Online resource learning behavior analysis and prediction method based on behavior sequence coding

Abstract

The present invention relates to the field of online resource learning behavior analysis and prediction, and provides an online resource learning behavior analysis and prediction method based on behavior sequence coding. The method uniformly encodes online resource learning behaviors, facilitating the construction of long-term online resource learning behavior coding sequences. The method converts online resource learning behavior codes into online resource learning behavior sequences, performs feature dimensionality reduction on online resource learning behaviors through rank correlation analysis, and effectively improves the classification accuracy of online resource learning behavior patterns. The method transforms the online resource learning behavior prediction problem into a multivariate time series prediction problem, using the learning behavior patterns of the previous cycle as dependent variables to predict the learning behavior patterns of the next cycle.

Description

Technical Field

[0001] The present invention relates to the field of online resource learning behavior analysis and prediction, and in particular to an online resource learning behavior analysis and prediction method based on behavior sequence coding. Background Art

[0002] Analyzing students' learning behaviors and their evolution on resource platforms helps us understand their learning patterns and provide guidance for resource platform development and updates. While offline learning behavior analysis can consider multimodal data such as audio, video, and physiological data, online learning behavior analysis can only identify student learning patterns through clickstream events.

[0003] At present, the difficulties in analyzing and predicting online resource learning behavior mainly include: 1) the behavior corresponding to click events is difficult to effectively define; 2) the patterns of online resource learning behavior are difficult to effectively divide; 3) the evolution of online resource learning behavior patterns is difficult to effectively predict. Summary of the Invention

[0004] Aiming at the problems that the evolution law of student groups' online resource learning behavior is unclear and the behavior patterns are difficult to predict, the present invention proposes an online resource learning behavior analysis and prediction method based on behavior sequence coding, which can effectively improve the accuracy of predicting students' online resource learning behavior.

[0005] The purpose of the present invention is achieved through the following technical measures.

[0006] A method for analyzing and predicting online resource learning behavior based on behavior sequence coding includes the following steps:

[0007] Step 1: Encode various online learning behaviors of students on the resource platform, track students' learning trajectories in the online learning space, and use linear data structures to describe the sequence of students' learning behaviors;

[0008] Step 2: We extracted 12 dimensions of indicators from the log data accessed by the resource platform to describe students’ learning behavior, reflecting their learning enthusiasm, learning concentration, learning interest, and learning effectiveness. After obtaining the students’ time feature series, we used the rank correlation coefficient r to s Measuring the correlation between indicator dimensions and reducing the dimensionality of time feature series;

[0009] Step 3: Categorize groups based on user behavior characteristics, determine the number of clusters with higher fit, and analyze the tendencies of different user groups;

[0010] Step 4: Statistically classify the user groups within the unit period, establish an ARIMAX time series model, and predict the evolution trend of the student group classification.

[0011] In the above technical solution, step 1 includes the following sub-steps:

[0012] Step 1.1: Decompose students' operation behaviors on the resource platform into four parts: Where, What, How, and Status. "Where" refers to which page or module the user visits on the platform, including navigation bar, search box, category list, course list, or video; "What" refers to what operations the user performs on the platform page elements, including links, buttons, or progress bars; "How" refers to how the user operates on the platform, including clicks, drags, etc.; "Status" refers to the user's video viewing status or page access status, including leaving, starting, ending, pausing, fast-forwarding, etc.

[0013] In step 1.2, based on the above encoding rules, each student's single action is converted into a single action trajectory. For example, if Student A clicks the navigation bar, this can be represented by "navigation_link_click." Each student's series of actions is converted into a set of action sequence data. For example, if Student A clicks the navigation bar to enter the course list, then accesses a course through the course list and clicks the start button to start online learning, this can be represented by "navigation_link_click, course_link_click, video_button_click_start." If the student closes the window or switches pages, a "page_leave" action is generated.

[0014] In the above technical solution, step 2 includes the following sub-steps:

[0015] Step 2.1: Extract 12 dimensional indicators from the log data of resource platform access to characterize students' learning behavior, namely: C01 platform visits, C02 offline times, C03 online time, C04 maximum page stay time, C05 number of columns visited, C06 number of learning subjects, C07 number of courses studied, C08 number of course visits, C09 number of video views, C10 number of after-class learning times, C11 behavior sequence length, C12 number of IP address changes. Among them, C01 and C02 are used to reflect learning enthusiasm, C03 and C04 are used to reflect learning concentration, C05 to C09 are used to reflect learning interest and learning preference, C10 is used to reflect students' independent learning ability, C11 is used to reflect students' activity on the platform, and C12 is used to reflect changes in students' learning environment.

[0016] Step 2.2: According to the defined 12-dimensional indicators, we first construct a student behavior feature sequence containing time series features = {user ID, access date, [C01, ..., C12]}, and then use the rank correlation coefficient r s The statistical correlation between the measurement indicator dimensions is used to screen indicators based on the size of the correlation coefficient between the indicators. s The calculation formula is:

[0017]

[0018] The correlation between any two variables in the two groups is measured by comparing the rank difference between them. The larger the rank correlation coefficient value, the closer the relationship between the two groups of variables. The correlation between any two variables is ranked in ascending order. i Represents x i The rank, Q i Represents y i The rank, R i -Q i is x i 、y i The rank difference, x i =y i =[C01,…,C12]. The rank correlation coefficient range is [1,-1], r s When it is positive, it indicates positive correlation, r s When it is negative, it indicates negative correlation. The feature dimensions with large correlation coefficient values ​​(>0.5) are eliminated, and the feature dimensions with large differences are retained.

[0019] In the above technical solution, step 3 includes the following sub-steps:

[0020] Step 3.1: Data standardization: The data of each student's filtered dimension is combined into a new feature vector, and the feature vector is normalized to obtain a unit vector with a length of 1;

[0021] In step 3.2, the K-means model is used for cluster analysis. All student data within a statistical period (one month / one quarter / one academic year) are clustered. The number of clusters is set to 3, 4, 5, 6, 7, and 8 respectively. The data are sorted from large to small according to the radius of the largest cluster after clustering. The K value corresponding to the last sorting is the final number of clusters.

[0022] In the above technical solution, step 4 includes the following sub-steps:

[0023] Step 4.1: Based on the determined K value, cluster all student data within two statistical periods (one month / one quarter / one academic year) to obtain two groups of student classification time series. Use the ADF model to test the stationarity of each group of time series. One group of student classification time series is used as the training set, and the other group of student classification time series is used as the test set.

[0024] In step 4.2, an ARIMAX (P, i, Q) time series model is established. Multiple rounds of iterations are performed using the training set and the test set. The difference between the predicted value and the actual value of the classification in the test set is compared. The appropriate P and Q values ​​are searched for to obtain the classification evolution model of the student group.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The method of the present invention uniformly encodes online resource learning behaviors, which helps to construct a long-term coding sequence of online resource learning behaviors;

[0027] 2. The method of the present invention converts the online resource learning behavior code into an online resource learning behavior sequence, and performs feature dimensionality reduction on the online resource learning behavior through rank correlation analysis, thereby effectively improving the classification accuracy of the online resource learning behavior pattern;

[0028] 3. The method of the present invention transforms the online resource learning behavior prediction problem into a multivariate time series prediction problem, and uses the learning behavior pattern of the previous cycle as the dependent variable to predict the learning behavior pattern of the next cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is the user behavior coding rule diagram of the present invention.

[0030] Figure 2 This is a flowchart of an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The technical solutions of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0032] like Figure 2 As shown, an online learning behavior analysis and prediction method based on behavior sequence feature coding provided by an embodiment of the present invention mainly includes the following steps:

[0033] Step 1: Encode various online learning behaviors of students on the resource platform, track students' learning trajectories in the online learning space, and use a linear data structure to describe the sequence of students' learning behaviors;

[0034] Step 2: Extract 12 dimensions of indicators from the log data accessed by the resource platform to describe students’ learning behavior, reflecting students’ learning enthusiasm, learning concentration, learning interest, and learning effectiveness. After obtaining the students’ time feature series, use the rank correlation coefficient r s Measuring the correlation between indicator dimensions and reducing the dimensionality of time feature series;

[0035] Step 3: Categorize groups based on user behavior characteristics, determine the number of clusters with higher fit, and analyze the tendencies of different user groups;

[0036] Step 4: Statistically classify the user groups within the unit period, establish an ARIMAX time series model, and predict the evolution trend of the student group classification.

[0037] Figure 1 This is the user behavior coding rule diagram proposed by the present invention.

[0038] The step 1 specifically includes the following steps:

[0039] Step 1.1: Decompose students' operation behaviors on the resource platform into four parts: Where, What, How, and Status. "Where" refers to which page or module the user visits on the platform, including navigation bar, search box, category list, course list, or video; "What" refers to what operations the user performs on the platform page elements, including links, buttons, or progress bars; "How" refers to how the user operates on the platform, including clicks, drags, etc.; "Status" refers to the user's video viewing status or page access status, including leaving, starting, ending, pausing, fast-forwarding, etc.

[0040] In step 1.2, based on the above encoding rules, each student's single action is converted into a single action trajectory. For example, if Student A clicks the navigation bar, this can be represented by "navigation_link_click." Each student's series of actions is converted into a set of action sequence data. For example, if Student A clicks the navigation bar to enter the course list, then accesses a course through the course list and clicks the start button to start online learning, this can be represented by "navigation_link_click, course_link_click, video_button_click_start." If the student closes the window or switches pages, a "page_leave" action is generated.

[0041] The step 2 specifically includes the following steps:

[0042] Step 2.1: Extract 12 dimensional indicators from the log data of resource platform access to characterize students' learning behavior, namely: C01 platform visits, C02 offline times, C03 online time, C04 maximum page stay time, C05 number of columns visited, C06 number of learning subjects, C07 number of courses studied, C08 number of course visits, C09 number of video views, C10 number of after-class learning times, C11 behavior sequence length, C12 number of IP address changes. Among them, C01 and C02 are used to reflect learning enthusiasm, C03 and C04 are used to reflect learning concentration, C05 to C09 are used to reflect learning interest and learning preference, C10 is used to reflect students' independent learning ability, C11 is used to reflect students' activity on the platform, and C12 is used to reflect changes in students' learning environment.

[0043] Step 2.2: According to the defined 12-dimensional indicators, we first construct a student behavior feature sequence containing time series features = {user ID, access date, [C01, ..., C12]}, and then use the rank correlation coefficient r s The statistical correlation between the measurement indicator dimensions is used to reduce the feature dimension according to the size of the correlation coefficient between the indicators. s The calculation formula is:

[0044]

[0045] The correlation between any two variables in the two groups is measured by comparing the rank difference between them. The larger the rank correlation coefficient value, the closer the relationship between the two groups of variables. The correlation between any two variables is ranked in ascending order. i Represents x i The rank, Q i Represents y i The rank, R i -Q i is x i 、y i The rank difference, x i =y i =[C01,…,C12]. The rank correlation coefficient range is [1,-1], r s When it is positive, it indicates positive correlation, r s When it is negative, it indicates negative correlation. The feature dimensions with large correlation coefficient values ​​(>0.5) are eliminated, and the feature dimensions with large differences are retained.

[0046] The step 3 specifically includes the following steps:

[0047] Step 3.1: Data standardization: The data of each student's filtered dimension is combined into a new feature vector, and the feature vector is normalized to obtain a unit vector with a length of 1;

[0048] In step 3.2, the K-means model is used for cluster analysis. All student data within a statistical period (one month / one quarter / one academic year) are clustered. The number of clusters is set to 3, 4, 5, 6, 7, and 8 respectively. The data are sorted from large to small according to the radius of the largest cluster after clustering. The K value corresponding to the last sorting is the final number of clusters.

[0049] The step 4 specifically includes the following steps:

[0050] Step 4.1: Based on the determined K value, cluster all student data within two statistical periods (one month / one quarter / one academic year) to obtain two groups of student classification time series. Use the ADF model to test the stationarity of each group of time series. One group of student classification time series is used as the training set, and the other group of student classification time series is used as the test set.

[0051] In step 4.2, an ARIMAX (P, i, Q) time series model is established. Multiple rounds of iterations are performed using the training set and the test set. The difference between the predicted value and the actual value of the classification in the test set is compared. The appropriate P and Q values ​​are searched for to obtain the classification evolution model of the student group.

[0052] Although the present invention has been described in detail and illustrated with embodiments, the present invention and its applicable embodiments are not limited thereto. Persons skilled in the art may make various modifications based on the principles of the present invention and may also apply portions of the methods of the present invention to other systems. Therefore, any modifications based on the principles of the present invention should be considered to fall within the scope of protection of the present invention.

Claims

1. An online resource learning behavior analysis and prediction method based on behavior sequence coding, characterized by: The following steps are involved: Encode all kinds of online learning behaviors of students on the resource platform, track students' learning trajectories in the online learning space, and use linear data structures to describe the sequence of students' learning behaviors; From the log data of the resource platform, 12 dimensional indicators are extracted to describe students' learning behavior, reflecting their learning enthusiasm, learning concentration, learning interest and learning effectiveness. After obtaining the students' time characteristic sequence, the rank correlation coefficient r is used to s Measuring the correlation between indicator dimensions and reducing the dimensionality of time feature series; Classify groups based on the temporal feature sequence of user behavior, determine the number of clusters with higher fitting degree, and analyze the tendencies of different user groups; It includes the following sub-steps: Data standardization: The data of each student's filtered dimensions are combined into a new feature vector, and the feature vector is normalized to obtain a unit vector with a length of 1; Cluster analysis was performed using the K-means model. All student data within a statistical period were clustered, and the number of clusters was set to 3, 4, 5, 6, 7, and 8 respectively. The data were sorted from large to small according to the radius of the largest cluster after clustering. The K value corresponding to the last sorting was the final number of clusters. Statistically classify user groups within a unit period, establish an ARIMAX time series model, and predict the evolution trend of student group classification; It includes the following sub-steps: Based on the determined K value, all student data in the two statistical periods are clustered to obtain two groups of student classification time series. The ADF model is used to test the stationarity of each group of time series. One group of student classification time series is used as the training set, and the other group of student classification time series is used as the test set. An ARIMAX (P, i, Q) time series model is established, and multiple rounds of iterations are performed using the training set and test set. The differences between the predicted values ​​and actual values ​​of the classification in the test set are compared, and suitable P and Q values ​​are traversed to obtain the classification evolution model of the student group.

2. The online resource learning behavior analysis and prediction method based on behavior sequence coding according to claim 1 is characterized by: The encoding of various online learning behaviors of students on the resource platform, tracking the learning trajectory of students in the online learning space, and using a linear data structure to describe the learning behavior sequence of students include the following sub-steps: Students' operational behaviors on the resource platform are broken down into four parts: Where, What, How, and Status. "Where" refers to which page or module the user accesses on the platform, including navigation bars, search boxes, category lists, course lists, or videos; "What" refers to what operations the user performs on the platform page elements, including links, buttons, or progress bars; "How" refers to how the user operates on the platform, including clicks and drags; "Status" refers to the user's video viewing status or page access status, including leaving, starting, ending, pausing, and fast-forwarding. Based on the above coding rules, each student's operation behavior is converted into a behavior trajectory data, and each student's series of operation behaviors is converted into a set of behavior sequence data.

3. The online resource learning behavior analysis and prediction method based on behavior sequence coding according to claim 1 is characterized by: The 12 dimensional indicators are extracted from the log data accessed from the resource platform to characterize the students' learning behavior, reflecting the students' learning enthusiasm, learning concentration, learning interest and learning efficiency. After obtaining the students' time feature sequence, the rank correlation coefficient r is used to s To measure the correlation between indicator dimensions, dimensionality reduction of time feature series includes the following sub-steps: Twelve dimensional indicators are extracted from the log data of resource platform access to characterize students' learning behavior, namely: C01 number of platform visits, C02 number of offline times, C03 online time, C04 maximum page stay time, C05 number of columns visited, C06 number of learning subjects, C07 number of courses studied, C08 number of course visits, C09 number of video views, C10 number of after-class learning times, C11 length of behavior sequence, and C12 number of IP address changes. Among them, C01 and C02 are used to reflect learning enthusiasm, C03 and C04 are used to reflect learning concentration, C05 to C09, a total of 5 indicators, are used to reflect learning interests and learning preferences, C10 is used to reflect students' independent learning ability, C11 is used to reflect students' activity on the platform, and C12 is used to reflect changes in students' learning environment. According to the defined 12-dimensional indicators, we first construct a student behavior feature sequence containing time series features = {user ID, access date, [C01, ..., C12]}, and then use the rank correlation coefficient r s The statistical correlation between the measurement indicator dimensions is used to reduce the feature dimension according to the size of the correlation coefficient between the indicators; the correlation between any two variables is ranked in ascending order, R i Represents x i The rank, Q i Represents y i The rank, R i -Q i is x i 、y i The rank difference, x i =y i =[C01,…,C12], the rank correlation coefficient range is [1,-1], r s When it is positive, it indicates positive correlation, r s When it is negative, it indicates negative correlation, and feature dimensions with correlation coefficient values ​​greater than 0.5 are eliminated.

Images