A method for predicting learners' abnormal learning status in online education

Abstract

The present invention provides a method for predicting abnormal learning states of learners in online education, including: preprocessing high-dimensional online education platform log information and learner registration information and encoding them based on a self-supervised learning method to construct learner portrait features; constructing learner state features, then constructing a state feature sequence based on the generation time sequence of state features, and constructing a state feature graph based on the cosine similarity between state features; constructing a long short-term memory-graph attention deep network that is suitable for predicting the degree of learning difficulties in online education, determining the number of network layers, the number of neurons in each layer, and the input and output dimensions; constructing pseudo-labels based on noise labels to iteratively train the network; and using the trained network to predict the learner's abnormal learning state and its degree during the learning stage to be predicted. The present invention uses learner registration information and learner log information to predict the degree of learner state abnormality, providing a reference for teachers to provide targeted guidance and assistance to learners.

Description

Technical Field

[0001] The present invention relates to the technical field of online education, and in particular to a method for predicting abnormal learning states of learners in online education. Background Art

[0002] With the widespread adoption of modern computer networks and the rapid development of home electronic devices, online education using modern technologies such as computer networks and artificial intelligence has become an essential component of family education. The online education industry saw a 35.5% year-on-year increase in market size in 2020. However, as a means of distance learning, online education, characterized by limited teacher-student interaction and difficulty in monitoring learning, has had a significant negative impact on teaching effectiveness. In recent years, the annual increase in the number of online education participants and online courses has accumulated a massive amount of registration information and log data from online learners. Effectively analyzing this data and building models to predict abnormal learning states is crucial for teachers to understand learners' progress and provide targeted guidance and supervision.

[0003] Existing research primarily focuses on detecting poor learners, treating this as a 0 / 1 classification problem (normal learners vs. poor learners). This problem extracts features based on learner registration information and log data generated during the learning process. Using binarized learning scores as labels, supervised learning methods are used to predict poor learners. The following literature provides reference technical solutions to the problem of detecting poor learners:

[0004] Reference 1: A prediction method for poor learners based on campus card data (201910833015.9).

[0005] Reference 2: A method for identifying poorly performing learners in online education (201610864980.9).

[0006] Reference 1 proposes a method for predicting poor learners based on campus card data. This method represents the learner's campus card data in matrix form, uses learning outcomes as labels, and uses a combination of convolutional neural networks (CNNs) and long short-term memory neural networks (LSTMs) to train a binary classifier to determine whether a learner is a poor learner.

[0007] Reference 2 designed a method for identifying poor-performing learners in online education. Based on online education data, it constructed high-dimensional features based on time windows and learning duration. The training dataset was divided into positive and negative examples based on whether the score was greater than 60 points. A random forest algorithm was used to train a binary classifier to detect poor-performing learners.

[0008] The above-mentioned methods for detecting learners with abnormal learning status have the following main problems: First, both References 1 and 2 focus on identifying learners with poor learning status. In practical applications, it is particularly important to provide personalized guidance for learners with different levels of abnormal learning status. For example, learners with occasional and recurrent poor learning behaviors require different levels of supervision and guidance. Second, the learning outcomes or test scores of a semester or course cannot fully reflect whether learners need targeted supervision during the learning process. Test scores depend on the different emphases of the course questions. Students who study more diligently in the sections where the questions are emphasized will receive higher scores. As a result, learning scores cannot fully reflect the learner's learning status in the current teaching section. In online education, teachers hope to promptly detect learners with abnormal learning status in any teaching section and provide targeted guidance. Therefore, directly using learning scores as training labels to measure the learning status of the entire process introduces label noise. At the same time, References 1 and 2 mine learner portrait features or temporal representations for learning. The lack of attention to different perspectives of features also makes it difficult to perform label disambiguation based on multi-perspective features to address the noise problem. However, obtaining accurate labels of poor learning status and its degree requires a large amount of manual labeling costs. How to construct labels based on existing learning performance and multi-perspective features to more accurately evaluate learners' abnormal learning status and degree at each stage has become an urgent problem to be solved. Summary of the Invention

[0009] The purpose of the present invention is to provide a method for predicting abnormal learning states of learners for online education, so as to solve the problems existing in the prior art.

[0010] In order to achieve the above objectives, the present invention adopts the following technical solutions:

[0011] A method for predicting abnormal learning status of learners in online education, comprising:

[0012] First, the high-dimensional online education platform log information and learner registration information are preprocessed and encoded based on the self-supervised learning method to construct learner portrait features; secondly, the learner log information is fragmented based on the time window and encoded separately to construct learner state features, and then the state feature sequence is constructed based on the generation time sequence of the state features, and the state feature graph is constructed based on the cosine similarity between the state features; thirdly, a long short-term memory-graph attention LSTM-GAT deep network is constructed to predict the degree of learning difficulties in online education. Based on the constructed learner state features, state feature sequences and state feature graphs, the number of network layers, the number of neurons in each layer, and the input and output dimensions are determined. ; From this time, the learning results mapped to [0, 1] are used as noise labels, and pseudo labels are constructed based on the noise labels to iteratively train the network. Each iteration first optimizes the network based on multi-view features and pseudo labels, and then selects reliable samples based on the temporal local continuity, spatial local consistency and sample prediction error of the state features, and constructs a subgraph centered on each unreliable sample in the state feature graph. Finally, the reliable sample labels are aggregated in each subgraph to reconstruct the pseudo labels of the central samples, and the reconstructed pseudo labels are used for the next iterative training; finally, the trained network is used to predict the learner's abnormal learning state and its degree in the learning stage to be predicted.

[0013] A further improvement of the present invention is that the method specifically comprises the following steps:

[0014] 1) Learner registration information and log information processing

[0015] The learner registration information and learner log information are preprocessed as the learner's initial features. The initial features are encoded to have a unified mathematical representation. Then, a masked autoencoder is trained based on a self-supervised method. The encoded features are re-encoded based on the masked autoencoder to achieve feature dimensionality reduction and thus obtain learner profile features.

[0016] 2) Log fragment information processing and state feature sequence and state feature graph construction

[0017] Divide the learner's log information into multiple log information segments based on a time window, and obtain learner state features after performing preprocessing and encoding similar to the log information processing in step 1); construct a state feature sequence based on the time series generated by the state features of the same learner based on their logs, calculate the feature similarity between all learner state features, and construct a state feature graph based on the k-nearest neighbor method;

[0018] 3) LSTM-GAT deep network construction

[0019] A deep network group consisting of parallel LTSM and GAT is constructed to extract temporal and spatial representations respectively. The inputs of the two are from the state feature sequence and state feature graph respectively. The vector output by the LTSM and GAT deep network group is connected with the learner profile feature vector and input into the fully connected layer group. The fully connected layer group outputs the probability of abnormal learning state of the learner.

[0020] 4) LSTM-GAT deep network training based on multi-view features and label reconstruction

[0021] The learning results are mapped to [0, 1] as noise labels, and the learner state features and learner profile features are used as training features. Each training iteration is divided into two parts: network parameter update and pseudo-label reconstruction. During model training, the network parameters are updated based on the pseudo-labels constructed based on the noise labels. When reconstructing pseudo-labels, reliable samples are first selected, and then a subgraph is constructed with unreliable samples as the center. The pseudo-labels of the reliable samples in the subgraph are aggregated to reconstruct the pseudo-labels of the central unreliable samples.

[0022] 5) Prediction of learners’ abnormal learning status

[0023] The registration information and log information of the learners to be predicted are processed and encoded as the input of the trained LSTM-GAT model. After being processed by the model, the probability of the learner's learning state being abnormal is output.

[0024] A further improvement of the present invention is that, in step 1), the processing of learner registration information and log information specifically includes the following steps:

[0025] Step 1: Processing learner registration information

[0026] To address the problem that the original learner registration information contains many redundant and missing fields, we selected gender, date of birth, identity type, highest level of education, graduation school, and location information that are highly relevant to the learner and have few missing fields from these fields as initial features. At the same time, the date of birth retains the year of birth and is mapped into categorical features with every 10 years as an interval. The location is mapped into six categories as features: first-tier cities, new first-tier cities, second-tier cities, third-tier cities, fourth-tier cities, and fifth-tier cities based on the "2020 City Business Charm Ranking" of the New First-tier City Research Institute, thereby obtaining a unified mathematical representation of time and category features. The above features are then encoded using one-hot encoding and concatenated as learner registration information features.

[0027] The one-hot encoding method is as follows:

[0028] s1.Remember the category feature set as Yes, there are ni The possible values ​​of the category feature c i Set an n i Bit status register M i ;

[0029] s2.M i Each bit in represents c i Is a value valid? If it is valid, it is 1; if it is invalid, it is 0.

[0030] s3. For all i∈{1,2,3,…,l c}, c i Encode in sequence to obtain the one-hot vector of all category features;

[0031] Step 2: Log information processing

[0032] The statistical characteristics of each semester's log information reflect the learner's general status during the semester. Therefore, using sum, average, and count aggregation functions, we construct statistical data on the total number of course video views, total course video viewing duration, number of courses participated in, average single viewing duration, total number of comment interactions, average number of single video comment interactions, number of video pauses, average video viewing interval, and average video viewing interval within the same chapter from each semester's log information. The numerical features are mapped based on z-score standardization to a mean of 0 and a standard deviation of 1. This converts the redundant features contained in the log information, which are difficult to use directly, into directly usable learner statistical log features with a unified mathematical representation.

[0033] The specific steps of the z-score method are:

[0034] s1. The numerical feature set is Logarithmic feature v i Calculate its sample mean μ i and the sample standard deviation σ i ;

[0035] s2. Set the feature υ i Normalization is performed according to the following z-score formula:

[0036]

[0037] s3. For all i∈{1,2,3,…,l v}, change v i Encode in sequence to obtain the one-hot vector of all category features;

[0038] Step 3: Constructing learner profile feature encoding

[0039] The learner registration information and the log information of the set semester are combined to describe the learner's general status in the set semester. The learner's general learning status plays an important reference role in predicting the learner's learning status. The learner registration information and log information are combined and encoded to construct a learner portrait feature that can describe the learner's general learning status; the learner portrait feature encoding construction first connects the features generated in Step 1 and Step 2 to form a learner initial feature to describe the learner, and then constructs a masked autoencoder composed of a linear encoder and a linear decoder to process sparse features, and mines the deep association of the learner initial feature by partially masking and restoring the learner initial feature; the encoder is set to consist of an input layer and two hidden layers, the first hidden layer is composed of neurons, and the second hidden layer consists of neurons, and the length of the learner's initial feature vector is l s , then the input layer dimension is l s , where the neurons in the first hidden layer are of size The second hidden layer consists of a linear layer and a tanh activation function, and the neurons of the second hidden layer are The decoder consists of two layers with opposite structure and order to the encoder. At the same time, a batch normalization layer is set before activation to improve the convergence of the model. The tanh activation function is formally expressed as:

[0040]

[0041] The feature vector is randomly masked and then input into the autoencoder network. The mask rate is set in the training sequence M with a step size of t. t Step growth, t is a hyperparameter that controls the growth rate of the mask rate, M t is the hyperparameter sequence that controls the growth rate; the loss function is set to mean square error loss, and the decoder output vector is set to The original eigenvector is x s , then the mean square error loss function is expressed as:

[0042]

[0043] After training, the encoder in the masked autoencoder is taken out as the learner portrait feature encoder, and all learner initial features are encoded to obtain the learner portrait features.

[0044] A further improvement of the present invention is that, in step 2), the log segment information processing and the state feature sequence and state feature graph construction specifically include the following steps:

[0045] Step 1: Log fragment information processing

[0046] The learner's learning status is reflected by the learning situation within a period of time. Log fragment information processing divides the log information into log information fragments according to the set time window, and the log information in the time fragment reflects the learner's learning status in the current time fragment;

[0047] Specifically, first, the log information of each learner is divided into log information segments according to the week. Secondly, similar to the log information processing, the total number of chapter video views, the total course video viewing time, the number of courses participated in, the average single viewing time, the total number of comment interactions, the average number of single video comment interactions, the number of video pauses, the average video viewing interval and the average video viewing interval in each chapter are constructed for the learner's weekly learning log. Then, the above statistics are encoded according to the log information processing method. Finally, a masked autoencoder consisting of a linear encoder and a linear decoder is constructed. The number of encoder layers and the number of neurons in each layer are the same as those in the learner portrait feature encoding construction. The length of the encoded feature vector is denoted as l p , then the input layer dimension is l p , the linear layer sizes of the two hidden layer neurons are and The rest of the network and training settings are the same as those mentioned above for the learner profile feature encoding. The trained encoder is used to embed the linked features into learner state features.

[0048] Step 2: State feature sequence construction

[0049] The generation of learner state features has a temporal sequence. In order to make full use of this information, a state feature sequence is constructed based on the generation time of the learning state. First, the learner state features are sorted according to the time sequence of their generation to construct the total sequence of learner state features. Then, for each learner state feature, the preceding l is selected. b Each state feature forms a state feature sequence with itself, so that the previous l b Weekly log fragment information assists prediction; for state features with insufficient previous state features, their previous sequences are padded with set values ​​to facilitate batch processing during training; thus, a sequence containing the state feature itself and the previous nodes is constructed for each learner state feature to facilitate data time series information mining during model training;

[0050] Step 3: State feature map construction

[0051] The log segment information is embedded into the representation space and the state feature graph is constructed based on the sample similarity to use the spatial local consistency of the log segment to perform label disambiguation and mine similar sample associations in learning. The number of learner state feature samples is denoted as n. is the state feature similarity matrix, S ij Representation status characteristics and The cosine similarity matrix between

[0052]

[0053] The graph is constructed based on cosine similarity and k-nearest neighbor method, that is, each sample point in the state feature embedding space is connected with the most similar k sample points to form an undirected graph. The adjacency matrix of the constructed graph is recorded as A n×n , then the graph construction is formally expressed as:

[0054]

[0055] in, Indicates S i The vector after the elements in are sorted in descending order, and the descending algorithm is called DSC, then The formal expression is:

[0056]

[0057] Thus, the state feature graph and its adjacency matrix representation are obtained.

[0058] A further improvement of the present invention is that, in step 3), the LSTM-GAT deep network construction specifically includes the following steps:

[0059] Step 1: LSTM network construction

[0060] After the learner state features are constructed through the state feature sequence, a learner state feature sequence based on time series is obtained. A two-layer LSTM network is constructed to learn the information in the learner state feature sequence to mine the data time series information. The network takes the state feature sequence as input and outputs the learner state time series representation; the LSTM neural unit input layer dimension is set to The hidden layer contains c lstm neurons, the output layer dimension is o lstm ;

[0061] Step 2: GAT network construction

[0062] After the learner state features are constructed through the state feature graph, a learner state feature graph based on the cosine similarity of the learner state features in the feature space is obtained. A two-layer GAT network with a multi-head attention mechanism is constructed to learn the learner state feature graph to mine the relationship information of the learner state features in the feature space. LeakyReLU is used as the activation function and the state feature graph as its input. For each node, similar node features are aggregated to obtain a learner state space representation.

[0063] Step 3: Fully connected layer group construction

[0064] Construct a fully connected layer group consisting of two fully connected layers to integrate all the encoded information to predict the abnormality of the learner state. The size of the first fully connected layer is Using tanh as the activation function, the second fully connected layer has a size of h concat ×1, using sigmoid as the activation function to output a probability representing the degree of abnormality of the learner's learning state; its input is the concatenation of the outputs of the LSTM, GAT network, and learner profile feature encoder; the feature vectors output by these three parts provide temporal information, spatial information, and general state information of the learner's state, respectively. The fully connected layer group simultaneously uses information from these three dimensions to predict the degree of abnormality of the learner's state;

[0065] The sigmoid function is formally expressed as follows:

[0066]

[0067] Among them, z is the output of the second fully connected layer, and sigmoid(z) is the probability prediction of the learner's abnormal learning state.

[0068] A further improvement of the present invention is that the vector of the l-1 layer of the GAT network The vector aggregation at level l is formally expressed as:

[0069]

[0070] in, is the sample point in the state feature graph, N i For samples The neighbor vector index list of α ij is the attention coefficient, W is the parameter matrix;

[0071] The attention coefficient calculation is formally expressed as:

[0072]

[0073] Among them, concat represents the concatenation operation, T Represents the transpose operation of the matrix, a is a weight matrix connecting two layers in a single-layer feedforward neural network;

[0074] The LeakyReLU function is formally expressed as:

[0075]

[0076] Among them, k is a hyperparameter representing the slope.

[0077] A further improvement of the present invention is that, in step 4), the LSTM-GAT deep network training based on multi-view features and label reconstruction specifically includes the following steps:

[0078] Step 1: Training data preparation

[0079] In order to predict the abnormal degree of learner status at different stages, each status feature in the semester is combined with the learner portrait feature of the current semester to form a training sample instance. The preprocessed and encoded training sample set is expressed as in Represents the learner state characteristics, for The learner profile features of the semester in which the log fragment is represented, for The noise label of the semester in which the log fragment is represented; the noise label is constructed by mapping the learner's average academic performance in the semester to [0, 1]. The set of the learner's academic performance in each subject in the kth semester is recorded as Then calculate the noise label of the kth semester The formal representation is:

[0080]

[0081] in represents the full score of the i-th subject in the k-th semester; After the sample features in are sequentially encoded, the state feature sequence is constructed, and the state feature graph is constructed, a training data set with a unified mathematical representation is obtained;

[0082] Step 2: Network parameter update

[0083] Initializing network parameters is very important for improving the training speed and convergence of deep learning networks. Therefore, the network parameters are initialized based on the Xavier initialization method to speed up the training speed and reduce gradient diffusion. Let the input dimension of the layer where the parameter w is located be , and the output dimension be o. The Xavier initialization method is formally expressed as follows:

[0084]

[0085] In order to reduce the influence of noise, pseudo labels are constructed based on noise labels during network training. At the beginning of training, the pseudo labels are initialized to noise labels. In each training iteration, the state feature map, state feature sequence, learner profile features and pseudo labels are input into the network. The loss function is set to mean square error, and the network parameters are updated based on the back propagation algorithm.

[0086] Step 3: Pseudo-label reconstruction

[0087] Since learners have different levels of abnormality at different learning stages, and the labels generated for each stage based on semester grades are the same, this introduces label noise into the training samples. Therefore, it is necessary to refine the labels during training and construct pseudo-labels that are closer to the abnormality level of the learning state to reduce label noise.

[0088] Specifically, after each iterative parameter update, the less reliable pseudo labels are updated based on the more reliable pseudo labels to reduce label noise. First, reliable samples are screened based on the temporal local continuity, spatial local consistency, and sample prediction error of state features, and their intersection is used as reliable training samples to achieve label disambiguation from the three state feature perspectives. The selected reliable sample set is formally expressed as follows:

[0089]

[0090] in and are respectively reliable sample sets selected based on temporal local continuity, spatial local consistency and sample prediction error; specifically, temporal local continuity means that the learner will have a continuous learning state in a short period of continuous time, so the continuous samples in the state feature sequence will have similar labels, so the pseudo-label difference of continuous samples on the state feature sequence is used to measure the reliability of the sample in time; spatial local consistency means that close samples in the state feature graph have similar state features, and samples with similar features have similar labels, so the pseudo-label difference of similar samples on the state feature graph is used to measure the reliability of the sample in space; at the same time, the mean square error is used to measure the error between the sample prediction and the pseudo-label and the difference between the sample pseudo-labels; then and The formal representation of a set is:

[0091]

[0092]

[0093]

[0094] Among them, τ g , τ s and τ r are three hyperparameters representing thresholds, which control the reliable sample selection based on temporal local continuity, spatial local consistency and sample prediction error respectively; is the sum of the elements in the i-th row of matrix S, is the sum of the elements in the i-th row of matrix D, x i represents the i-th training sample, y′ iIndicates that it is based on Constructed pseudo-label, SEQ i Represents sample x i The index list of the pre- and post-sequence samples on the sequence, D ij Represents sample x i and x j The number of time segments that differ in the sequence, thus obtaining a reliable sample set Used for subsequent training and pseudo-label reconstruction;

[0095] Then, after each iteration, the information of surrounding reliable samples is aggregated for the unreliable sample to update its label. The unreliable sample is selected as the center, and the adjacent neighboring samples on its state feature graph are selected to construct a subgraph. If all the neighbors of an unreliable sample are unreliable, it means that the local label centered on this sample has too much noise and little effective information, so the pseudo-label reconstruction of this sample is skipped. The pseudo-label reconstruction on the subgraph is carried out based on the label propagation method, which is formally expressed as follows:

[0096]

[0097] in Represents the noise sample x i A list of reliable samples in the neighborhood.

[0098] A further improvement of the present invention is that, in step 5), the prediction of the learner's abnormal learning state specifically includes the following steps:

[0099] The learner registration information and learner log information to be predicted are processed through steps 1) and 2) to obtain the state feature code x e , learner profile feature encoding x s And the state feature sequence SEQ as the network input, x e and its pre-order input LSTM network part, and the output of the LSTM part is used as the state feature time series representation of the sample; x e Embed the training sample into the graph space to find its k nearest neighbors, and aggregate the GAT network output of its k nearest neighbors based on distance to represent the state feature space representation of the sample. Its formal representation is:

[0100]

[0101] Among them, GAT(x) represents the output of sample x after being processed by the GAT network, N train Represents sample x eIn the neighbor set of the training set, S(x1, x2) represents the cosine similarity between samples x1 and x2. The state feature temporal representation output by the LSTM part, the state feature spatial representation output by the GAT part, and the learner portrait feature encoding are connected as the input of the fully connected layer group, and a learner state abnormality probability p is obtained through the fully connected layer group. x ∈[0, 1], p x The larger the value, the more abnormal the learner's state is, and vice versa.

[0102] The present invention has at least the following beneficial technical effects:

[0103] The present invention provides a method for predicting abnormal learning states of learners in online education. It uses learner registration information and learner log information to predict the degree of abnormality in learner states, providing a reference for teachers to provide targeted guidance and assistance to learners. By constructing different state feature perspectives based on time-space representation and implementing label reconstruction based on different state feature perspectives, the method improves on existing technologies, enabling the construction of a label noise-robust prediction model for learners' abnormal learning states without the need to label each learner state segment. Compared with existing technologies, the present invention has the following advantages:

[0104] (1) The present invention treats the prediction of learners' abnormal learning states as a regression problem, mapping learning results into continuous noisy labels to train the prediction model. Unlike the prior art that directly binarizes learning results as labels, the noisy labels constructed by the present invention can better reflect the degree of abnormality of learners' learning states at different stages in the training data. The learners' abnormal learning state prediction model trained using this data is more suitable for scenarios where teachers can provide targeted guidance to learners with different degrees of abnormality.

[0105] (2) The present invention proposes a noisy label learning method based on multi-view reliable sample disambiguation and label propagation, and constructs iteratively updated pseudo labels based on noisy labels. Specifically, the method divides reliable samples and unreliable samples based on the temporal local continuity, spatial local consistency and sample prediction error of state features, and then constructs a subgraph centered on the unreliable sample and aggregates the pseudo labels of reliable samples in the subgraph to reconstruct the pseudo labels of unreliable samples. Different from the prior art that directly uses noisy labels for supervised network training, the present invention realizes the reconstruction of pseudo labels of samples, effectively reducing the impact of label noise on network training.

[0106] (3) The present invention proposes a temporal-spatial representation extraction method, which constructs a temporal structure of state features based on the temporal sequence of learner state features, constructs a graph structure based on the similarity between learner state features, and learns its temporal-spatial representation based on LSTM and GAT, respectively. Different from the existing technology that usually mines learner profile features or temporal representations for learning, the present invention combines static learner profile features with temporal and spatial representations, mines deep multi-perspective features, and improves the accuracy of predicting learner abnormal states. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 This is the overall framework flow chart.

[0108] Figure 2 Flowchart for learner registration and log information processing.

[0109] Figure 3 Construct flow charts for log segment information processing and state feature sequences and state feature graphs.

[0110] Figure 4 Construct a flowchart for the LSTM-GAT deep network.

[0111] Figure 5 Flowchart for LSTM-GAT deep network training based on multi-view features and label reconstruction.

[0112] Figure 6 Flowchart for predicting learners' abnormal learning status.

[0113] Figure 7 Schematic diagram for predicting learners' abnormal learning status. DETAILED DESCRIPTION

[0114] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, unless there is a conflict, the embodiments of the present disclosure and the features described in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0115] Example

[0116] We selected the registration information of all learners on an online education platform in 2017 and their log information from 2017 to 2020. The present invention is further described in detail below with reference to the accompanying drawings, experimental cases, and specific implementations. All technologies implemented based on the present invention fall within the scope of this invention.

[0117] like Figure 1 As shown, in a specific implementation of the present invention, a method for predicting abnormal learning status of learners for online education includes the following steps:

[0118] Step 1. Processing learner registration information and log information

[0119] Learner registration information and log information contain many redundant fields that are irrelevant to the prediction of learner abnormal learning status and a large number of data missing fields; at the same time, there are many useful time and category fields that cannot be directly used for model training. The learner registration information and log information processing deletes the fields with low relevance to the training target in the data and constructs a unified mathematical representation for the fields that cannot be directly used for model training to facilitate subsequent feature extraction and model training. The implementation process of learner registration information and log information processing is as follows: Figure 2 , specifically including the following steps:

[0120] S101. Learner registration information processing

[0121] The learner registration information includes {sex, date of birth, identity type, highest academic qualification, institution, city, profile, and job} fields waiting to be processed.

[0122] Specifically, in this embodiment, since the personal profile and occupation fields are optional fields when registering, a large number of samples have missing corresponding fields. Furthermore, the personal profile information is not highly relevant to the target question and is difficult to use as long text information, these two fields are deleted from the learner registration information, and the remaining fields are used as features. Since date features cannot be used directly, this embodiment uses a 10-year interval as an interval and maps birthdays to discrete categorical features, with the birth date year being denoted as year. The mapping method is formally expressed as:

[0123]

[0124] Thus, yearType is obtained to represent the year category after mapping. In addition, the location is mapped to {1, 2, 3, 4, 5, 6} for first-tier cities, new first-tier cities, second-tier cities, third-tier cities, fourth-tier cities, and fifth-tier cities according to the "2020 City Business Charm Ranking" of the New First-tier City Research Institute, thereby obtaining the city category. At this point, the learner registration information is converted into category features, and then these category features are encoded based on the unique hot encoding method. In the embodiment, gender (male / female), year category, identity type (student / employed), highest degree (below junior college / junior college / undergraduate / master / doctor), graduation school (double first-class college / ordinary undergraduate / other) and city category are encoded separately, and then the codes of the above fields are connected as learner registration information features.

[0125] S102. Log Information Processing

[0126] The statistical characteristics of the log information in each semester reflect the general status of the learners during the semester, but the log information contains a large number of behavior fields and time fields.

[0127] Specifically, in this embodiment, the behavior fields include logging into the platform, watching videos, pausing videos, and commenting, and the time fields include watching video time, pausing video time, logging into the platform time, and logging out of the platform time. These fields have a large amount of data and are difficult to use directly for machine learning model training. Therefore, some aggregation functions are used to construct their statistics as features. In the embodiment, statistical data such as the total number of course video views in the current semester, the total duration of course video viewing, the number of courses participated in, the average single viewing duration, the total number of comment interactions, the average number of single video comment interactions, the number of video pauses, the average video viewing interval, and the average video viewing interval in the same chapter are constructed from the above fields as features of the log information to be processed.

[0128] Since the above statistics are numerical features with different dimensions, the numerical features are mapped based on z-score standardization. In the embodiment, the sample means {μ1, μ2, ..., μ9} and sample variances {σ1, σ2, ..., σ9} of the above 9 columns of features are first calculated, and then the z-score formula is used to standardize the numerical features. Update its features {υ1,υ2,…,υ9}, thereby converting the redundant features contained in the log information that are difficult to use directly into features with a unified mathematical representation that can be used directly. The above features are connected to form the learner statistical log features.

[0129] S103. Construction of learner profile feature coding

[0130] The combination of learner registration information and log information of a set semester can provide a macroscopic description of the learner's general status in the set semester. The learner's general learning status plays an important reference role in predicting the learner's learning status. In the embodiment, the invention combines the learner registration information and log information and encodes them to construct a learner portrait feature that can describe the learner's general learning status.

[0131] Specifically, in this embodiment, the learner portrait feature encoding construction first connects the learner registration information features and the learner statistical log features to form a vector of length 40 as the learner's initial features, and then constructs a masked autoencoder consisting of a linear encoder and a linear decoder to process sparse features. The deep associations of the learner's initial features are mined by partially masking and recovering the learner's initial features. The encoder consists of an input layer and two hidden layers. The input layer dimension is 40. The first hidden layer consists of 6 neurons and the second hidden layer consists of 4 neurons. The neurons in the first hidden layer are composed of a linear layer with a size of 40×32 and a tanh activation function, and the neurons in the second hidden layer are composed of a linear layer with a size of 32×16 and a tanh activation function. The decoder consists of two layers with the opposite structure and order to the encoder. At the same time, a batch normalization layer is set before activation to improve the convergence of the model.

[0132] In the masked autoencoder training of this embodiment, the mask rate of the feature vector is increased step by step in the order of {0, 15%, 30%, 45%} with a step size of 10 during training, and the loss function is set to the mean square error: The encoder in the trained masked autoencoder is used as the learner profile feature encoder. The 40-dimensional learner initial features are then encoded into 16-dimensional learner profile features.

[0133] Step 2. Log fragment information processing and state feature sequence and state feature graph construction

[0134] The learning state of a learner is reflected by the learning situation over a period of time. The learner log information is divided into multiple log information segments based on the time window. After preprocessing and encoding the log information segments, the learner state features reflecting the learner's learning situation over a period of time are obtained; the state features of the same learner are constructed into a state feature sequence based on the time series generated by their logs, the feature similarity between all learner state features is calculated, and a state feature graph is constructed based on the k-nearest neighbor method, so as to facilitate the mining of the temporal and spatial relationships between learner state features in the subsequent model training. The implementation process of log segment information processing and state feature sequence and state feature graph construction is as follows: Figure 3 , specifically including the following steps:

[0135] S201. Log fragment information processing

[0136] Each learner's log information is divided into log segments using the week as a time window. Next, statistical features of each week with similar log processing are constructed and encoded according to the log processing method. A masked autoencoder consisting of a linear encoder and a linear decoder is constructed to further encode the statistical features to mine their deep feature correlation information.

[0137] Specifically, in this embodiment, the number of encoder layers and the number of neurons in each layer are the same as those in the learner profile feature encoding construction, the length of the encoded feature vector is 9, the linear layer sizes of the two hidden layer neurons are 9×24 and 24×16 respectively, and the rest of the network and training settings are the same as those in the learner profile feature encoding construction mentioned above; the trained encoder is used to embed the linked features to obtain the final state features.

[0138] S202. State feature sequence construction

[0139] The generation of learner state features has a temporal sequence. In order to fully utilize this information, the present invention constructs a state feature sequence based on the generation time of the learning state.

[0140] Specifically, in this embodiment, each semester consists of 16 weeks, and the state characteristics of a learner in a certain semester are recorded. and the day of the week to which it belongs for according to right The total sequence of state features for this learner in this semester is obtained by sorting them. Then, for each learner's state feature, the five preceding state features are selected and combined with the state feature itself to form a six-state feature sequence, so that the log fragment information of the first five weeks can be used to assist prediction during training. For state features with insufficient preceding state features, their preceding sequences are padded with 0s. In this way, 16 state feature sequences are constructed for each learner in each semester to facilitate the mining of time series information in the data during the subsequent learning process.

[0141] S203. State feature map construction

[0142] Log segment information is embedded into the representation space and a state feature graph is constructed based on sample similarity to exploit the spatial local consistency of log segments for label disambiguation and mining similar sample associations in learning.

[0143] Specifically, in this embodiment, the state feature similarity matrix S is first calculated, where Representation status characteristics and Then, sort each column in the matrix S to get SDSC ,in is the sequence obtained by sorting the i-th column of S in descending order; finally, a graph is constructed based on cosine similarity and the k-nearest neighbor method. In the embodiment, k=20, that is, each sample point in the state feature embedding space is connected to the 20 most similar sample points to form an undirected graph. The constructed graph is represented by the adjacency matrix A, then The state feature graph and its adjacency matrix representation are obtained to mine spatial information in the data.

[0144] Step 3. LSTM-GAT deep network construction

[0145] Construct a deep network group consisting of parallel LTSM and GAT to extract temporal and spatial representations respectively. The inputs of the two come from the state feature sequence and state feature graph respectively. The vector output after processing by the LTSM and GAT deep network group is connected with the learner portrait feature vector and input into the fully connected layer group. The fully connected layer group outputs the probability of abnormal learning state of the learner. The implementation process of LSTM-GAT deep network construction is as follows: Figure 4 , specifically including the following steps:

[0146] S301.LSTM network construction

[0147] After the learner state features are processed in S202, a learner state feature sequence based on time series is obtained. An LSTM network is constructed to learn the sequence to mine the data time series information, thereby obtaining a learner state time series representation.

[0148] Specifically, in this embodiment, a two-layer LSTM is constructed, the LSTM neural unit input layer dimension is 16, the hidden layer contains 4 neurons, and the output layer dimension is 16.

[0149] S302.GAT network construction

[0150] After the learner state features are constructed through the state feature graph, a learner state feature graph based on the cosine similarity of the learner state features in the feature space is obtained. A two-layer GAT network with a multi-head attention mechanism is constructed to learn the state feature graph to mine the learner state space representation.

[0151] Specifically, in this embodiment, LeakyReLU is used as the activation function, and The vector of the l-1 layer of the GAT network The vector aggregation at level l is formally expressed as in, is the sample point in the state feature graph, N i For samples A list of neighbor vector indices of , is the attention coefficient, W is the parameter matrix, concat represents the concatenation operation, T Represents the transpose operation of the matrix, and a is the weight matrix connecting two layers in a single-layer feedforward neural network.

[0152] S303. Construction of fully connected layer group

[0153] A fully connected layer group consisting of two fully connected layers is constructed to integrate all encoded information to predict the degree of abnormality of the learner state. Its input is the connection of three vectors: the learner state temporal representation, the learner state spatial representation, and the learner portrait feature.

[0154] Specifically, in this embodiment, the size of the first fully connected layer is 48×16, using tanh as the activation function, and the size of the second fully connected layer is 16×1. Let the output of the fully connected layer group be z, and use As an activation function, the output of the second fully connected layer is mapped into a probability representing the degree of abnormality of the learner's learning state.

[0155] Step 4. LSTM-GAT deep network training based on multi-view features and label reconstruction

[0156] LSTM-GAT deep network training is divided into three steps: training data preparation, network parameter update, and pseudo-label reconstruction. In training data preparation, the learning results are mapped to [0, 1] as noise labels, and the learner state features and learner portrait features are used as training features. Each training iteration includes two parts: network parameter update and pseudo-label reconstruction update. The network parameters are updated during network training; when reconstructing pseudo-labels, reliable samples are first selected, and then a subgraph is constructed with unreliable samples as the center. The labels of the reliable samples in the subgraph are aggregated to reconstruct the labels of the central unreliable samples. The implementation process of LSTM-GAT deep network training is as follows: Figure 5 , specifically including the following steps:

[0157] S401. Training data preparation

[0158] In order to predict the abnormal degree of learner status at different stages, each status feature within the semester is combined with the learner portrait feature of the current semester to form a training sample instance.

[0159] Specifically, in this embodiment, the data from 2017 to 2019 are divided into a training set and a validation set in a ratio of 9:1, and the data from 2020 is used as a test set. The model is trained using the constructed training set, the model is selected using the validation set, and the model effect is tested using the test set. The specific training process is: a training sample set after preprocessing and encoding can be expressed as in represents the 16-dimensional learner state feature vector, for The 16-dimensional learner portrait features of the semester in which the log segment is represented, is a noise label mapped to [0, 1], where State characteristics The corresponding learners The corresponding semester's grade set, represents the full score of the i-th subject in the k-th semester. The sample features in the LTSM-GAT network are sequentially encoded, the state feature sequence is constructed, and the state feature graph is constructed to obtain a training set, a validation set, and a test set that can be processed in the LTSM-GAT network.

[0160] S402. Network parameter update

[0161] Each training iteration of LSTM-GAT includes two parts: network parameter update and pseudo-label reconstruction. The network parameter update part is based on the pseudo-labels obtained by reconstructing the noise labels to update the network parameters.

[0162] Specifically, in this embodiment, network parameters are first initialized using the Xavier initialization method to accelerate training and reduce gradient spread. Next, the state feature sequence is input into the LSTM network, and the state feature graph is input into the GAT network, resulting in a temporal representation of the learner state and a spatial representation of the learner state. Subsequently, the temporal representation, spatial representation, and profile features are concatenated and input into a fully connected layer group to obtain a predicted probability of abnormal learning state for the learner. Finally, the mean squared error (MSE) is calculated between the abnormal learning state probability and the pseudo-label reconstructed from the noisy label, and the network parameters are updated using the backpropagation algorithm.

[0163] S403. Pseudo-label reconstruction

[0164] Each training iteration of LSTM-GAT includes two parts: LSTM-GAT network training and pseudo-label reconstruction. The pseudo-label reconstruction part iteratively reconstructs the noisy labels based on the output of the LSTM-GAT network during training, thereby constructing pseudo-labels that are closer to the degree of abnormality in the learning state to reduce label noise.

[0165] Specifically, in this embodiment, for the sample First, the pseudo-label difference is calculated based on temporal local continuity, spatial local consistency and sample prediction error. in is the sum of the elements in the i-th row of matrix S, is the sum of the elements in the row of matrix D, y′ i Indicates based on Constructed pseudo-label, SEQ iRepresents sample x i The index list of the pre- and post-sequence samples on the sequence, D ij Represents sample x i and x j The number of time segments that differ in the sequence. This gives us a reliable sample set. For subsequent training and pseudo-label reconstruction. Then, if (e s <0.01)∧(e g <0.01)∧(e l <0.01), the sample is considered reliable, otherwise the sample is unreliable. If it is an unreliable sample, it is taken as the central sample and the adjacent samples on its state feature graph are selected to construct a subgraph, and the calculation is Update the center sample pseudo label with aggregated subgraph information.

[0166] Step 5. Prediction of learner’s abnormal learning state

[0167] After step 4, a trained learner abnormal learning state prediction model is obtained ( Figure 7 ), based on this model, the probability of abnormal learning status of the learner to be predicted can be estimated.

[0168] Specifically, in this embodiment, first, the learner registration information and learner log information to be predicted are pre-processed in step 1) and step 2) to obtain the state feature code x e , learner profile feature encoding x s And the state feature sequence SEQ( Figure 6 -S501). Then, x e and its predecessor as input, and the state feature time series representation of the sample is obtained through the LSTM network ( Figure 6 -S502). Again, calculate x e The similarity with the state characteristics of each sample in the training set is used to find its 10 nearest neighbors in the training set, and the output of the GAT network of its 10 nearest neighbors is used to represent the state feature space representation of the sample to be predicted based on the distance aggregation ( Figure 6 -S503). Finally, the state feature temporal representation, state feature spatial representation and learner profile feature are connected as the input of the fully connected layer group, and a learner state abnormality probability p is obtained through the fully connected layer group. x ( Figure 6 -S504).

[0169] It will be easily understood by those skilled in the art that the above description is merely an embodiment of the method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting abnormal learning status of learners in online education, characterized by: include: Firstly, the high-dimensional online education platform log information and learner registration information are preprocessed and encoded based on the self-supervised learning method to construct learner profile features; Secondly, based on the time window, the learner log information is segmented and encoded separately to construct the learner state features, and then the state feature sequence is constructed based on the generation time sequence of the state features, and the state feature graph is constructed based on the cosine similarity between the state features; thirdly, a long short-term memory-graph attention LSTM-GAT deep network is constructed to predict the degree of learning difficulties in online education. The number of network layers, the number of neurons in each layer, and the input and output dimensions are determined based on the constructed learner state features, state feature sequences, and state feature graphs; from this, the learning results mapped to [0, 1] are used as noise labels, and pseudo labels are constructed based on the noise labels to iteratively train the network. Each iteration first optimizes the network based on multi-view features and pseudo labels, and then selects reliable samples based on the temporal local continuity, spatial local consistency, and sample prediction error of the state features. For each unreliable sample not selected in the state feature graph, a subgraph centered on this sample is constructed. Finally, the reliable sample labels are aggregated in each subgraph to reconstruct the pseudo labels of the central samples, and the reconstructed pseudo labels are used for the next iterative training; finally, the trained network is used to predict the learner's abnormal learning state and its degree in the learning stage to be predicted; the method specifically includes the following steps: 1) Learner registration information and log information processing The learner registration information and learner log information are preprocessed as the learner's initial features. The initial features are encoded to have a unified mathematical representation. Then, a masked autoencoder is trained based on a self-supervised method. The encoded features are re-encoded based on the masked autoencoder to achieve feature dimensionality reduction and thus obtain learner profile features. 2) Log fragment information processing and state feature sequence and state feature graph construction Divide the learner's log information into multiple log information segments based on a time window, and obtain learner state features after performing preprocessing and encoding similar to the log information processing in step 1); construct a state feature sequence based on the time series generated by the state features of the same learner based on their logs, calculate the feature similarity between all learner state features, and construct a state feature graph based on the k-nearest neighbor method; 3) LSTM-GAT deep network construction A deep network group consisting of parallel LTSM and GAT is constructed to extract temporal and spatial representations respectively. The inputs of the two are from the state feature sequence and state feature graph respectively. The vector output by the LTSM and GAT deep network group is connected with the learner profile feature vector and input into the fully connected layer group. The fully connected layer group outputs the probability of abnormal learning state of the learner. 4) LSTM-GAT deep network training based on multi-view features and label reconstruction The learning results are mapped to [0, 1] as noise labels, and the learner state features and learner profile features are used as training features. Each training iteration is divided into two parts: network parameter update and pseudo-label reconstruction. During model training, the network parameters are updated based on the pseudo-labels constructed based on the noise labels. When reconstructing pseudo-labels, reliable samples are first selected, and then a subgraph is constructed with unreliable samples as the center. The pseudo-labels of the reliable samples in the subgraph are aggregated to reconstruct the pseudo-labels of the central unreliable samples. 5) Prediction of learners’ abnormal learning status The registration information and log information of the learners to be predicted are processed and encoded as the input of the trained LSTM-GAT model. After being processed by the model, the probability of the learner's learning state being abnormal is output.

2. A method for predicting abnormal learning status of learners for online education according to claim 1, characterized in that: In step 1), the learner registration information and log information processing specifically includes the following steps: Step 1: Processing learner registration information To address the problem that the original learner registration information contains many redundant and missing fields, we selected gender, date of birth, identity type, highest level of education, graduation school, and location information that are highly relevant to the learner and have few missing fields from these fields as initial features. At the same time, the date of birth retains the year of birth and is mapped into categorical features with every 10 years as an interval. The location is mapped into six categories as features: first-tier cities, new first-tier cities, second-tier cities, third-tier cities, fourth-tier cities, and fifth-tier cities based on the "2020 City Business Charm Ranking" of the New First-tier City Research Institute, thereby obtaining a unified mathematical representation of time and category features. The above features are then encoded using one-hot encoding and concatenated as learner registration information features. The one-hot encoding method is as follows: s1.Remember the category feature set as Yes, there are n i The possible values ​​of the category feature c i Set an n i Bit status register M i ; s2.M i Each bit in represents c i Is a value valid? If it is valid, it is 1; if it is invalid, it is 0. s3. For all i∈{1,2,3,…,l c }, c i Encode in sequence to obtain the one-hot vector of all category features; Step 2: Log information processing The statistical characteristics of each semester's log information reflect the learner's general status during the semester. Therefore, using sum, average, and count aggregation functions, we construct statistical data on the total number of course video views, total course video viewing duration, number of courses participated in, average single viewing duration, total number of comment interactions, average number of single video comment interactions, number of video pauses, average video viewing interval, and average video viewing interval within the same chapter from each semester's log information. The numerical features are mapped based on z-score standardization to a mean of 0 and a standard deviation of 1. This converts the redundant features contained in the log information, which are difficult to use directly, into directly usable learner statistical log features with a unified mathematical representation. The specific steps of the z-score method are: s1. The numerical feature set is Logarithmic feature v i Calculate its sample mean μ i and the sample standard deviation σ i ; s2. Set feature v i Normalization is performed according to the following z-score formula: s3. For all i∈{1,2,3,…,l v }, change v i Encode in sequence to obtain the one-hot vector of all category features; Step 3: Constructing learner profile feature encoding The learner registration information and the log information of the set semester are combined to describe the learner's general status in the set semester. The learner's general learning status plays an important reference role in predicting the learner's learning status. The learner registration information and log information are combined and encoded to construct a learner portrait feature that can describe the learner's general learning status; the learner portrait feature encoding construction first connects the features generated in Step 1 and Step 2 to form a learner initial feature to describe the learner, and then constructs a masked autoencoder composed of a linear encoder and a linear decoder to process sparse features, and mines the deep association of the learner initial feature by partially masking and restoring the learner initial feature; the encoder is set to consist of an input layer and two hidden layers, the first hidden layer is composed of neurons, and the second hidden layer consists of neurons, and the length of the learner's initial feature vector is l s , then the input layer dimension is l s , where the neurons in the first hidden layer are of size The second hidden layer consists of a linear layer and a tanh activation function, and the neurons of the second hidden layer are The decoder consists of two layers with opposite structure and order to the encoder. At the same time, a batch normalization layer is set before activation to improve the convergence of the model. The tanh activation function is formally expressed as: The feature vector is randomly masked and then input into the autoencoder network. The mask rate is set in the training sequence M with a step size of t. t Step growth, t is a hyperparameter that controls the growth rate of the mask rate, M t is the hyperparameter sequence that controls the growth rate; the loss function is set to mean square error loss, and the decoder output vector is set to The original eigenvector is x s , then the mean square error loss function is expressed as: After training, the encoder in the masked autoencoder is taken out as the learner portrait feature encoder, and all learner initial features are encoded to obtain the learner portrait features.

3. The method for predicting abnormal learning status of learners for online education according to claim 2, characterized in that: In step 2), the log segment information processing and the state feature sequence and state feature graph construction specifically include the following steps: Step 1: Log fragment information processing The learner's learning status is reflected by the learning situation within a period of time. Log fragment information processing divides the log information into log information fragments according to the set time window, and the log information in the time fragment reflects the learner's learning status in the current time fragment; Specifically, first, the log information of each learner is divided into log information segments according to the week. Secondly, similar to the log information processing, the total number of chapter video views, the total course video viewing time, the number of courses participated in, the average single viewing time, the total number of comment interactions, the average number of single video comment interactions, the number of video pauses, the average video viewing interval and the average video viewing interval in each chapter are constructed for the learner's weekly learning log. Then, the above statistics are encoded according to the log information processing method. Finally, a masked autoencoder consisting of a linear encoder and a linear decoder is constructed. The number of encoder layers and the number of neurons in each layer are the same as those in the learner portrait feature encoding construction. The length of the encoded feature vector is denoted as l p , then the input layer dimension is l p , the linear layer sizes of the two hidden layer neurons are and The rest of the network and training settings are the same as those mentioned above for the learner profile feature encoding. The trained encoder is used to embed the linked features into learner state features. Step 2: State feature sequence construction The generation of learner state features has a temporal sequence. In order to make full use of this information, a state feature sequence is constructed based on the generation time of the learning state. First, the learner state features are sorted according to the time sequence of their generation to construct the total sequence of learner state features. Then, for each learner state feature, the preceding l is selected. b Each state feature forms a state feature sequence with itself, so that the previous l b Weekly log fragment information assists prediction; for state features whose preceding state features are insufficient, their preceding state features are padded with set values ​​to facilitate batch processing during training; Thus, a sequence containing the state feature of each learner and its predecessor nodes is constructed to facilitate the mining of data time series information in model training; Step 3: State feature map construction The log segment information is embedded into the representation space and the state feature graph is constructed based on the sample similarity to use the spatial local consistency of the log segment to perform label disambiguation and mine similar sample associations in learning. The number of learner state feature samples is denoted as n. is the state feature similarity matrix, S ij Representation status characteristics and The cosine similarity matrix between The graph is constructed based on cosine similarity and k-nearest neighbor method, that is, each sample point in the state feature embedding space is connected with the most similar sample points to form an undirected graph. The adjacency matrix of the constructed graph is recorded as A n×n , then the graph construction is formally expressed as: in, Indicates S i The vector after the elements in are sorted in descending order, and the descending algorithm is called DSC, then The formal expression is: Thus, the state feature graph and its adjacency matrix representation are obtained.

4. A method for predicting abnormal learning status of learners for online education according to claim 3, characterized in that: In step 3), the LSTM-GAT deep network construction specifically includes the following steps: Step 1: LSTM network construction After the learner state features are constructed through the state feature sequence, a learner state feature sequence based on time series is obtained. A two-layer LSTM network is constructed to learn the information in the learner state feature sequence to mine the data time series information. The network takes the state feature sequence as input and outputs the learner state time series representation; the LSTM neural unit input layer dimension is set to The hidden layer contains c lstm neurons, the output layer dimension is o lstm ; Step 2: GAT network construction After the learner state features are constructed through the state feature graph, a learner state feature graph based on the cosine similarity of the learner state features in the feature space is obtained. A two-layer GAT network with a multi-head attention mechanism is constructed to learn the learner state feature graph to mine the relationship information of the learner state features in the feature space. LeakyReLU is used as the activation function and the state feature graph as its input. For each node, similar node features are aggregated to obtain a learner state space representation. Step 3: Fully connected layer group construction Construct a fully connected layer group consisting of two fully connected layers to integrate all the encoded information to predict the abnormality of the learner state. The size of the first fully connected layer is Using tanh as the activation function, the second fully connected layer has a size of h concat ×1, using sigmoid as the activation function to output a probability representing the degree of abnormality of the learner's learning state; its input is the concatenation of the outputs of the LSTM, GAT network, and learner profile feature encoder; the feature vectors output by these three parts provide temporal information, spatial information, and general state information of the learner's state, respectively. The fully connected layer group simultaneously uses information from these three dimensions to predict the degree of abnormality of the learner's state; The sigmoid function is formally expressed as follows: Among them, z is the output of the second fully connected layer, and sigmoid(z) is the probability prediction of the learner's abnormal learning state.

5. The method for predicting abnormal learning status of learners for online education according to claim 4, characterized in that: The vector of the l-1 layer of the GAT network The vector aggregation at level l is formally expressed as: in, is the sample point in the state feature graph, N i For samples The neighbor vector index list of α ij is the attention coefficient, W is the parameter matrix; The attention coefficient calculation is formally expressed as: Among them, concat represents the concatenation operation, T Represents the transpose operation of the matrix, a is a weight matrix connecting two layers in a single-layer feedforward neural network; The LeakyReLU function is formally expressed as: Among them, k is a hyperparameter representing the slope.

6. A method for predicting abnormal learning status of learners for online education according to claim 5, characterized in that: In step 4), the LSTM-GAT deep network training based on multi-view features and label reconstruction specifically includes the following steps: Step 1: Training data preparation In order to predict the abnormal degree of learner status at different stages, each status feature in the semester is combined with the learner portrait feature of the current semester to form a training sample instance. The preprocessed and encoded training sample set is expressed as in Represents the learner state characteristics, for The learner profile features of the semester in which the log fragment is represented, for The noise label of the semester in which the log fragment is represented; the noise label is constructed by mapping the learner's average academic performance in the semester to [0, 1]. The set of the learner's academic performance in each subject in the kth semester is recorded as Then calculate the noise label of the kth semester The formal representation is: in represents the full score of the i-th subject in the k-th semester; After the sample features in are sequentially encoded, the state feature sequence is constructed, and the state feature graph is constructed, a training data set with a unified mathematical representation is obtained; Step 2: Network parameter update Initializing network parameters is very important for improving the training speed and convergence of deep learning networks. Therefore, the network parameters are initialized based on the Xavier initialization method to speed up training and reduce gradient diffusion. Let the input dimension of the layer where the parameter w is located be i and the output dimension be o. The Xavier initialization method is formally expressed as follows: In order to reduce the influence of noise, pseudo labels are constructed based on noise labels during network training. At the beginning of training, the pseudo labels are initialized to noise labels. In each training iteration, the state feature map, state feature sequence, learner profile features and pseudo labels are input into the network. The loss function is set to mean square error, and the network parameters are updated based on the back propagation algorithm. Step 3: Pseudo-label reconstruction Since learners have different levels of abnormality at different learning stages, and the labels generated for each stage based on semester grades are the same, this introduces label noise into the training samples. Therefore, it is necessary to refine the labels during training and construct pseudo-labels that are closer to the abnormality level of the learning state to reduce label noise. Specifically, after each iterative parameter update, the less reliable pseudo labels are updated based on the more reliable pseudo labels to reduce label noise. First, reliable samples are screened based on the temporal local continuity, spatial local consistency, and sample prediction error of state features, and their intersection is used as reliable training samples to achieve label disambiguation from the three state feature perspectives. The selected reliable sample set is formally expressed as follows: in and are respectively reliable sample sets selected based on temporal local continuity, spatial local consistency and sample prediction error; specifically, temporal local continuity means that the learner will have a continuous learning state in a short period of continuous time, so the continuous samples in the state feature sequence will have similar labels, so the pseudo-label difference of continuous samples on the state feature sequence is used to measure the reliability of the sample in time; spatial local consistency means that close samples in the state feature graph have similar state features, and samples with similar features have similar labels, so the pseudo-label difference of similar samples on the state feature graph is used to measure the reliability of the sample in space; at the same time, the mean square error is used to measure the error between the sample prediction and the pseudo-label and the difference between the sample pseudo-labels; then and The formal representation of a set is: Among them, τ g , τ s and τ r are three hyperparameters representing thresholds, which control the reliable sample selection based on temporal local continuity, spatial local consistency and sample prediction error respectively; is the sum of the elements in the i-th row of matrix S, is the sum of the elements in the i-th row of matrix D, x i represents the i-th training sample, y′ i Indicates that it is based on Constructed pseudo label, SEQi represents sample x i The index list of the pre- and post-sequence samples on the sequence, D ij Represents sample x i and x j The number of time segments that differ in the sequence, thus obtaining a reliable sample set Used for subsequent training and pseudo-label reconstruction; Then, after each iteration, the information of surrounding reliable samples is aggregated for the unreliable sample to update its label. The unreliable sample is selected as the center, and the adjacent neighboring samples on its state feature graph are selected to construct a subgraph. If all the neighbors of an unreliable sample are unreliable, it means that the local label centered on this sample has too much noise and little effective information, so the pseudo-label reconstruction of this sample is skipped. The pseudo-label reconstruction on the subgraph is carried out based on the label propagation method, which is formally expressed as follows: in Represents the noise sample x i A list of reliable samples in the neighborhood.

7. The method for predicting abnormal learning status of learners in online education according to claim 6, characterized in that: In step 5), the prediction of the learner's abnormal learning state specifically includes the following steps: The learner registration information and learner log information to be predicted are processed through steps 1) and 2) to obtain the state feature code x e , learner profile feature encoding x s And the state feature sequence SEQ as the network input, x e and its pre-order input LSTM network part, and the output of the LSTM part is used as the state feature time series representation of the sample; x e Embed the training sample into the graph space to find its k nearest neighbors, and aggregate the GAT network output of its k nearest neighbors based on distance to represent the state feature space representation of the sample. Its formal representation is: Among them, GAT(x) represents the output of sample x after being processed by the GAT network, N train Represents sample x e In the neighbor set of the training set, S(x1, x2) represents the cosine similarity between samples x1 and x2. The state feature temporal representation output by the LSTM part, the state feature spatial representation output by the GAT part, and the learner portrait feature encoding are connected as the input of the fully connected layer group, and a learner state abnormality probability p is obtained through the fully connected layer group. x ∈[0, 1], p x The larger the value, the more abnormal the learner's state is, and vice versa.

Images