponding second discrimination result, perform a game based on the first discrimination result and the second discrimination result, adjust parameters of the trend reasoning network and the discrimination network, and obtain the trend reasoning model.

[0065] When the discrimination ability of the discrimination network becomes more and more accurate, the trend reasoning network will again improve the generation ability to try to generate pseudo interaction information that cannot be discriminated; at the same time, because the pseudo information is more and more close to the real information, the discrimination network will further improve its discrimination ability in order to distinguish the two. The optimization of the two networks is repeated and alternated to form a game. This is a method based on zero-sum game. Through the discrimination network, the optimization of the trend reasoning network is realized, and compared with directly training the trend reasoning network, the problem of being unable to label individual characteristics can be avoided.

[0066] In an embodiment, performing a game based on the first discrimination result and the second discrimination result comprises:

[0067] constructing a first loss function of the discrimination network by minimizing the first discrimination result and maximizing the second discrimination result;

[0068] constructing a second loss function of the trend reasoning network by minimizing the second discrimination result;

[0069] updating network parameters of the discrimination network and the trend reasoning network according to the first loss function and the second loss function.

[0070] The mathematical expression of the discrimination network can be written as formula 1 and formula 2:

[0071] D: y = d(x, θ D ) (1)

[0072]

[0073] wherein D represents the discrimination network, x is real interaction data, is generated pseudo interaction data. θ D is a parameter in the discrimination network; y and is the output of the discriminator, which is also a discrimination result (i.e. the first discrimination result and the second discrimination result mentioned above). In the training process, y and Two kinds of loss functions are used for optimization, and the loss function is shown in formula 3:

[0074]

[0075] The physical meaning of formula 3 is that by minimizing d(x, θ D), the discriminator is expected to determine x as true, so as to learn the characteristics of the real interaction information and enhance the discrimination ability. By letting max, the discriminator is expected to determine x as false, so as to enhance the discrimination ability of the discriminator to the generated information.

[0076] The mathematical expression of the trend inference network is shown in equation 4.

[0077]

[0078] where G represents the trend inference network, z is the information of the student; θ G is the parameter in the model. is the output of the model, that is, the generated pseudo information. It is sent to the discriminator network, and the model is optimized through the identification result. The loss function is shown in equation 5.

[0079]

[0080] The physical meaning of equation 5 is that by letting min, the discriminator is expected to determine x as true, that is, the discriminator cannot find the difference between the generated interaction information and the real interaction information.

[0081] The trend inference model is obtained by gradient descent training based on the two loss functions.

[0082] In step S220, the knowledge state of the target object for the target question is predicted according to the knowledge change vector, the question vector corresponding to the historical practice interaction sequence, and the interaction time of each question.

[0083] In an embodiment, it is assumed that the historical practice interaction sequence is input into the trend inference model to obtain the knowledge change vector of the target object, which includes:

[0084] The feature obtained by the last time down-sampling of the trend inference network in the trend inference model is taken as the model output to obtain the knowledge change vector of the target object.

[0085] Please refer to Figure 5 , Figure 5 is the architecture schematic diagram for predicting the knowledge state based on the trend inference model in an embodiment of the present application. Specifically, it is assumed that the student a has n interaction information, and the question information and practice time of the i-th interaction information are respectively q i , t i , where q i represents the question of the i-th interaction, and t i ​​denotes the time of the ith interaction, i takes value 0 < i < n. This application uses the output of the trend inference network l layer (i.e. the output of the last down-sampling) as the personalized features of student a, and uses f a denotes, predicting time, f a , q i , t v , q v , t v .

[0086] wherein:

[0087]

[0088] q v = w q × q i

[0089] t v = w t × t i

[0090] In an embodiment, the knowledge state of the target object for the target question is predicted according to the knowledge change vector, the question vector corresponding to the historical practice interaction sequence, and the interaction time of each question, comprising:

[0091] providing an interaction prediction network;

[0092] composing an input matrix of the knowledge change vector, the question vector, and the corresponding interaction time as the input of the interaction prediction network, and giving different weights to the data of different dimensions in the input matrix through the interaction prediction network;

[0093] normalizing the knowledge change vector, the question vector, and the corresponding interaction time to obtain the knowledge state.

[0094] Specifically, the interaction prediction network can adopt a self-attention network, and the specific network architecture can be adjusted according to actual application requirements.

[0095] The question vector q i is fused with the knowledge change vector to obtain K a , which obtains relevant information required for the current interaction of q i , and filters out knowledge noise. Here, K a represents the ability of student a on q i .

[0096] In an embodiment, the normalization method according to the knowledge change vector, the question vector, and the corresponding interaction time is represented as:

[0097]

[0098] where K is the knowledge state, F is the knowledge change vector, Q is the question vector, d k is the dimension of the history feature vector or question vector, and T is the time sequence corresponding to the interaction time.

[0099] Selecting the square root F dot Q can be made to satisfy the distribution with expectation 0 and variance 1, similar to normalization.

[0100] Since the knowledge state of the student is not constant, the evolution of the knowledge state at each time node also needs to be predicted. By adding time information, the final knowledge state of the student is obtained.

[0101] Based on the above technical solution, through the trend inference deep knowledge tracking architecture, reliable connections between exercises at different times can be learned, the generalization of the model is further improved, and the interpretability of the deep knowledge tracking model is enhanced; the game mechanism formed by the trend inference network and the discriminative network solves the mapping problem of the high-dimensional feature of the knowledge curve. Benefiting from the knowledge features of different time dimensions, the proposed framework can consider the correlation of interactions at each stage. In addition, another advantage of the framework is that it can capture higher-level semantic information of student individualization in the knowledge tracking process.

[0102] Please refer to Figure 6 The embodiment provides a trend inference deep knowledge tracking system for executing the trend inference deep knowledge tracking method described in the foregoing method embodiment. Since the technical principles of the system embodiment are similar to those of the foregoing method embodiment, the same technical details will not be repeated.

[0103] In an embodiment, the trend inference deep knowledge tracking system comprises: a data acquisition module 10 configured to acquire a historical exercise interaction sequence of a target object, the historical exercise interaction sequence being configured to represent the answering conditions of different questions at different time nodes; a trend inference module 11 configured to input the historical exercise interaction sequence into a trend inference model to obtain a knowledge change vector of the target object, the trend inference model being obtained by playing a game based on a knowledge change curve generated based on the historical exercise interaction sequence and interpolation information; and an interaction prediction module 12 configured to predict a knowledge state of the target object for a target question according to the knowledge change vector, a question vector corresponding to the historical exercise interaction sequence, and an interaction time of each question.

[0104] The above embodiments are only illustrative of the principles of the present application and its efficacy, and are not intended to limit the present application. Any modification or change made by any person skilled in the art without departing from the spirit and scope of the present application shall be covered by the claims of the present application.

Claims

1. A trend-based reasoning deep knowledge tracking method, characterized in that, include: Obtain the historical practice interaction sequence of the target object, which is used to represent the answering situation of different exercises at different time points; The historical practice interaction sequence is input into the trend reasoning model to obtain the knowledge change vector of the target object. The knowledge change vector is used to characterize the knowledge change process of the target object. The trend reasoning model is built by playing a game between the pseudo-interaction information generated by the historical practice interaction sequence and the real interaction information corresponding to the historical practice interaction sequence. The steps for constructing the trend reasoning model based on the pseudo-interaction information generated from the historical practice interaction sequence and the real interaction information corresponding to the historical practice interaction sequence include: providing a trend reasoning network and a discriminant network; inputting the historical practice interaction sequence into the discriminant network to obtain a first discriminant result corresponding to the real interaction information; inputting the historical practice interaction sequence into the trend reasoning network to obtain a knowledge change curve corresponding to the pseudo-interaction information; inputting the knowledge change curve into the discriminant network to obtain a corresponding second discriminant result; and adjusting the parameters of the trend reasoning network and the discriminant network by playing a game based on the first discriminant result and the second discriminant result to obtain the trend reasoning model. Based on the knowledge change vector, the exercise vector corresponding to the historical practice interaction sequence, and the interaction time of each exercise, the knowledge state of the target object for the target exercise is predicted.

2. The trend reasoning deep knowledge tracking method according to claim 1, characterized in that, The trend inference network includes a first input layer and multiple concatenated first convolutional units. Each first convolutional unit includes multiple convolutional layers and at least one multi-scale convolutional layer. The input layer is connected to each convolutional layer in the first convolutional unit. The multi-scale convolutional layer fuses and pools the output features of each convolutional layer in the current convolutional unit as the input to each convolutional layer in the next convolutional unit. By downsampling and then upsampling multiple first convolutional units, the output of the last convolutional unit is integrated with the features of the input layer to obtain pseudo-interaction features.

3. The trend reasoning deep knowledge tracking method according to claim 2, characterized in that, The discriminant network includes a second input layer, multiple convolutional units connected in series, and a fully connected layer. The structure of the second convolutional unit is the same as that of the first convolutional unit in terms of structure and connection method. After passing through the second input layer, the second convolutional unit, and the fully connected layer in sequence, the discrimination result between the real interaction features and the pseudo interaction features is obtained.

4. The trend reasoning deep knowledge tracking method according to claim 1, characterized in that, The game is based on the first and second discrimination results, including: The first loss function of the discrimination network is constructed by minimizing the first discrimination result and maximizing the second discrimination result; The second loss function of the trend inference network is constructed by minimizing the second discrimination result; The network parameters of the discriminant network and the trend inference network are updated based on the first loss function and the second loss function.

5. The trend reasoning deep knowledge tracking method according to claim 2, characterized in that, The historical practice interaction sequence is input into the trend reasoning model to obtain the knowledge change vector of the target object, including: The features obtained from the last downsampling of the trend inference network in the trend inference model are used as the model output to obtain the knowledge change vector of the target object.

6. The trend reasoning deep knowledge tracking method according to claim 1, characterized in that, Based on the knowledge change vector, the exercise vector corresponding to the historical practice interaction sequence, and the interaction time of each exercise, predict the knowledge state of the target object for the target exercise, including: Provide interactive prediction networks; The knowledge change vector, the exercise vector, and the corresponding interaction time are combined to form an input matrix, which is used as the input to the interaction prediction network. The interaction prediction network assigns different weights to the data in different dimensions of the input matrix. The knowledge state is obtained by normalizing the knowledge change vector, the exercise vector, and the corresponding interaction time.

7. The trend reasoning deep knowledge tracking method according to claim 6, characterized in that, The normalized representation based on the knowledge change vector, exercise vector, and corresponding interaction time is as follows: Where K represents the knowledge state, F represents the knowledge change vector, Q represents the exercise vector, and d k T represents the dimension of the historical feature vector or problem vector, and T represents the time series corresponding to the interaction time.

8. A trend-based reasoning deep knowledge tracking system, characterized in that, include: The data acquisition module is used to acquire the historical practice interaction sequence of the target object, which is used to represent the answering situation of different exercises at different time points; The trend reasoning module is used to input the historical practice interaction sequence into the trend reasoning model to obtain the knowledge change vector of the target object. The trend reasoning model is obtained by playing a game based on the knowledge change curve generated by the historical practice interaction sequence and interpolation information. The steps for constructing the trend reasoning model based on the pseudo-interaction information generated from the historical practice interaction sequence and the real interaction information corresponding to the historical practice interaction sequence include: providing a trend reasoning network and a discriminant network; inputting the historical practice interaction sequence into the discriminant network to obtain a first discriminant result corresponding to the real interaction information; inputting the historical practice interaction sequence into the trend reasoning network to obtain a knowledge change curve corresponding to the pseudo-interaction information; inputting the knowledge change curve into the discriminant network to obtain a corresponding second discriminant result; and adjusting the parameters of the trend reasoning network and the discriminant network by playing a game based on the first discriminant result and the second discriminant result to obtain the trend reasoning model. The interaction prediction module is used to predict the knowledge state of the target object for the target exercise based on the knowledge change vector, the exercise vector corresponding to the historical exercise interaction sequence, and the interaction time of each exercise.

Images