Process data score prediction device and method based on online learning platform

Abstract

The invention discloses a process data score prediction device and method based on an online learning platform. The device comprises the following contents: (1) a data collection module, (2) a data preprocessing module, (3) a time sequence data construction module, (4) a data conversion module, (5) an image fusion module, (6) an image classification module and (7) a feedback module. According to the online learning platform process data score prediction device and method provided by the invention, the relationship between the time sequence characteristics and the multi-dimensional data in the learning process data can be better captured, so that the accuracy and reliability of learning score prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intersection of artificial intelligence and education, and specifically relates to a device and method for predicting scores based on process data of an online learning platform. Background Art

[0002] With the rapid development of Internet technology, online learning has become an important way for more and more people to acquire knowledge and improve their abilities. However, due to the autonomy and flexibility of online learning, students are easily affected by various factors, resulting in poor learning results. Therefore, how to use online learning process data to predict students' academic performance, identify problems and take effective measures have become urgent issues to be solved in the field of education.

[0003] The popularity of online education platforms and systems allows students to study at any time and any place. This flexible learning method makes learning more and more autonomous, and online learning has also been widely welcomed. However, with the popularity of online learning, how to ensure the quality and efficiency of learning has become a key issue. In the traditional education model, teachers can evaluate students' learning status by observing their classroom performance, homework completion and test results, but in online learning, the direct interaction between teachers and students is reduced, which makes it more difficult to evaluate students' learning status and predict their academic performance.

[0004] To solve this problem, researchers have begun to explore the use of data generated during online learning to predict students' academic performance. These data include but are not limited to online class attendance, video viewing times and duration, forum participation, homework submissions, and test scores. By analyzing these data, students' learning habits, participation, and learning outcomes can be revealed, thereby providing teachers with a basis for evaluating students' learning status and providing students with personalized learning advice and assistance.

[0005] Existing learning performance prediction methods are mostly based on traditional machine learning techniques, such as decision trees, random forests, support vector machines, etc. These methods usually input learning process data as features into the model, and obtain a prediction model through training to predict students' learning performance. However, these methods often ignore the time series characteristics of learning process data, that is, the changing patterns of students' learning behavior over time.

[0006] In addition, existing methods usually directly input multi-dimensional learning process data into the model, ignoring the intrinsic connection and structure between the data. This makes it difficult for the model to capture the complex relationships in the data, thus affecting the accuracy and reliability of the prediction. Summary of the invention

[0007] In view of the above defects or improvement needs of the prior art, the present invention provides a device and method for predicting grades based on process data of an online learning platform, the core of which is to convert students' learning behavior data into image data, and use image processing and classification technology to predict students' academic performance; this method constructs multi-dimensional time series data, converts one-dimensional time series data into two-dimensional image data, and classifies the images to predict grades. This method can better capture the relationship between the time series characteristics in the learning process data and the multi-dimensional data, thereby improving the accuracy and reliability of academic performance prediction.

[0008] The specific technical solutions are as follows:

[0009] A device for predicting process data scores based on an online learning platform, comprising:

[0010] (1) Data collection module: collects students’ learning behavior data from the online learning platform, including online class attendance, video viewing times and duration, forum participation, homework submission and test scores, video viewing process data (including actions, video time points corresponding to actions, and specific time when actions occur), etc.

[0011] (2) Data preprocessing module: cleans and formats the collected data, removes invalid data, unifies the data format, and ensures data quality and consistency;

[0012] (3) Time series data construction module: The preprocessed data is constructed into time series data in chronological order. A set of learning behavior data of each student forms a time series. The viewing action of each video, the video time point corresponding to the action, and the specific time when the action occurs form a time series respectively.

[0013] (4) Data conversion module: converting one-dimensional time series data into two-dimensional image data by using a method of converting one-dimensional time series data into two-dimensional image data, and generating an image data set;

[0014] (5) Image fusion module: splicing multiple previously generated images representing different behavioral characteristics;

[0015] (6) Image classification module: Use the training set in the constructed image dataset to train the academic performance classification network, extract the multi-dimensional features of the data and the correlation information between the data, and combine the unique advantages of each module in the deep learning network structure to achieve accurate analysis and classification of the data and obtain the prediction results of academic performance;

[0016] (7) Feedback module: Provide students with personalized learning suggestions based on the predicted results of their academic performance.

[0017] A method for predicting scores based on process data of an online learning platform, comprising:

[0018] (1) Data collection: Collect students’ learning behavior data from the online learning platform, including online class attendance, video viewing times and duration, forum participation, homework submission and test scores, video viewing process data (including actions, video time points corresponding to actions, and specific times when actions occur), etc.

[0019] (2) Data preprocessing: cleaning and formatting the collected data, eliminating invalid data, unifying the data format, and ensuring data quality and consistency;

[0020] (3) Constructing time series data: The preprocessed data is constructed into time series data in chronological order. A set of learning behavior data of each student forms a time series.

[0021] (4) Data conversion: converting one-dimensional time series data into two-dimensional image data, using a method for converting one-dimensional time series data into two-dimensional image data, and generating an image data set; the imaging methods include but are not limited to the Gram angular field method (GAF), the Markov transition field method (MTF), and an imaging method based on pixel filling.

[0022] (5) Image fusion: stitching together multiple previously generated images representing different behavioral characteristics.

[0023] (6) Image classification: Use the training set in the constructed image dataset to train the academic performance classification network, extract the multi-dimensional features of the data and the correlation information between the data, and combine the unique advantages of each module in the deep learning network structure to achieve accurate analysis and classification of the data and obtain the prediction results of academic performance;

[0024] (7) Feedback: Provide students with personalized learning suggestions based on the predicted results of their academic performance.

[0025] Furthermore, the data collection module collects learning behavior data including the following steps:

[0026] (11) Determine the API or database interface of the online learning platform;

[0027] (12) Collecting students’ learning behavior data: Automatically collect students’ learning behavior data through the API or database interface of the online learning platform, including online class attendance, video viewing times and duration, forum participation, homework submission and test scores, and video viewing process data (including actions, video time points corresponding to the actions, and specific times when the actions occurred);

[0028] (13) Store these data in the data warehouse: Store these data in the data warehouse for subsequent data preprocessing and analysis.

[0029] Furthermore, the working process of the data preprocessing module includes the following steps:

[0030] (21) Clean the collected raw data: remove null values, outliers and duplicate data;

[0031] (22) Normalize the data: convert the data to the same scale for subsequent analysis and modeling;

[0032] (23) Fill in and process missing values ​​of time series data to ensure data integrity and continuity;

[0033] (24) Perform feature engineering on the data as needed: extract key features that help predict learning performance.

[0034] Furthermore, the time series data construction module constructs the time series data, including the following steps:

[0035] (31) constructing time series data from the preprocessed data in chronological order;

[0036] (32) Online class attendance, video viewing times and duration, forum participation, homework submission and test scores, and each video viewing process data (including actions, video time points corresponding to the actions, and specific times when the actions occurred) are processed in chronological order to form multiple learning behavior data sets. The original data sets are organized according to the tree structure of course-student-learning behavior data. For online class attendance, a time series data set is constructed in chronological order with 1 and 0 representing whether or not a student has signed in. For the number and duration of video viewing, the actions occurring at each time point are represented in numerical form. For forum participation, the number of forum posts is counted based on each class in the course, so as to subsequently construct time series data. For homework submission and test scores, the homework and test scores are arranged in chronological order, and the data are normalized, so as to subsequently construct time series data. For video viewing actions, 1, 2, 3, and 4 are defined to represent the four actions of play, pause, jump, and complete, respectively. The three feature information of video viewing actions, the video time points corresponding to the actions, and the specific time when the actions occurred are processed, and a time series data set is constructed for the three dimensional features of each video.

[0037] (33) Aggregate each student’s learning behavior data to the corresponding time granularity to form time series data.

[0038] Furthermore, the working process of the data conversion module includes the following steps:

[0039] (41) The Gram Angular Field (GAF) method is used to transform the time series data of learning behavior in six dimensions, including online class attendance, video viewing times and duration, forum participation, homework submission, and test scores, to form two 32×32×3 (W×H×C) three-channel time series images. The above two images are spliced ​​into a 64×32×3 (W×H×C) three-channel time series image 1.

[0040] (42) Extract information of three key dimensions from each video viewing process data (including action, video time point corresponding to the action, and specific time when the action occurred): fill this information into a three-channel image, where the image has three color channels, usually in RGB mode, where R represents red, G represents green, and B represents blue;

[0041] (43) Quantize and normalize the three sequences of each video viewing process data (including the action, the video time point corresponding to the action, and the specific time when the action occurred): Then quantize and normalize the three sequences so that their ranges are all within the range of 0 to 255, which is convenient for subsequent pixel filling;

[0042] (44) Padding and trimming data: In order to accurately store information without losing too much information, the three sets of sequence data are padded and trimmed to make the length of the data uniform to 64. This step ensures the alignment of all sequence data in length.

[0043] (45) Converted into a three-channel image: Finally, the three sets of sequence data of each video viewing process data (including action, video time point corresponding to the action, and specific time when the action occurs) are converted into a three-channel image by pixel filling. The image size corresponding to a single video learning data is 64×1×3 (W×H×C). A maximum of 32 video viewing time series image data are selected and spliced ​​into a 64×32×3 (W×H×C) time series image. Data less than 32 are filled with blanks. In this way, a three-channel image 2 of the video viewing process data (including action, video time point corresponding to the action, and specific time when the action occurs) is obtained.

[0044] Furthermore, the working process of the image fusion module includes the following steps:

[0045] The time series image 1 and the time series image 2 are spliced ​​together to obtain a three-channel total image with an image size of 64×64×3 (W×H×C), which includes online class attendance, video viewing times and duration, forum participation, homework submission and test scores, and video viewing process data (including actions, video time points corresponding to the actions, and the specific time when the actions occurred).

[0046] Furthermore, the working process of the image classification module includes the following steps:

[0047] (61) Use deep learning or other image classification algorithms to classify the processed images: the input image data is extracted by the residual connection module of the Self Attention Model module and the residual connection module of the Channel Attention Model module, and then input into a Transformer Model for feature extraction, and finally classified by the classifier to obtain the result of score prediction.

[0048] In the Self-Attention Model module, the first step is to generate a query, key, and value. For each pixel in the input image, a learnable weight matrix W is used. Q , W K and W V Generate the corresponding Query, Key and Value vectors respectively, and the calculation formula is shown in Formula 1:

[0049] Q=XW Q , K = XW K ,V=XW V (Formula 1)

[0050] Where X is the pixel value of the input image, Q, K, and V represent query, key, and value respectively.

[0051] The second step is to calculate the attention weight: calculate the similarity between each query vector Q and all key vectors K, and normalize it through the softmax function to obtain the attention weight. The calculation formula is shown in Formula 2.

[0052]

[0053] Where dk is the dimension of the key vector, which is used to scale the dot product to avoid the vanishing gradient problem.

[0054] The third step is output calculation: the normalized weight is multiplied by the value vector to obtain the weighted sum output, that is, the attention value. The calculation formula is shown in Formula 3.

[0055] Attention(Q,K,V)=Weight ts·V (Formula 3)

[0056] Attention(Q,K,V) can be regarded as a new representation of the input sequence, which contains the importance of each element relative to the entire sequence.

[0057] Before entering the Channel-Attention module, the data will first pass through a residual connection module. The calculation formula is shown in Formula 4.

[0058]

[0059] The Channel Attention Model module is divided into two parts: compressed channel features and feature recalibration:

[0060] The calculation formula of the compressed channel feature is shown in Formula 5, where X represents the input feature map, C1D C Represents a one-dimensional convolution with kernel size C.

[0061] z=σ(C1D C (X)) Formula 5)

[0062] Feature recalibration is usually achieved through a simple element-by-element multiplication operation, that is, the value of each channel is multiplied by the corresponding attention weight. After compressing the channel features, it is necessary to merge them with the original feature map X, that is, to directly multiply this vector with the original feature map, as shown in Formula 6.

[0063] F(z,X)=zX (Formula 6)

[0064] The data processed by the Channel-Attention module must pass through a residual connection before entering the Transformer module, as shown in Formula 7.

[0065]

[0066] Finally, after feature extraction by the Transformer Model and classification by the classifier, the prediction results are divided into five categories: excellent, good, medium, pass, and fail.

[0067] (62) Use the training data set to train the classification model until the model achieves a satisfactory accuracy;

[0068] (63) Use the validation data set to validate the model and adjust the model parameters to improve the generalization ability of the model;

[0069] (64) Use the trained model to predict new student image data and obtain the predicted results of learning performance.

[0070] Furthermore, the feedback module is implemented as follows:

[0071] (71) Providing personalized learning suggestions to students based on the predicted results of their academic performance;

[0072] (72) The predicted results of academic performance will be fed back to teachers and educational institutions so that they can better understand students’ learning status and provide more effective educational services.

[0073] The present invention provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor runs the computer program, the steps of the above-mentioned method for predicting scores based on process data of an online learning platform are executed.

[0074] The present invention provides a computer-readable storage medium storing a computer program, wherein the computer program enables the computer-readable storage medium to execute the steps of the above-mentioned method for predicting scores based on process data of an online learning platform.

[0075] The present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for predicting scores based on process data of an online learning platform.

[0076] The present invention has the following beneficial technical effects:

[0077] 1. The device and method for predicting grades based on process data of an online learning platform described in the present invention can more accurately predict students' academic performance by converting online learning process data into image data and combining image processing and classification technology. This image-based classification method can capture complex patterns and features in learning behavior data, thereby improving the accuracy and reliability of academic performance prediction.

[0078] 2. The device and method for predicting grades based on process data of an online learning platform described in the present invention adopts a multi-dimensional time series data construction method, which can comprehensively consider the changing rules of students' learning behaviors over time. By mapping learning behavior data of different dimensions into images, the correlation and trend between learning behaviors can be better observed, providing richer information for learning performance prediction.

[0079] 3. The device and method for predicting scores based on process data of an online learning platform described in the present invention provide personalized learning suggestions and guidance. By predicting students' academic performance, personalized learning suggestions can be provided according to their specific circumstances, helping students identify their own strengths and weaknesses, and making targeted adjustments and improvements to their learning plans, thereby improving learning outcomes.

[0080] 4. The device and method for predicting scores based on process data of an online learning platform described in the present invention can be widely applied to various online learning platforms and systems. Whether it is higher education, vocational education or lifelong learning, the method of the present invention can provide valuable references for educators, help improve teaching quality and learning effects, and enhance the overall level of online education.

[0081] 5. The device and method for predicting grades based on process data of an online learning platform described in the present invention have good practicality and feasibility. By collecting learning behavior data on an online learning platform and combining image processing and classification technology, the present invention provides an efficient and accurate method for predicting learning grades. At the same time, the method has good scalability and can be combined with other types of data and more advanced image processing technology to further improve the accuracy and reliability of learning grade prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the prior art description:

[0083] Figure 1 It is a flow chart of a method for predicting scores based on process data of an online learning platform according to the present invention;

[0084] Figure 2 is a flow chart of the data collection module of the present invention;

[0085] Figure 3 is a flow chart of the data preprocessing module of the present invention;

[0086] Figure 4 is a flow chart of a time series data construction module according to the present invention;

[0087] Figure 5 It is a schematic diagram of a three-channel image pixel filling method of a data conversion module according to the present invention;

[0088] Figure 6 is a flow chart of the image fusion module of the present invention;

[0089] Figure 7 It is a flow chart of the image classification module described in the present invention. DETAILED DESCRIPTION

[0090] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the following Figure 1-7 The present invention is described in further detail.

[0091] like Figure 1 As shown, the present invention provides a device for predicting process data scores based on an online learning platform, comprising:

[0092] (1) Data collection module: Figure 2 As shown in Table 1, students’ learning behavior data are automatically collected through the API or database interface of the online learning platform. These data include but are not limited to online classroom sign-in status, video viewing times and duration, forum participation, homework submission status and test scores, and video viewing process data (including actions, video time points corresponding to actions, and specific times when actions occur). The corresponding numbers are shown in Table 1.

[0093] Table 1 Categories of learning behavior data collected on online platforms

[0094]

[0095] (2) Data preprocessing module: Figure 3 As shown, check the learning behavior data of each course. If the amount of data for a course is too small, or the amount of data is large but the proportion of students who generate data is small, then remove the data of such courses from the data set. If it is found that the behavioral data of a course in the data set is far more than that of other courses, it also needs to be removed from the data set. Preprocess the collected raw data, including removing null values, outliers, and duplicate data, normalizing the data, converting the data to the same scale for subsequent analysis and modeling; and filling and processing missing values ​​for time series data. In addition, feature engineering can be performed on the data as needed to extract key features that are helpful for predicting academic performance.

[0096] (3) Time series data building module: Figure 4As shown in the figure, the preprocessed data is constructed into time series data in chronological order. The online class sign-in status, video viewing times and duration, forum participation, homework submission status and test scores, and each video viewing process data (including actions, video time points corresponding to actions, and specific times when actions occur) are processed in chronological order to form multiple learning behavior data sets. The original data sets are organized according to the tree structure of course-student-learning behavior data. For online class attendance, we construct a time series data set with 1 and 0 to represent whether or not to sign in in chronological order; for the number and duration of video viewing, we represent the actions occurring at each time point in the form of numerical values; for forum participation, we count the number of forum posts per class in the course so that we can subsequently construct time series data; for homework submission and test scores, we arrange the homework and test scores in chronological order and normalize the data so that we can subsequently construct time series data; for video viewing actions, we define 1, 2, 3, and 4 to represent the four actions of play, pause, jump, and complete, and process the three feature information of video viewing actions, the video time point corresponding to the action, and the specific time when the action occurs, and construct a time series data set for the three dimensional features of each video. Aggregate each student's learning behavior data to the corresponding time granularity to form time series data.

[0097] Table 2 Characterization of video viewing behavior categories

[0098]

[0099] (4) Data conversion module, whose working process includes the following steps:

[0100] (41) The Gram Angular Field (GAF) method is used to transform the time series data of learning behavior in six dimensions, namely, online class attendance, video viewing times and duration, forum participation, homework submission, and test scores, to form two 32×32×3 (W×H×C) three-channel time series images. The above two images are spliced ​​into a 64×32×3 (W×H×C) three-channel time series image 1.

[0101] (42) Extract information of three key dimensions from each video viewing process data (including action, video time point corresponding to the action, and specific time when the action occurred): fill this information into a three-channel image, where the image has three color channels, usually in RGB mode, where R represents red, G represents green, and B represents blue;

[0102] (43) Quantize and normalize the three sequences of each video viewing process data (including the action, the video time point corresponding to the action, and the specific time when the action occurred): Then quantize and normalize the three sequences so that their ranges are all within the range of 0 to 255, which is convenient for subsequent pixel filling;

[0103] (44) Padding and trimming data: In order to accurately store information without losing too much information, the three sets of sequence data are padded and trimmed to make the length of the data uniform to 64. This step ensures the alignment of all sequence data in length.

[0104] (45) Convert to a three-channel image: Finally, fill in the pixels, such as Figure 5 As shown, the three sets of sequence data of each video viewing process data (including action, video time point corresponding to the action, and specific time when the action occurs) are converted into a three-channel image, and the image size corresponding to the single video learning data is 64×1×3 (W×H×C). A maximum of 32 video viewing time-series image data are selected and spliced ​​into a 64×32×3 (W×H×C) time-series image, and data less than 32 are filled with blanks, so that the three-channel image 2 of the video viewing process data (including action, video time point corresponding to the action, and specific time when the action occurs) is obtained.

[0105] (5) Image fusion module, such as Figure 6 As shown, its working process includes the following steps:

[0106] The time series image 1 and the time series image 2 are spliced ​​together to obtain a three-channel total image with an image size of 64×64×3 (W×H×C), which includes online class attendance, video viewing times and duration, forum participation, homework submission and test scores, and video viewing process data (including actions, video time points corresponding to the actions, and the specific time when the actions occurred).

[0107] (6) Image classification module: Its working process includes the following steps:

[0108] (61) Use deep learning or other image classification algorithms to classify the processed images. The structure is as follows: Figure 7 As shown in the figure: the input image data is extracted by the residual connection module of the Self-Attention Model module and the residual connection module of the ChannelAttention Model module, and then input into a Transformer Model for feature extraction. Finally, it is classified by the classifier to obtain the result of grade prediction. In the Self-Attention Model module, the first step is to generate a query (Query), key (Key) and value (Value). For each pixel in the input image, a learnable weight matrix W is used. Q , W K and W VGenerate the corresponding Query, Key and Value vectors respectively, and the calculation formula is shown in Formula 1:

[0109] Q=XW Q ,K=XW K ,V=XW V (Formula 1)

[0110] Where X is the pixel value of the input image, Q, K, and V represent query, key, and value respectively.

[0111] The second step is to calculate the attention weight: calculate the similarity between each query vector Q and all key vectors K, and normalize it through the softmax function to obtain the attention weight Weigh ts. The calculation formula is shown in Formula 2.

[0112]

[0113] Where dk is the dimension of the key vector, which is used to scale the dot product to avoid the vanishing gradient problem.

[0114] The third step is output calculation: the normalized weight is multiplied by the value vector to obtain the weighted sum output, that is, the attention value. The calculation formula is shown in Formula 3.

[0115] Attention(Q,K,V)=Weights·V (Formula 3)

[0116] Attention(Q,K,V) can be regarded as a new representation of the input sequence, which contains the importance of each element relative to the entire sequence.

[0117] Before entering the Channel-Attention module, the data will first pass through a residual connection module. The calculation formula is shown in Formula 4.

[0118]

[0119] The Channel Attention Model module is divided into two parts: compressed channel features and feature recalibration:

[0120] The calculation formula of the compressed channel feature z is shown in Formula 5, where X represents the input feature map, C1D C represents a one-dimensional convolution with a kernel size of C, and σ represents the use of the Sigmoid function as the activation function of the neural network to map the variable to between [0,1].

[0121] z=σ(C1D C (X)) (Formula 5)

[0122] Feature recalibration is usually achieved through a simple element-by-element multiplication operation, that is, the value of each channel is multiplied by the corresponding attention weight. After compressing the channel features, it is necessary to merge them with the original feature map X, that is, to directly multiply this vector with the original feature map to obtain the feature recalibration result F(z,X), as shown in Formula 6.

[0123] F(z,X)=zX (Formula 6)

[0124] The data processed by the Channel-Attention module must go through a residual connection before entering the Transformer module. As shown in Formula 7:

[0125]

[0126] Finally, after feature extraction by the Transformer Model and classification by the classifier, the prediction results are divided into five categories: excellent, good, medium, pass, and fail.

[0127] The Transformer used in the present invention is composed of a linear projection, a learnable embedding, a position embedding, a Transformer encoder, and a classification head.

[0128] Specifically, in capturing the connection between video learning data, the calculation formula of each step of the Transformer used in the present invention is as follows:

[0129]

[0130] z x represents the total embedding of a layer of the Transformer Model in the present invention, and z0 represents the total embedding of the first layer in the Transformer Model. class is a learnable embedding vector, x p E represents the flattened one-dimensional sequence x P Through the fully connected layer E, it is linearly transformed from P 2 The result after reducing dimension C to dimension D, E pos It is the location embedding information.

[0131] Further explanation is as follows:

[0132] The first is linear mapping. The standard Transformer model usually accepts a one-dimensional sequence as input. In order to enable the Transformer to process two-dimensional images, we first need to transform the two-dimensional image Reshape into a flattened one-dimensional sequence Where (H, W) is the size of the original image, C is the number of channels of the original image, and (P, P) is the size of each image block, so the number of image blocks generated is N = HW / P 2 , the one-dimensional sequence length is P 2 C, which is also the number of sequences that are valid input to the Transformer module. In addition, the Transformer uses a trainable linear projection (fully connected layer FC) to transform the dimension P 2 ·C is mapped to D dimension, while keeping the number of image blocks N unchanged, as shown in Formula 8. The role of linear mapping is to flatten the one-dimensional sequence Through the fully connected layer Perform a linear transformation from P 2 ·The C dimension is reduced to the D dimension, and finally we get

[0133] Then comes the learnable embedding. In order to be consistent with the Transformer in the field of natural language processing (NLP), this module introduces the learnable embedding method. This method captures the features recognized by the network by extracting the cls token at the end and uses it as the output of the classification task. These features are fed into the MLP for the final category.

[0134] First, define an embedding z, which is the final classification result in the Transformer module, and then define Where X class is a learnable embedding vector, represents the first vector embedded in z. Specifically, in Formula 8, a classification vector is added in front of the vector of length N, which is responsible for learning information related to the category during training. The one-dimensional sequence obtained after the original image is divided into N image blocks Input to the Transformer encoder, for classification purposes, an additional and learnable embedding vector X is introduced class , and is input into the Transformer encoder together with other one-dimensional vectors. The total length of the embedding vector input into the Transformer is N+1, and the added learnable embedding vector is finally selected as the classification prediction result. This newly added category vector can be regarded as the category identification sought by N image blocks. The learnable embedding is randomly generated at the beginning of training, and then continuously adjusted and optimized through the training process.

[0135] Next, position embedding is performed. Position embedding is introduced to give each image block its relative position information in the original image to ensure that these key spatial relationships are maintained in the subsequent feature extraction process. Unlike CNN, the Transformer architecture requires position embedding to obtain the position information between image blocks. This is mainly due to the perturbation invariance of self-attention, which makes it possible to change the order of elements in the sequence without affecting the final output. The operation of this process is to embed the position information The one-dimensional sequence obtained by adding the image block is shown in Formula 8. If the model is not provided with position information about the image block, the model will have to rely on the content semantics of the image block to infer its position in the image, which will undoubtedly increase the learning difficulty of the model. There are mainly the following types of position encoding schemes: non-position embedding; 1-D position embedding (1D-PE), this method regards the two-dimensional image block as a one-dimensional sequence; 2-D position embedding (2D-PE), this method considers the two-dimensional position of the image block in the image; relative position embedding (RPE): consider the relative position of the image block. The present invention only needs to consider the order of the videos in the course, so it is possible to choose to use 1-D position embedding. The calculation method of position embedding is shown in Formulas 9 and 10. Among them, pos is the position of the image block in the figure, i corresponds to the position of embedding, d represents the length of the image block generation sequence, and the results PE calculated by the two formulas are (pos,2i) and PE (pos,2i+1) The corresponding embedding vectors of position pos in dimensions 2i and 2i+1 respectively.

[0136]

[0137] z' l =MSA(LN(z l-1 ))+z l-1 ,l=1…L (Formula 11)

[0138] z' l+1 =MLP(LN(z' l ))+z' l ,l=1…L (Formula 12)

[0139] Finally, the classification result is output through the Transformer Encoder and the classification head. The Transformer encoder is composed of multiple alternating multi-head self-attention blocks (MSA, Multi-Head Attention) and multi-layer perceptron blocks (MLP) stacked together, as shown in formulas 11 and 12, z l-1 、z' l 、z' l+1Represents the input and output of multiple alternating multi-head self-attention blocks (MSA, Multi-HeadAttention) and multi-layer perceptron blocks (MLP), and LN represents the Layer Normalization operation. The self-attention layer is the core component of the Transformer module. For the input sequence z∈R N×D For each element in , we first calculate the weighted sum of all values ​​v in the sequence, as shown in Formula 13, where zs represents the embedding after multiple alternating multi-head self-attention blocks (MSA, Multi-Head Attention) and multi-layer perceptron blocks (MLP), U qkv represents the weight matrix of q, k, v, D h Usually set to D / k. Attention weight A ij Based on calculating each query vector q i With all key vectors k j The similarity between them is expressed as shown in Formula 14. Then the weight matrix is ​​multiplied by the value vector to obtain the weighted sum output, that is, the attention value SA(z), as shown in Formula 15. Multi-head self-attention (MSA) is an extension of the self-attention mechanism, running k self-attention operations SA in parallel. k (z), each self-attention operation is called a “head”, and the outputs of these processes are merged and projected, as shown in Formula 16, [SA1(z); SA2(z); ···; SA k (z)] represents a merge operation, [SA1(z); SA2(z); ···; SA k (z)]U msa Indicates projection of the merged result.

[0140]

[0141] SA(z)=Av (Formula 15

[0142]

[0143] This module has a normalization operation before each MSA and MLP block, and introduces a residual connection after each block. In this module, MSA is followed by a feed-forward network (FFN), which consists of two fully connected layers (FC). The first FC maps the features from dimension D to 4D, and the second FC restores the features from dimension 4D back to D. The nonlinear activation function between the two is GeLU (Gaussian Error Linear Unit). After completing all the work of feature extraction, MLP Head is used for the final classification. The output of the entire Encoder is Only the front cls token is retained as the input of the full connection, which is Then the fully connected layer and the number of classifications are connected, the cross entropy loss function is used to calculate the loss, and the back propagation algorithm is used to adjust the weights and parameters.

[0144] During the implementation of the Transformer module, the input image is first flattened through linear mapping, and then the cls token is added through learnable embedding to obtain the classification result. Then, position embedding is used to prevent the loss of position information in the subsequent feature extraction. Finally, the data is input into the Transformer encoder to extract the connection between the video learning data, and finally input into the classification head to complete the classification task.

[0145] (62) Use the training data set to train the classification model until the model achieves a satisfactory accuracy;

[0146] (63) Use the validation dataset to validate the model and adjust the model parameters to improve the generalization ability of the model.

[0147] (7) Feedback module: Based on the predicted results of learning performance, it provides students with personalized learning suggestions, such as improving learning methods, adjusting learning plans, etc. At the same time, the predicted results can be fed back to teachers and educational institutions so that they can better understand the students' learning status and provide more effective educational services.

[0148] The present invention also provides a method for predicting process data scores based on an online learning platform, comprising:

[0149] (1) Data collection: Collect students’ learning behavior data from the online learning platform, including online class attendance, video viewing times and duration, forum participation, homework submission and test scores, video viewing process data (including actions, video time points corresponding to actions, and specific times when actions occur), etc.

[0150] (2) Data preprocessing: cleaning and formatting the collected data, eliminating invalid data, unifying the data format, and ensuring data quality and consistency;

[0151] (3) Constructing time series data: The preprocessed data is constructed into time series data in chronological order. A set of learning behavior data of each student forms a time series.

[0152] (4) Data conversion: converting one-dimensional time series data into two-dimensional image data, using a method for converting one-dimensional time series data into two-dimensional image data, and generating an image data set; the imaging methods include but are not limited to the Gram angular field method (GAF), the Markov transition field method (MTF), and an imaging method based on pixel filling.

[0153] (5) Image fusion: stitching together multiple previously generated images representing different behavioral characteristics.

[0154] (6) Image classification: Use the training set in the constructed image dataset to train the academic performance classification network, extract the multi-dimensional features of the data and the correlation information between the data, and combine the unique advantages of each module in the deep learning network structure to achieve accurate analysis and classification of the data and obtain the prediction results of academic performance;

[0155] (7) Feedback: Provide students with personalized learning suggestions based on the predicted results of their academic performance.

[0156] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor runs the computer program, the steps of the above-mentioned method for predicting scores based on process data of an online learning platform are executed.

[0157] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program enables the computer-readable storage medium to execute the steps of the above-mentioned method for predicting scores based on process data of an online learning platform.

[0158] The present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for predicting scores based on process data of an online learning platform.

[0159] The above description is only a preferred embodiment of the present invention and is not intended to limit the present application. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A device for predicting process data scores based on an online learning platform, characterized in that: include: (1) Data collection module: collects students’ learning behavior data from the online learning platform, including online class attendance, video viewing times and duration, forum participation, homework submission and test scores, and video viewing process data; (2) Data preprocessing module: cleans and formats the collected data, removes invalid data, unifies the data format, and ensures data quality and consistency; (3) Time series data construction module: The preprocessed data is constructed into time series data in chronological order. A set of learning behavior data of each student forms a time series. The viewing action of each video, the video time point corresponding to the action, and the specific time when the action occurs form a time series respectively. (4) Data conversion module: converting one-dimensional time series data into two-dimensional image data by using a method of converting one-dimensional time series data into two-dimensional image data, and generating an image data set; (5) Image fusion module: splicing multiple previously generated images representing different behavioral characteristics; (6) Image classification module: Use the training set in the constructed image dataset to train the academic performance classification network, extract the multi-dimensional features of the data and the correlation information between the data, and combine the unique advantages of each module in the deep learning network structure to achieve accurate analysis and classification of the data and obtain the prediction results of academic performance; (7) Feedback module: Provide students with personalized learning suggestions based on the predicted results of their academic performance.

2. The device for predicting process data scores based on an online learning platform according to claim 1, characterized in that: The data collection module collects learning behavior data including the following steps: (11) Determine the API or database interface of the online learning platform; (12) Collecting students’ learning behavior data: Automatically collect students’ learning behavior data through the API or database interface of the online learning platform, including online class attendance, video viewing times and duration, forum participation, homework submission and test scores, and video viewing process data; (13) Store these data in the data warehouse: Store these data in the data warehouse for subsequent data preprocessing and analysis.

3. The device for predicting process data scores based on an online learning platform according to claim 1, characterized in that: The working process of the data preprocessing module includes the following steps: (21) Clean the collected raw data: remove null values, outliers and duplicate data; (22) Normalize the data: convert the data to the same scale for subsequent analysis and modeling; (23) Fill in and process missing values ​​of time series data to ensure data integrity and continuity; (24) Perform feature engineering on the data as needed: extract key features that help predict learning performance.

4. The device for predicting process data scores based on an online learning platform according to claim 1, characterized in that: The time series data construction module constructs time series data, including the following steps: (31) constructing time series data from the preprocessed data in chronological order; (32) The online class attendance, video viewing times and duration, forum participation, homework submission and test scores, and each video viewing process data are processed in chronological order to form multiple learning behavior data sets. The original data sets are organized according to the tree structure of course-student-learning behavior data; For online class sign-in, a time series data set is constructed in chronological order with 1 and 0 representing whether the sign-in is present or not; For the number and duration of video viewing, the actions occurring at each time point are represented in numerical form; For forum participation, the number of forum posts was counted for each class in the course so that time series data can be constructed later; For homework submission and test scores, arrange the homework submission and test scores in chronological order and normalize the data to facilitate the subsequent construction of time series data; For video viewing actions, define 1, 2, 3, and 4 to represent the four actions of play, pause, jump, and complete, respectively. Process the three-dimensional feature information of the video viewing action, the video time point corresponding to the action, and the specific time when the action occurs, and construct a time series data set for the three-dimensional features of each video; (33) Aggregate each student’s learning behavior data to the corresponding time granularity to form time series data.

5. The device for predicting process data scores based on an online learning platform according to claim 1, characterized in that: The working process of the data conversion module includes the following steps: (41) The Gram angular field method is used to transform the time series data of learning behaviors in six dimensions, namely, online class attendance, video viewing times and duration, forum participation, homework submission, and test scores, to form two 32×32×3 three-channel time series images. The above two images are spliced ​​into a 64×32×3 three-channel time series image 1; (42) Extracting information of three key dimensions from each video viewing process data: action, video time point corresponding to the action, and specific time when the action occurred. Filling this information into a three-channel image. Three channels means that the image has three color channels, usually in RGB mode, where R represents red, G represents green, and B represents blue. (43) Quantizing and normalizing the three sequences in each video viewing process data, namely, the action, the video time point corresponding to the action, and the specific time when the action occurred. Then, quantizing and normalizing the three sequences so that their ranges are all within the range of 0 to 255, which is convenient for subsequent pixel filling. (44) Padding and trimming data: In order to accurately store information without losing too much information, the three sets of sequence data are padded and trimmed to make the length of the data uniform to 64. This step ensures the alignment of all sequence data in length. (45) Converted into a three-channel image: Finally, by pixel filling, the three sets of sequence data in each video viewing process data, including the action, the video time point corresponding to the action, and the specific time when the action occurs, are converted into a three-channel image. The image size corresponding to the single video learning data is 64×1×3. A maximum of 32 video viewing time-series image data are selected to be spliced ​​into a 64×32×3 time-series image. Data less than 32 are filled with blanks. In this way, a three-channel image 2 containing the action, the video time point corresponding to the action, and the specific time when the action occurs is obtained.

6. The device for predicting process data scores based on an online learning platform according to claim 5, characterized in that: The working process of the image fusion module includes the following steps: The time series image 1 and the time series image 2 are spliced ​​together to obtain a three-channel total image with an image size of 64×64×3, which contains online class attendance, video viewing times and duration, forum participation, homework submission and test scores, and video viewing process data.

7. The device for predicting process data scores based on an online learning platform according to claim 1, characterized in that: The image classification module works in the following steps: (61) using deep learning or other image classification algorithms to classify the processed images; (62) Use the training data set to train the classification model until the model achieves a satisfactory accuracy; (63) Use the validation data set to validate the model and adjust the model parameters to improve the generalization ability of the model; (64) Use the trained model to predict new student image data and obtain the predicted results of learning performance.

8. The device for predicting process data scores based on an online learning platform according to claim 1, characterized in that: Feedback module, its implementation steps are as follows: (71) Providing personalized learning suggestions to students based on the predicted results of their academic performance; (72) The predicted results of academic performance will be fed back to teachers and educational institutions so that they can better understand students’ learning status and provide more effective educational services.

9. A method for predicting process data scores based on an online learning platform, characterized in that: include: (1) Data collection: Collect students’ learning behavior data from the online learning platform, including online class attendance, video viewing times and duration, forum participation, homework submission and test scores, and video viewing process data; (2) Data preprocessing: cleaning and formatting the collected data, eliminating invalid data, unifying the data format, and ensuring data quality and consistency; (3) Constructing time series data: The preprocessed data is constructed into time series data in chronological order. A set of learning behavior data of each student forms a time series. (4) Data conversion; Converting one-dimensional time series data into two-dimensional image data, using a method of converting one-dimensional time series data into two-dimensional image data, and generating an image data set; (5) Image fusion: stitching together multiple previously generated images representing different behavioral characteristics; (6) Image classification: Use the training set in the constructed image dataset to train the academic performance classification network, extract the multi-dimensional features of the data and the correlation information between the data, and combine the unique advantages of each module in the deep learning network structure to achieve accurate analysis and classification of the data and obtain the prediction results of academic performance; (7) Feedback: Provide students with personalized learning suggestions based on the predicted results of their academic performance.

10. An electronic device, characterized in that: The electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the steps of the method according to claim 9 are executed when the processor runs the computer program.

Images