Anomaly detection method for MOOC datasets of large-scale online education platforms

By combining PCA dimensionality reduction and the isolated forest model with the LightGBM algorithm, the problem of identifying learner cheating behavior on MOOC platforms was solved, achieving efficient anomaly detection and supervision, and improving learning efficiency.

CN114757265BActive Publication Date: 2024-11-19BEIJING TIMES XINCHUANG TECHNOLOGY DEVELOPMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210296385.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-24
Publication Date
2024-11-19
Estimated Expiration
2042-03-24

AI Technical Summary

Technical Problem

The lack of supervision in large-scale online education platforms (MOOCs) has led to some learners cheating to complete learning tasks, making it difficult to effectively identify abnormal learners.

Method used

An unsupervised anomaly detection model of isolated forest after PCA dimensionality reduction is used to initially label the MOOC dataset. Combined with the LightGBM supervised learning algorithm, abnormal learning behaviors are identified and warned through feature extraction and model training.

Benefits of technology

It improves the accuracy and efficiency of anomaly detection, reduces the cost of manual annotation, and enhances learners' learning efficiency and course effectiveness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114757265B_ABST
    Figure CN114757265B_ABST
Patent Text Reader

Abstract

This paper proposes an anomaly detection method for MOOC datasets on a large-scale online education platform. The isolation forest algorithm solves the problem of unlabeled MOOC data, preliminarily labels the data, and provides a labeled dataset for subsequent supervised learning algorithms to train, reducing the cost of manual labeling. The supervised learning algorithm LightGBM solves the problem of poor detection results of unsupervised learning algorithms, and further improves the accuracy of anomaly detection. The model combining the unsupervised isolation forest algorithm with the supervised LightGBM algorithm improves the detection efficiency and accuracy of cheating behaviors of learners in MOOC online learning courses, and can improve the overall learning effect and efficiency of the course.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer and application technology, and in particular to an anomaly detection method for a large-scale online education platform MOOC data set. Background Art

[0002] With the rapid development of modern Internet technology and communication technology, using online teaching resources for teaching has gradually become a mainstream trend. Compared with the traditional education model, online education resources have the characteristics of convenience and replayability, which greatly improve the efficiency of teaching and reduce the cost of learning. For example, the massive online education platform (MOOC) aggregates and saves the online course resources of major universities, thereby helping more students to retrieve the required online learning courses more conveniently.

[0003] However, this new online learning model still has many drawbacks. For example, without teachers to supervise learners' behavior offline, a small number of learners often "cheat" to complete corresponding learning tasks and online exams. Possible "cheating" situations include cheating, etc., which will result in teachers not being able to understand whether learners have truly completed their learning tasks. In order to reduce the occurrence of "cheating behavior", it is necessary to discover "abnormal learners" on the MOOC platform and issue warnings to play a supervisory role, so as to improve learners' learning efficiency and ensure that teachers receive real learning feedback. Summary of the invention

[0004] In order to solve the current situation of lack of supervision in MOOC learning, the present invention proposes an anomaly detection method for MOOC datasets of large-scale online education platforms.

[0005] An anomaly detection method for a large-scale online education platform MOOC data set includes the following steps:

[0006] Step 1: Collect the learning behavior data of MOOC learners and the relevant data of learners and learning courses, then extract features from the MOOC dataset, and standardize and normalize the data. Since the data dimension is too high, we perform PCA dimensionality reduction to obtain the data feature set;

[0007] Step 2: Establish an isolation forest unsupervised anomaly detection model, detect some MOOC data, annotate the detected abnormal learners, and provide a label basis for the subsequent LightGBM model;

[0008] Step 3, using the pseudo-label data obtained in step 2, the supervised learning algorithm LightGBM is used to train the model data to discover abnormal learning behaviors in MOOCs, thereby warning abnormal learners and playing a supervisory role.

[0009] Furthermore, the specific steps of step 1 are as follows:

[0010] Step 1-1: Analyze the MOOCDATA open source dataset, build a learning behavior feature system for learners, and extract features. The features include video viewing behavior, courseware click behavior, comment interaction behavior, and user and course information.

[0011] Step 1-2: Based on the MOOC data feature set extracted in step 1-1, establish a behavior feature vector X for each learner’s learning behavior feature, x i is the i-th feature in X, 1≤i≤K, including the number of times the video is loaded x1, the number of times the video is played x2, the number of times the video is paused x3, the number of times the video is closed x4, the number of times the courseware is clicked x5, the number of times the courseware is closed x6, the number of times the comment is created x7, the number of times the comment is closed x8, the number of times the question is generated x9, and the number of times the question is checked x 10 、Video viewing time x 11 ;

[0012] In step 1-3, since the features in the extracted MOOC data feature set have different dimensions and dimensional units, this situation will affect the results of data analysis. In order to eliminate the dimensional impact between indicators, min-max standardization is performed on them. The calculation formula for standardization is:

[0013]

[0014] Among them, min is the minimum value of this feature attribute, and max is the minimum value of this feature attribute;

[0015] Step 1-4: After the data standardization in step 1-3, PCA dimensionality reduction is performed on the obtained data feature set to facilitate the subsequent training and processing of the isolation forest model. The number of features after dimensionality reduction is reduced from K to M.

[0016] Furthermore, the specific steps of step 2 are as follows:

[0017] Step 2-1, based on the MOOC data set, assuming that the number of training data sets is N and the number of features is M, construct iTree;

[0018] Step 2-2, for each X, traverse each isolated tree from the t iTrees obtained above and calculate X i The average height h(X) in the forest is normalized to the average height; the outlier score is calculated as follows:

[0019]

[0020] Where n is the size of the sample set; h(X) is the path length of sample X; c(n) is the average path length, which is used to standardize h(X) to obtain the anomaly score of instance X; H(n-1) is the harmonic number, which can be estimated as ln(n-1)+0.5772156649, where the constant is the Euler constant; E(h(X)) represents the average path length of data X in multiple iTrees.

[0021] Step 2-3, through step 2-2, obtain the anomaly score s of each sample X. When E(h(x))→c(n), s→0.5, that is, when the average path length of sample x is close to the average path length of the tree, it cannot be distinguished whether it is abnormal; when E(h(X))→0, s→1, that is, when the anomaly score of x is close to 1, it is judged to be abnormal; when E(h(X))→n-1, s→0, it is judged to be normal. According to this, part of the data in a large number of unlabeled MOOC datasets can be labeled to provide labeled data for the training of the LightGBM model.

[0022] Furthermore, in step 2-1, the process of constructing iTree is as follows:

[0023] First, n samples are randomly selected from N training data sets as sub-data sets;

[0024] Then, from X={x1,x2,…,x M Select one feature from the M feature vectors of} as the root node;

[0025] Next, a partition value p is selected between the maximum and minimum values ​​of the feature, and the sub-dataset is binary partitioned, where data smaller than the value is divided into the left child node, and data larger than the value is divided into the right child node;

[0026] Finally, recurse the left and right child nodes until there is only one instance of the leaf node or the tree reaches a limited height;

[0027] Furthermore, the specific steps of step 3 are as follows:

[0028] Step 3-1, the LightGBM model uses the Histogram algorithm to set and adjust its main parameters to achieve algorithm control and optimization: after the model passes the set and adjusted parameters, the model training starts, and cross-validation is used for training; after the training is completed, the model is saved; finally, the trained model is used for prediction;

[0029] Step 3-2, complete the judgment of abnormal MOOC learners, conduct warning supervision, and end the process.

[0030] Furthermore, in step 3-1, the adjusted parameters include, num_leaves: the number of leaves in each tree; learning_rate: learning rate; max_depth: maximum learning depth, which limits the maximum depth of the tree model. When max_depth<0, there is no limit on the learning depth; min_data: the minimum number of data in a leaf; feature_fraction: the proportion of selected features to the total number of features, with a value between 0 and 1. When feature_fraction<0, LightGBM will randomly select some features in each iteration. Feature_fraction is used to control the proportion of the total number of features selected; bagging_fraction: the proportion of selected data to the total amount of data, with a value between 0 and 1, random and non-repetitive, and used for observations of corresponding proportions.

[0031] The beneficial effects achieved by the present invention are as follows: In this method, the isolation forest unsupervised algorithm solves the problem of unlabeled MOOC data, preliminarily labels the data, provides a label data set for subsequent supervised learning algorithms to train, and reduces the cost of manual labeling; the supervised learning algorithm LightGBM solves the problem of poor detection effect of unsupervised learning algorithms, and further improves the accuracy of anomaly detection. The model combining the unsupervised isolation forest algorithm with the supervised LightGBM algorithm improves the detection efficiency and accuracy of cheating behavior of learners in MOOC online learning courses, and can improve the overall learning effect and learning efficiency of the course. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The figure is a flowchart of the steps of collecting and preprocessing MOOC related data in an embodiment of the present invention.

[0033] Figure 2 The figure is a flowchart of the steps of detecting and marking abnormal MOOC learners using the isolation forest anomaly detection model established in an embodiment of the present invention.

[0034] Figure 3 This is a MOOC abnormal learning behavior detection model built using the LightGBM algorithm in an embodiment of the present invention, and a flowchart of the steps for detecting learners with abnormal learning behavior DETAILED DESCRIPTION

[0035] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0036] The present invention discloses an anomaly detection method for MOOC data sets, which specifically includes the following three parts.

[0037] 1. Collect and preprocess MOOC related data, and perform feature extraction and selection on MOOC data sets. The specific steps are as follows:

[0038] Step 1-1: Analyze the MOOCDATA open source dataset, build a learning behavior feature system for learners, and extract features. The features include video viewing behavior, courseware clicking behavior, comment interaction behavior, and user and course information.

[0039] Step 1-2: Based on the MOOC data feature set extracted in step 1-1, establish a behavior feature vector X for each learner’s learning behavior feature, x i is the i-th feature in X, 1≤i≤K.

[0040] In step 1-3, since the features in the extracted MOOC data feature set have different dimensions and dimensional units, this situation will affect the results of data analysis. In order to eliminate the dimensional impact between indicators, min-max standardization is performed on them. The calculation formula for standardization is:

[0041]

[0042] Among them, min is the minimum value of this feature attribute, and max is the minimum value of this feature attribute;

[0043] Step 1-4: After the data standardization in step 1-3, PCA dimensionality reduction is performed on the obtained data feature set to facilitate the subsequent training and processing of the isolation forest model. The number of features after dimensionality reduction is reduced from K to M.

[0044] Second, establish an isolation forest model, detect some MOOC data, and annotate the detected abnormal learners to provide a label basis for the subsequent LightGBM model. The specific steps are as follows:

[0045] Step 1: Based on the MOOC data set, set the number of training data sets to N and the number of features to M, and construct iTree;

[0046] First, n samples are randomly selected from N training data sets as sub-data sets;

[0047] Then, from X={x1,x2,…,x M Select one feature from the M feature vectors of} as the root node;

[0048] Next, a partition value p is selected between the maximum and minimum values ​​of the feature, and the sub-dataset is binary partitioned, where data smaller than the value is divided into the left child node, and data larger than the value is divided into the right child node;

[0049] Finally, recurse the left and right child nodes until there is only one instance of the leaf node or the tree reaches a limited height;

[0050] Step 2: For each X, traverse each isolated tree from the t iTrees obtained above and calculate X i The average height h(X) in the forest is normalized to the average height; the outlier score is calculated as follows:

[0051]

[0052] Where n is the size of the sample set; h(X) is the path length of sample X; c(n) is the average path length, which is used to standardize h(X) to obtain the anomaly score of instance X; H(n-1) is the harmonic number, which can be estimated as ln(n-1)+0.5772156649, where the constant is the Euler constant; E(h(X)) represents the average path length of data X in multiple iTrees.

[0053] Step 3, in step 2, obtain the anomaly score s of each sample X. When E(h(x))→c(n), s→0.5, that is, when the average path length of sample x is close to the average path length of the tree, it cannot be distinguished whether it is abnormal; when E(h(X))→0, s→1, that is, when the anomaly score of x is close to 1, it is judged to be abnormal; when E(h(X))→n-1, s→0, it is judged to be normal. According to this, part of the data in a large number of unlabeled MOOC datasets can be labeled to provide labeled data for the training of the LightGBM model.

[0054] Third, using the pseudo-labeled dataset obtained in the second part, the LightGBM training model is trained through supervised learning to discover abnormal MOOC learners, thereby warning abnormal learners and playing a supervisory role. The specific steps are as follows:

[0055] Step 1: The labeled MOOC data D obtained above ij , where i represents the i-th sample and j represents the j-th feature. The LightGBM model is used to detect abnormal learning behaviors in MOOC.

[0056] The LightGBM model uses the Histogram algorithm, the idea of ​​which is to discretize continuous floating-point features into k discrete values ​​and construct a Histogram with a width of k. Then traverse the training data and count the cumulative statistics of each discrete value in the histogram. When performing feature selection, you only need to traverse and find the optimal split point based on the discrete values ​​of the histogram. The Histogram algorithm abandons the level-wise decision tree growth strategy used by most GBDTs now, and uses a leaf-wise strategy with depth restrictions. Level-wise data can split the leaves of the same layer at the same time, which makes it easy to perform multi-threaded optimization, control model complexity, and is not easy to overfit.

[0057] Step 2: Control and optimize the algorithm by adjusting the following key parameters:

[0058] num_leaves: the number of leaves in each tree; learning_rate: learning rate; max_depth: the maximum learning depth, which limits the maximum depth of the tree model and is used to control overfitting. When max_depth<0, there is no limit on the learning depth; min_data: the minimum number of data in a leaf, which can be used to control overfitting; feature_fraction: the proportion of selected features to the total number of features, with a value between 0 and 1. When feature_fraction<0, LightGBM will randomly select some features in each iteration. feature_fraction is used to control the proportion of selected features to the total number of features. This parameter can be used to speed up training and control overfitting; bagging_fraction: the proportion of selected data to the total amount of data, with a value between 0 and 1, which is similar to feature_fraction, but the corresponding proportion of observations are selected randomly and non-repeatedly. Note that it should be set to a ratio greater than 0. This parameter can be used to speed up training and assist in controlling overfitting.

[0059] After passing in the above parameters, we start training the model using cross validation. After training, we can save the model. Finally, we use the trained model to make predictions.

[0060] Step 3: Complete the judgment of abnormal MOOC learners, conduct warning supervision, and end the process.

[0061] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A method for detecting anomalies in a large-scale online education platform MOOC dataset, characterized by: The steps include: Step 1: Collect the learning behavior data of MOOC learners and the relevant data of learners and learning courses, extract features from the MOOC data set, standardize and normalize the data, perform PCA dimensionality reduction on it, and obtain the data feature set; Step 2: Establish an isolation forest unsupervised anomaly detection model, detect some MOOC data, annotate the detected abnormal learners, and provide a label basis for the subsequent LightGBM model; The specific steps of step 2 are as follows: Step 2-1, based on the MOOC data set, assuming that the number of training data sets is N and the number of features is M, construct iTree; Step 2-2, get t iTrees, for each X, traverse each isolated tree and calculate X i The average height h(X) in the forest is normalized to the average height; the outlier score is calculated as follows: Where n is the size of the sample set; h(X) is the path length of sample X; c(n) is the average value of the path length, which is used to standardize h(X) to obtain the anomaly score of instance X; H(n-1) is the harmonic number, which is calculated as ln(n-1)+0.5772156649, which is the Euler constant; E(h(X)) represents the average path length of data X in multiple iTrees; Step 2-3, through step 2-2, obtain the anomaly score s of each sample X. When E(h(x))→c(n), s→0.5, that is, when the average path length of sample x is close to the average path length of the tree, it cannot be distinguished whether it is abnormal; when E(h(X))→0, s→1, that is, when the anomaly score of x is close to 1, it is judged to be abnormal; when E(h(X))→n-1, s→0, it is judged to be normal; according to this, part of the data in a large number of unlabeled MOOC data sets is labeled to provide label data for the training of the LightGBM model; In step 2-1, the process of constructing iTree is as follows: First, n samples are randomly selected from N training data sets as sub-data sets; Then, from X={x1,x2,…,x M Select one feature from the M feature vectors of} as the root node; Next, a partition value p is selected between the maximum and minimum values ​​of the feature, and the sub-dataset is binary partitioned, where data smaller than the value is divided into the left child node, and data larger than the value is divided into the right child node; Finally, recurse the left and right child nodes until there is only one instance of the leaf node or the tree reaches a limited height; Step 3, using the pseudo-label data obtained in step 2, the supervised learning algorithm LightGBM is used to train the model data to discover abnormal learning behaviors in MOOCs, thereby warning abnormal learners and playing a supervisory role.

2. The method for detecting anomalies in a large-scale online education platform MOOC dataset according to claim 1, characterized in that: The specific steps of step 1 are as follows: Step 1-1: Analyze the MOOCDATA open source dataset, build a learning behavior feature system for learners, and extract features. The features include video viewing behavior, courseware clicking behavior, comment interaction behavior, and user and course information. Step 1-2: Based on the MOOC data feature set extracted in step 1-1, establish a behavior feature vector X for each learner’s learning behavior feature, x i is the i-th feature in X, 1≤i≤K; Steps 1-3, in order to eliminate the dimensional influence between features, min-max normalization is performed on them; the normalization calculation formula is: Among them, min is the minimum value of this feature attribute, and max is the minimum value of this feature attribute; Step 1-4: After the data standardization in step 1-3, PCA dimensionality reduction is performed on the obtained data feature set to facilitate the subsequent training and processing of the isolation forest model. The number of features after dimensionality reduction is reduced from K to M.

3. The method for detecting anomalies in a large-scale online education platform MOOC dataset according to claim 2, characterized in that: In step 1-2, the behavior feature vector X includes the number of times the video is loaded x1, the number of times the video is played x2, the number of times the video is paused x3, the number of times the video is closed x4, the number of times the courseware is clicked x5, the number of times the courseware is closed x6, the number of times the comment is created x7, the number of times the comment is closed x8, the number of times the question is generated x9, and the number of times the question is checked x 10 、Video viewing time x 11 .

4. The method for detecting anomalies in a large-scale online education platform MOOC dataset according to claim 1, characterized in that: The specific steps of step 3 are as follows: Step 3-1, the LightGBM model uses the Histogram algorithm to set and adjust its main parameters to achieve algorithm control and optimization: after the model passes the set and adjusted parameters, the model training starts, and cross-validation is used for training; after the training is completed, the model is saved; finally, the trained model is used for prediction; Step 3-2, complete the judgment of abnormal MOOC learners, conduct warning supervision, and end the process.

5. The method for detecting anomalies in a MOOC dataset of a large-scale online education platform according to claim 4, characterized in that: In step 3-2, the adjusted parameters include, num_leaves: the number of leaves in each tree; learning_rate: learning rate; max_depth: maximum learning depth, limiting the maximum depth of the tree model. When max_depth<0, there is no limit on the learning depth; min_data: the minimum number of data in a leaf; feature_fraction: the proportion of selected features to the total number of features, with a value between 0 and 1. When feature_fraction<0, LightGBM will randomly select some features in each iteration. Feature_fraction is used to control the proportion of the total number of features selected; bagging_fraction: the proportion of selected data to the total amount of data, with a value between 0 and 1, random and non-repetitive, used for observations of the corresponding proportion.