A cognitive intervention system of dynamic cognitive diagnosis combined with deep learning model

Abstract

The application discloses a cognitive intervention system of a dynamic cognitive diagnosis combined with a deep learning model, provides a classical scale to score the cognitive level of a patient, and provides a cognitive intervention project to diagnose the cognitive level of the patient; the cognitive intervention project is pre-trained by using a QuesNet model before use, and unsupervised heterogeneous intervention project representation is realized; when a task of the cognitive intervention project is performed, a user knowledge state tracking is performed on the pre-training result of the cognitive intervention project by using an EKT model in combination with the QuesNet model; by using a user-based collaborative filtering algorithm, the similar users can be recommended to the project with good effects on the similar users according to the knowledge tracking of the existing users, and the project with poor effects can be improved and eliminated according to the result of the QuesNet and the EKT model. The application can realize comprehensive evaluation on the intervention project, dynamic diagnosis on the cognitive level of the user in the intervention process, and dynamic design of a personalized intervention scheme.

Description

Technical Field

[0001] This invention belongs to the field of cognitive diagnostic technology, specifically relating to a cognitive intervention system that combines dynamic cognitive diagnosis with a deep learning model. Background Technology

[0002] The proliferation of online learning websites in China has disrupted traditional offline learning models. Leveraging technologies such as big data, artificial intelligence, and deep learning, online learning enables precise user targeting and needs analysis, resulting in better course and knowledge delivery for students. This model can be applied to similar online applications, Q&A delivery, intelligent customer service, and personalized medical rehabilitation. However, current online learning methods primarily utilize natural language processing to directly convert question statements into syntactic patterns or semantic codes, followed by supervised learning to optimize the model. However, the lack and low accuracy of labeled data significantly impacts the personalized data analysis and delivery performance of online learning systems.

[0003] Cognitive diagnostics is a necessary and fundamental task applied in fields such as education, gaming, and medical diagnosis. Typical dynamic cognitive diagnostic frameworks discuss the necessary user and project factors, expanding each user's comprehensive state vector into a knowledge state matrix that updates over time, where each vector represents their ability level in a certain cognitive dimension.

[0004] Current methods for processing project data primarily involve raw text processing. However, project data used in intervention systems is characterized by coherence and heterogeneity. The corpus contains questions in various forms, including text, images, and logical prior knowledge. Therefore, corpus analysis cannot simply be based on literal translation of text within the linguistic context (extracting low-level features); it must also consider the logical information it contains (extracting high-level features). Furthermore, the existing cognitive diagnostic methods and their combination with subsequent algorithms are insufficient, making it difficult to more effectively improve data analysis results and enhance the application effectiveness of online products.

[0005] Recommendation algorithms leverage user logs to automatically connect users with relevant items, helping users discover suitable interventions in information-overloaded environments. By analyzing users' historical behavior and the similarity between users, recommendation systems proactively recommend items most likely to meet their needs, effectively uncovering long-tail information. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a cognitive intervention system that combines dynamic cognitive diagnosis with a deep learning model. By pre-training the intervention items, the system can achieve dynamic cognitive diagnosis of users and cognitive intervention using recommendation algorithms, thereby solving the problems existing in the current online education or online medical intervention systems.

[0007] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0008] A cognitive intervention system combining dynamic cognitive diagnosis and a deep learning model, wherein when a cognitively impaired patient uses the system, a classic scale is provided to score the patient's cognitive level, and cognitive intervention items are provided to diagnose the patient's cognitive level.

[0009] Before use, the cognitive intervention project is pre-trained using the QuesNet model in deep learning: metadata, natural language processing and computer vision data analysis are performed on the project to achieve unsupervised heterogeneous intervention project representation.

[0010] When performing cognitive intervention projects, the cognitive intervention system collects and records data on user IDs, received intervention projects, correctness, and progress bar status. It uses a dynamic cognitive diagnosis method and combines the EKT model with the QuesNet model to track user knowledge status based on the pre-training results of cognitive intervention projects, thus creating a user profile.

[0011] The cognitive intervention system is equipped with a recommendation module, which uses a user-based collaborative filtering algorithm to recommend items that are effective to similar users based on their existing knowledge tracking. It can also improve and remove items that are ineffective based on the results of QuesNet and EKT models.

[0012] To optimize the above technical solution, the specific measures also include:

[0013] The QuesNet model described above consists of three layers: the embedding layer, the content layer, and the statement layer.

[0014] The embedding layer is used to embed heterogeneous corpus data. Its target input is the heterogeneous information representation of a single intervention item q, and its target output is the sentence representation mapping obtained after training the QuesNet model.

[0015] The heterogeneous information of a single intervention project q is represented as follows:

[0016] x = {x0, x1, ..., x} T};

[0017] x0 = m represents supplementary information, T is the item sequence length, and the rest of x... i The word can be a word or a W×H sized image.

[0018] The sentence representation mapping obtained after training the QuesNet model is (υ (i) υ (s) );

[0019] in, υ (s) It is an N-dimensional single-sentence representation vector;

[0020] The embedding layer is used to embed heterogeneous corpus data by processing the target input, resulting in an embedding representation vector sequence e = {e0, e1, ..., e...}. T}, meaning the project data input reaches the first embedded layer, according to x t The data types are processed according to the following conditions to map heterogeneous input data into a sequence of embedded representation vectors:

[0021]

[0022] Where, θ we With θ te The parameters for the word embedding model and the graph embedding model determine the modulus N of the vector after mapping words or images. e ;

[0023] After the heterogeneous corpus information is mapped to a unified domain in the embedding layer, it enters the content layer, that is, let

[0024] The content layer uses a multi-layer bidirectional LSTM to extract low-level semantic features and low-level language features. Finally, each process t is connected to extract bidirectional hidden states, obtaining the content representation for each layer.

[0025]

[0026] in, In a positive hidden state, For reverse hidden state, c t In cellular state, h t For the hidden state, θ LSTM These are the parameters of the LSTM network;

[0027]

[0028] The statement layer uses a self-attention mechanism to fuse high-level logical knowledge information with low-level semantic features. Specifically, a multi-head self-attention mechanism is used to aggregate the features of the input items to obtain statement representations.

[0029] The current N-dimensional representation (Query) of the corpus is denoted as matrix Q, the labels (Key) of all corpus are denoted as matrix K, and the actual corpus representation (Value) is denoted as matrix V. A multi-head self-attention mechanism is used to aggregate the features of the input items.

[0030] MultiHead(Q,K,V)=concatenate(head1,...,head H WO ,

[0031]

[0032]

[0033] Where H represents the number of attention heads, W O , It is a mapping matrix.

[0034] In practice, a self-attention mechanism is used to integrate the content representation sequence υ (i) Location-embedded information pe (i) .

[0035] The position embedding formula can use the trigonometric function position encoding method in the core transformer layer of the BERT model to generate a position vector (or position embedding, abbreviated as PE) for each token at a different position.

[0036]

[0037] Where p is the position, i is the dimension of the embedding vector, and d model is the dimension of the position vector.

[0038] After integration The system then enters a multi-head attention system, and finally uses max pooling to integrate the actual corpus representations at each time step, i.e.:

[0039]

[0040] In the pre-training stage of the QuesNet model, an unsupervised two-layer hierarchical training algorithm is used to further extract question information from the vectors. In the embedding layer, Word2Vec is used to calculate the initial word-vector mapping for the entire corpus. For the low-level features, HLM is used to extract semantic information from the large-scale corpus.

[0041] The loss function for the Quesnet model during the pre-training phase is as follows:

[0042]

[0043] in, This represents the hidden state outside of time t.

[0044] The aforementioned EKT model first focuses on model learning for users and projects, including:

[0045] 1) Knowledge Embedding: For the user set |U|, the intervention item set |E|, and the cognitive dimension |K|, the intervention process for each user is tracked and characterized as follows:

[0046] u={(k1, e1, r1), (k2, e2, r2),..., (k T e T r T )}(u∈U,e i ∈E, k i ∈K)

[0047] Where, r i With intervention project e i The contribution coefficient of the matched factor to the corresponding cognitive dimension;

[0048] Using a knowledge matrix Quantitatively characterize the different cognitive dimensions of a specific user to obtain the influence weight β for each stage. t ;

[0049] 2) Project Embedding: The target input is a user's usage, processed in stages. In stage t, the project information is input in a low-dimensional form for conceptual encoding. t =W k T k t ,,in The embedding matrix is ​​used, and Softmax is selected as the activation function. The influence weights are obtained by combining the matrix with the knowledge state matrix.

[0050]

[0051] Among them, M i For knowledge memory vectors;

[0052] 3) User data embedding: The knowledge state of a specific user at a specific time is... Using the LSTM model, it is represented as in For e i With r i The comprehensive For the parameters on the LSTM network in the cognitive dimension i.

[0053] The EKT model described above updates data based on Markov chains and selects ReLU and Sigmoid functions as activation functions:

[0054]

[0055]

[0056]

[0057] Where W1 and W2 are weight matrices, b1 and b2 are biases, and y T+1 This represents the overall performance at time T+1. For specific intervention projects e T+1 Performance prediction.

[0058] The recommendation module described above uses the following Pearson correlation coefficient to characterize the similarity between users:

[0059]

[0060] Among them, R i，p Let p represent user i's level on the cognitive dimension. This represents the average level of user i across all dimensions.

[0061] The present invention has the following beneficial effects:

[0062] This invention enables comprehensive evaluation of intervention projects, dynamic diagnosis of users' cognitive levels during the intervention process, and dynamic design of personalized intervention programs. Attached Figure Description

[0063] Figure 1 Here is a block diagram of the QuesNet model;

[0064] Figure 2 A flowchart illustrating a dynamic cognitive diagnostic method using EKT in conjunction with pre-training of intervention projects using the QuesNet model;

[0065] Figure 3 To develop personalized cognitive intervention plans for the recommendation module using recommendation algorithms;

[0066] Figure 4 This is a flowchart of the overall cognitive intervention system. Detailed Implementation

[0067] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0068] Technical basis of the present invention:

[0069] Cognitive psychology analysis can not only clarify the knowledge, skills, problem-solving strategies, and processing procedures required for subjects to answer correctly, but also clarify the relationship between item characteristics, stimulus conditions, and response, thereby significantly improving the predictability of difficulty and other performance aspects during test development. Therefore, it is a prerequisite for tests to achieve their cognitive and diagnostic functions through intervention tasks.

[0070] Psychometric models should be able to integrate cognitive psychology theories into psychometric models (cognitive diagnostic models), that is, to incorporate information such as the cognitive attributes contained in the items into appropriate mathematical models, thereby mathematically modeling an individual's cognitive structure. This allows for a combined qualitative and quantitative examination of differences in cognitive ability structures and among patients with chemotherapy-related cognitive impairment in breast cancer, enabling the diagnosis of an individual's cognitive status.

[0071] At the data processing level, in-depth analysis of the implementation data and estimation of unknown parameters in the model are necessary. Deep learning models can be used to achieve good application results by selecting appropriate cognitive diagnostic models that are complex enough to provide sufficient skill diagnostic information, have sufficiently concise model parameters, and fully fit the data.

[0072] See Figure 1 The present invention provides a cognitive intervention system that combines dynamic cognitive diagnosis with a deep learning model. When used by patients with cognitive impairment, the cognitive intervention system provides a classic scale to score the patient's cognitive level and provides a variety of cognitive intervention items to diagnose the patient's cognitive level.

[0073] Before use, the cognitive intervention project is pre-trained using the QuesNet model in deep learning: metadata, natural language processing and computer vision data analysis are performed on the project to achieve unsupervised heterogeneous intervention project representation.

[0074] When performing cognitive intervention projects, the cognitive intervention system collects and records data such as user ID, the received personalized intervention projects, the corresponding correctness and error status, and the time progress bar. It adopts a dynamic cognitive diagnosis method and uses the EKT model combined with the QuesNet model to track the user's knowledge state based on the pre-training results of the cognitive intervention projects, thus creating a user profile.

[0075] The cognitive intervention system is equipped with a recommendation module, which uses a user-based collaborative filtering algorithm to recommend items that are effective to similar users based on their existing knowledge tracking. It can also improve and remove items that are ineffective based on the results of QuesNet and EKT models.

[0076] The details are as follows:

[0077] I. Unsupervised heterogeneous intervention project representation based on the QuesNet model:

[0078] For the design of the cognitive intervention program set |E|: Based on extensive and in-depth research on classic foreign cognitive training software such as Brain-HQ, Cognifitim, Mindsparke, and Gogme, and taking into account the characteristics of the target user group, the team independently developed several cognitive training programs based on WeChat mini programs, including attention, executive function, spatial rotation, and working memory, as cognitive intervention programs. Because the project data comes from real application systems, it has good heterogeneity.

[0079] This invention utilizes QuesNet, an understanding-based deep learning algorithm based on a unified specific domain. QuesNet can aggregate heterogeneous data for a given problem into a unified specific domain, and further extract semantic information and logical knowledge. Combining QuesNet with various subsequent algorithms can effectively improve data analysis results and enhance the application effectiveness of online products. There are two methods for intelligent analysis of intervention projects: one is based on rule representation as a tree structure, and the other is based on natural language processing decomposition into semantic vectors. This invention chooses the latter. Furthermore, for text processing, a corpus can be obtained through feature extraction or preprocessing fine-tuning.

[0080] The sentences, images, and metadata of each intervention item are extracted for unsupervised heterogeneous representation, i.e., the number of each cognitive intervention item is determined.

[0081] By using the QuesNet model to process the cognitive intervention project library, the project information can be trained and processed to achieve effects such as knowledge mapping and difficulty estimation.

[0082] The overall architecture of the QuesNet model is mainly divided into three layers: the embedding layer, the content layer, and the statement layer. A training preprocessing stage is also designed.

[0083] 1. Embedded layer:

[0084] The intervention project implemented the embedding of heterogeneous corpus data one by one, and the heterogeneous input data was mapped into a sequence of embedded representation vectors:

[0085] The target input is a heterogeneous information representation of a single intervention item q, and the target output is a sentence representation mapping obtained after training the QuesNet model.

[0086] according to Figure 1 and Figure 2 The QuesNet model part of the document establishes a deep learning framework and pre-trains the aforementioned created project.

[0087] Inputting existing cognitive intervention project sets |E| into the QuesNet deep learning framework:

[0088] x = {x0, x1, ..., x} TThe additional information is x0 = m, and the item sequence length T, and the rest x i The word can be a word or a W×H sized image.

[0089] Single sentence representation vector υ (s) Given N dimensions, the final expected training result for item q is mapped to (υ) using an unsupervised heterogeneous item representation method. (i) υ (s) ).

[0090] The embedding layer performs heterogeneous corpus data embedding by processing the target input, resulting in an embedded representation vector sequence e = {e0, e1, ..., e...} T}, meaning the project data input reaches the first embedded layer, according to x t The data type is determined according to the following conditions

[0091] The process maps heterogeneous input data into a sequence of embedded representation vectors:

[0092]

[0093] The parameters θ of the word embedding model and the graph embedding model are... we With θ te Determine the modulus N of the vector after mapping words or images e

[0094] 2. Content layer:

[0095] Then, the heterogeneous corpus information is mapped to a unified domain in the embedding layer and then enters the content layer, that is, let

[0096] The content layer uses a multi-layer bidirectional LSTM to extract low-level semantic features and low-level language features. Finally, each process t is connected to extract bidirectional hidden states, obtaining the content representation for each layer.

[0097]

[0098] in, In a positive hidden state, For reverse hidden state, c t In cellular state, h t For the hidden state, θ LSTM These are the parameters of the LSTM network;

[0099]

[0100] 3. Statement level:

[0101] The statement layer uses a self-attention mechanism to fuse high-level logical knowledge information with low-level semantic features. Specifically, a multi-head self-attention mechanism is used to aggregate the features of the input items to obtain statement representations.

[0102] The current N-dimensional representation (Query) of the corpus is denoted as matrix Q, the labels (Key) of all corpus are denoted as matrix K, and the actual corpus representation (Value) is denoted as matrix V. A multi-head self-attention mechanism is used to aggregate the features of the input items.

[0103] MultiHead(Q,K,V)=concatenate(head1,...,head H W O ,

[0104]

[0105]

[0106] Where H represents the number of attention heads, W O , It is a mapping matrix.

[0107] In practice, a self-attention mechanism is used to integrate the content representation sequence υ (i) Location-embedded information pe (i) .

[0108] The position embedding formula can use the trigonometric function position encoding method in the core transformer layer of the BERT model to generate a position vector (or position embedding, abbreviated as PE) for each token at a different position.

[0109]

[0110] Where p is the position, i is the dimension of the vector, and d model is the dimension of the position vector.

[0111] After integration The system then enters a multi-head attention system, and finally uses max pooling to integrate the actual corpus representations at each time step, i.e.:

[0112]

[0113] Finally, by synthesizing the results of the unsupervised heterogeneous intervention project representations, a high-level logical feature is established, the target output is obtained, and the information mapping of the intervention project is derived (υ). (i) υ (s) )

[0114] 4. Pre-training phase:

[0115] In the pre-training stage, an unsupervised two-layer hierarchical training algorithm is used to further extract problem information from the vectors.

[0116] The initial word-vector mapping is computed for the entire corpus using Word2Vec in the embedding layer.

[0117] To address low-level features, a Holed Language Model (HLM) is proposed to extract semantic information from a large-scale corpus. The objective function of HLM considers the semantic environment of the context.

[0118] for Figure 2 2.11 indicates that the encoding function for image data in actual pre-training is emb. i The parameter is θ ie Set the decoding function It also has trainable parameters θ id

[0119] At this point, the loss function for the autoencoder of the image data is: Where l can be selected as the mean squared error as the loss function, and the metadata method is similar.

[0120] Set the initial weights for the image data and metadata encoder. and Enter a multi-layer fully connected neural network.

[0121] The initial weights are obtained from the extreme values ​​of the loss function of the corresponding autoencoder:

[0122]

[0123] For text data, then use Figure 2 The 2.12 pore language model, that is, when processing each position t of the item input sequence, needs to consider the semantics of its two sides, i.e., its context:

[0124]

[0125] like Figure 2 Output of the pore language model This involves combining the hidden state vectors on both sides of the given position, which in data processing is represented by concatenating the two:

[0126]

[0127] After embedding various types of data, the data is processed through a fully connected neural network layer. w Obtain the output; the parameters for this layer are θ. owAnd for each type of output, cross-entropy or mean squared error is selected as the loss function:

[0128]

[0129] in, The hidden state outside of time t;

[0130] By summing the outputs at each position, we can obtain the low-level features (relationships and content, etc.) of the project.

[0131]

[0132] For high-level features, a domain-specific objective function is used to extract the logic and knowledge understanding for each problem.

[0133] Encode the options for item q to obtain υ opt =enc(opt), and model the accuracy of the options:

[0134]

[0135] The activation function is the Sigmoid function, D(·, θ). D ) is a one-layer fully connected neural network, and its output is one-dimensional, υ opt This refers to the answer choices for objective questions. This is the stem representation of the intervention project obtained by training using the QuesNet framework.

[0136] At this point, the high-level features of project q have been optimized using a domain-specific objective function, and after integrating the low-level features:

[0137]

[0138] Then, by applying optimization and update algorithms such as stochastic gradient descent to the pre-trained model, the parameter θ can be adjusted. D With θ ow Perform updated training.

[0139] Arrive after Figure 2 Section 2.13 employs multi-head self-attention and max pooling to aggregate input features and obtain sentence representations.

[0140] The current N-dimensional representation (Query) of the corpus is denoted as matrix Q, the labels (Key) of all corpus are denoted as matrix K, and the actual corpus representation (Value) is denoted as matrix V. A multi-head self-attention mechanism is used to aggregate the features of the input items.

[0141] MultiHead(Q,K,V)=concatenate(head1,...,headH W O ,

[0142]

[0143]

[0144] Where H represents the number of attention heads, W O , It is a mapping matrix.

[0145] In practice, a self-attention mechanism is used to integrate the content representation sequence υ (i) Location-embedded information pe (i) .

[0146] The position embedding formula can use the trigonometric function position encoding method in the core transformer layer of the BERT model to generate a position vector (or position embedding, abbreviated as PE) for each token at a different position.

[0147]

[0148] Where p is the position, i is the dimension of the vector, and d model is the dimension of the position vector.

[0149] After integration The system then enters a multi-head attention system, and finally uses max pooling to integrate the actual corpus representations at each time step, i.e.:

[0150]

[0151] Finally, by synthesizing the representations of unsupervised heterogeneous intervention projects, a high-level logical feature is established, and the target output is obtained, which is the information mapping of the intervention project (υ). (i) υ (s) )

[0152] After pre-training the intervention project information, we recruited patients with breast cancer chemotherapy-related cognitive impairment to form a user set |U|, and began to put this cognitive intervention system, which combines dynamic cognitive diagnosis with a deep learning model, into use.

[0153] Fine-tuning can be made for different specific types of projects.

[0154] Existing research has selected traditional supervised learning algorithms, ELMo, BERT, and H-BERT, and compared them with QuesNet for experimental purposes, conducting experiments on three tasks: knowledge mapping, difficulty estimation, and student performance prediction. The results are shown in Table 1, where QuesNet demonstrates superior performance in all three tasks.

[0155] Table 1

[0156]

[0157] II. Using the EKT model for dynamic cognitive diagnosis to track users' cognitive states:

[0158] The EKT model, which utilizes dynamic cognitive diagnosis, aims to achieve more accurate predictions of user performance and more interpretable analysis. By extending the EERNN framework, the EKT framework expands the comprehensive state vector into a knowledge state matrix, which can effectively handle the problem of predicting user performance in future cognitive intervention projects. It can also be broken down into various domains (such as attention, memory, etc.), thus clearly identifying cognitive improvements in different aspects. Memory networks are used to quantify the degree of impact on different aspects of multiple cognitive levels of users during each cognitive intervention.

[0159] 1. The EKT model is designed for user and project-based model learning, including:

[0160] 1) Knowledge Embedding: For the user set |U|, the intervention item set |E|, and the cognitive dimension |K|, the intervention process for each user is tracked and characterized as follows:

[0161] u={(k1, e1, r1), (k2, e2, r2),..., (k T e T r T )}(u∈U,e i ∈E, k i ∈K)

[0162] Where, r i With intervention project e i The contribution coefficient of the matched factor to the corresponding cognitive dimension;

[0163] Using a knowledge matrix The aim is to quantitatively characterize the different cognitive dimensions of a specific user in order to obtain the influence weight β at each stage. t Since different cognitive dimensions are not independent during cognitive intervention, the correlation weights between cognitive dimensions can also be considered.

[0164] Information about the intervention item set |E| and the setting of the cognitive dimension |K| parameters can both be obtained using QuesNet and the results of pre-training.

[0165] 2) Figure 2 Section 2.22 is the project embedding layer. The target input is the user's usage, which is processed in stages. At stage t, the project information input is reduced in dimensionality for conceptual encoding. t =W k T kt ,in The embedding matrix is ​​used, and Softmax is selected as the activation function. The influence weights are obtained by combining the matrix with the knowledge state matrix.

[0166]

[0167] M i For the knowledge memory vector, β t The weight of the impact of time t on cognitive diagnosis is derived from the information collected on the items received by the user at time t.

[0168] 3) Figure 2 Section 2.21 involves the embedding of user data, where the knowledge state of a specific user at a specific time is...

[0169] Using the LSTM model, it is represented as

[0170] in For e i With r i The weighted sum, For the parameters on the LSTM network for the cognitive dimension i.

[0171] arrive Figure 2 Section 2.23 is the Markov-based performance prediction part, and its expected results are shown in [link to section]. Figure 4 Section 4.22 of the Behavior Prediction and Cognitive Diagnosis describes the methods for obtaining performance predictions and cognitive diagnoses for this user.

[0172] 2. Characteristics of using the Markov property assumption:

[0173] Based on the transitions between time steps, data is updated according to the Markov chain, and ReLU and Sigmoid functions are selected as activation functions:

[0174] It is the user's cognitive diagnosis at time T+1;

[0175]

[0176]

[0177] Where W1 and W2 are weight matrices, b1 and b2 are biases, and y T+1 This represents the overall performance at time T+1. For specific intervention projects e T+1 Performance prediction;

[0178] x T+1 For predicting overall user performance, for eT+1 Performance predictions.

[0179] 3. The pre-training results of the representations from the aforementioned unsupervised heterogeneous intervention projects can be combined:

[0180] QuesNet, when combined with various subsequent algorithms, can effectively improve data analysis results and enhance the application effectiveness of online products. Traditional EKT's e i While the previous method only processed the raw text information of the project, the results pre-trained using QuesNet offer greater flexibility and better utilize the coherent and heterogeneous nature of the data, extracting logical information (high-level features) on top of low-level features. This enables subsequent data processing and meets the needs of online applications.

[0181] Third, a collaborative filtering recommendation algorithm is used, combined with the user process information mentioned in points one and two above, to guide the design of personalized intervention programs for the cognitive intervention system:

[0182] use Figure 3 The recommendation model shown selects the recommendation algorithm as needed; this invention uses [the algorithm described]. Figure 4 The main user-based collaborative filtering recommendation algorithms used in versions 4.13 and 4.23 are:

[0183] Utilize each user Knowledge matrix analysis, using Pearson correlation coefficient to calculate user similarity:

[0184]

[0185] Among them, R i，p Let p represent user i's level on the cognitive dimension. The average level of user i across all dimensions can be selected from the knowledge state matrix in the user's dynamic cognitive diagnosis process described above. Since the calculation of the Pearson correlation coefficient becomes more complex with the increase in the number of users, the user's scores on the classic scale can also be selected and substituted into the Pearson correlation coefficient to characterize the similarity relationship between users.

[0186] In summary, the data collected and trained through dynamic cognitive diagnosis can be used to further train the online cognitive intervention system. For example, personalized intervention plans can be designed using a collaborative filtering recommendation algorithm that integrates the cognitive level records of all users, or intervention items with poor performance can be improved or removed. This will result in a more targeted and effective cognitive intervention system for patients with chemotherapy-related cognitive impairment in breast cancer.

[0187] Based on the items that a user received during the previous dynamic cognitive diagnosis process and which had a high weighting on improving cognitive level, i.e., the items that had a good effect on that user, they are recommended to similar users to obtain better cognitive intervention results.

[0188] The model input can be selected from the process data of the user's dynamic cognitive diagnosis. Since the calculation of the Pearson correlation coefficient is complex when the number of users is large, the user's scores on the classic scale can also be selected and substituted into the Pearson correlation coefficient to characterize the similarity relationship between users.

[0189] The model recommendation method involves observing a user's dynamic cognitive diagnosis process and basing it on the influence weight β. t Alternatively, it can display a prediction curve and recommend projects that this user has previously used and that have yielded good results to similar users.

[0190] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1.A cognitive intervention system of dynamic cognitive diagnosis combined with deep learning model, characterized in that, The cognitive intervention system provides a classical scale to score the cognitive level of the patient and a cognitive intervention project to diagnose the cognitive level of the patient when the patient with cognitive impairment uses the system; The cognitive intervention project is pre-trained by using a QuesNet model in deep learning before use: meta data, natural language processing and computer vision data analysis of the project are performed to realize unsupervised heterogeneous intervention project representation; The cognitive intervention system collects and records the user number, the received intervention project, and the corresponding correct or incorrect situation and time progress bar situation when the user performs the task of the cognitive intervention project, and uses a dynamic cognitive diagnosis method to track the knowledge state of the user by using an EKT model combined with the pre-training result of the QuesNet model to depict the user portrait; The cognitive intervention system is provided with a recommendation module, which can recommend projects with good effects to similar users according to the knowledge tracking of existing users by using a user-based collaborative filtering algorithm, and can also improve and eliminate projects with poor effects according to the results of the QuesNet and EKT models; The QuesNet model is divided into three layers: an embedding layer, a content layer and a sentence layer; The embedding layer is used to realize heterogeneous corpus data embedding, and the target input is a single intervention project q heterogeneous information representation, and the target output is a sentence representation mapping obtained after training of the QuesNet model: The single intervention project q heterogeneous information representation is: x = {x0, x1,..., x T}; x0= m is the side information, T is the length of the item sequence, and the rest of x i is a word or a picture of size W x H; The sentence representation obtained after the QuesNet model is trained is mapped to (v (i) , v (s) ); wherein, v (s) is an N-dimensional single sentence representation vector; An embedding layer is configured to implement heterogeneous corpus data embedding by processing the target input to obtain an embedding representation vector sequence e = {e0, e1,..., en}, where n is the length of the embedding representation vector sequence e. T}, i.e., the project data input reaches the first layer embedding layer, and is processed according to the data type of x t to map the heterogeneous input data into an embedding representation vector sequence according to the following conditions: Wherein, θ we With θ te The parameters of the word embedding model and the graph embedding model determine the modulus N e of the vector mapped after the word or picture; The heterogeneous corpus information is mapped to a unified domain in the embedding layer, that is, The content layer applies a multi-layer bidirectional LSTM to extract low-level semantic features and low-level language features, and finally connects each process t to extract a bidirectional hidden state to obtain a content representation of each layer: wherein is a forward hidden state, is a backward hidden state, c t is a cell state, h t is a hidden state, θ LSTM are parameters of the LSTM network; The sentence layer fuses high-level logical knowledge information and low-level semantic features based on a self-attention mechanism, specifically, a multi-head self-attention mechanism is used to aggregate the features of input items to obtain a sentence representation: The current N-dimensional representation of the corpus is denoted as a matrix Q, the labels Key of all corpora are denoted as a matrix K, and the actual corpus representation Value is denoted as a matrix V, and a multi-head self-attention mechanism is used to aggregate the features of input items: MultiHead(Q, K, V) = concatenate(head1,..., head H )W o , wherein H is the number of attention heads, W O , is a mapping matrix; In practice, the content representation sequence v is integrated using a self-attention mechanism (i) with the position embedding information pe (i) ; The position embedding formula can select a trigonometric function position encoding mode in a core transformer layer of a Bert model to generate a position vector pe for each token in different positions (i) ; where p is the position, i is the dimension number of the embedding vector, d model is the dimension number of the position vector; post-integration The multi-head attention system is entered, and finally the actual corpus representation at each moment is integrated by using maximum pooling, that is: 2.The cognitive intervention system of claim 1, wherein, In the pre-training stage of the QuesNet model, an unsupervised two-level hierarchical training algorithm is used to further extract problem information from the vector, Word2Vec is used in the embedding layer to calculate the initial word-to-vector mapping for the entire corpus, and HLM is used for low-level features to extract semantic information from a large-scale corpus; The loss function in the pre-training stage of the QuesNet model is as follows: wherein, is the hidden state at time t. 3.The cognitive intervention system of claim 1, wherein, The EKT model first learns the model for users and projects, including: 1) Knowledge embedding: for a user set |U|, an intervention project set |E| and a cognitive dimension |K|, each user intervention process is depicted as: u={(k1,e1,r1),(k2,e2,r2),...,(k T ,e T ,r T )}(u∈U,e i ∈E,k i ∈K) wherein r i With the intervention project e i The contribution coefficient of the matched pair to the corresponding cognitive dimension; Adopting knowledge matrix Quantitative description is made to different cognitive dimension levels of specific users to obtain influence weight β of each stage t ; 2) Project embedding: the target input is processed in stages for the use of a certain user, at the t-th stage, the project information is input into low dimension for concept coding v t = W k T k t where is the embedding matrix, and Softmax is selected as the activation function, and the influence weight is obtained with the knowledge state matrix: wherein M i is a knowledge memory vector; 3) User data embedding: the knowledge state of a specific user at a specific time is represented as is represented by an LSTM model, denoted as where is e i is the combination of r i is the parameter on the LSTM network for the cognitive dimension i.​ 4.The cognitive intervention system of claim 1, wherein, The EKT model updates data according to a Markov chain, and selects ReLU and Sigmoid functions as activation functions: where W1W2 are weight matrices, b1b2 are biases, y T+1 is the overall performance at time T+1, is the performance prediction for a specific intervention project e T+1 . 5.The cognitive intervention system of claim 1, wherein, The recommendation module uses the following Pearson correlation coefficient to depict the similarity relationship of users: wherein R i,p is the level of user i in cognitive dimension p, is the average level of user i in each dimension.

Images