An AI-based online course learning behavior monitoring system

Abstract

The application relates to the technical field of network class monitoring, in particular to a network class learning behavior monitoring system based on AI, which comprises a data collection module, a behavior monitoring module, a recommendation analysis module, a range determination module and a range optimization module. The recommendation analysis module is used for determining a network class recommendation state according to the number of abnormal behavior users and the course difference degree of the abnormal behavior users, and determining a recommendation mode according to the network class recommendation state to determine a recommendation item. The range determination module is used for determining a selection priority coefficient of each network class to be selected according to a graph correlation degree and a group mapping correlation degree, and determining a network class recommendation range based on the network class recommendation quantity corresponding to the recommendation item and the selection priority coefficient. The range optimization module is used for determining whether to perform range optimization according to the recommendation range freshness of the network class recommendation range and a negative feedback coefficient. The network class recommendation module is used for performing network class recommendation on each recommendation item based on the network class recommendation range. The application can guarantee the accuracy of behavior monitoring and the intelligent level of system recommendation in a complex network class environment.

Description

Technical Field

[0001] This invention relates to the field of online class monitoring technology, and in particular to an AI-based online class learning behavior monitoring system. Background Technology

[0002] Online education, with its advantages of flexibility in time and space and abundant resources, has become an important part of the education system, especially with the normalization of blended learning, resulting in explosive growth in user scale and the number of courses. However, in complex scenarios with dense deployment of multiple services and diverse user behavior patterns, traditional solutions suffer from problems such as insufficient integration of multi-source monitoring data, lack of learning status assessment mechanisms, and insufficient adaptive capabilities of recommendation strategies. Therefore, how to identify users with abnormal learning behaviors in complex online learning environments and achieve accurate and adaptive course recommendations to ensure the accuracy of behavior monitoring and the intelligence level of system recommendations is a technical problem that urgently needs to be solved by those skilled in the art.

[0003] Chinese Patent Publication No. CN116992142A discloses a course intelligent recommendation system and method based on user big data. The system includes: a course pre-recommendation model construction module, a pre-recommendation course scheme course experience benefit analysis module, an early warning condition value setting module, and a pre-recommendation course scheme adjustment module. The pre-recommendation course scheme course experience benefit analysis module is used to monitor the course experience benefit of the pre-recommendation course scheme for the current user in real time, and analyze the applicability rate between the current user and the pre-recommendation course scheme based on the benefit results. However, the above solution has the following problems: it cannot guarantee recommendation efficiency when the number of users is large, and it cannot achieve adaptive matching of recommendation methods and dynamic determination and optimization adjustment of recommendation range according to actual application scenarios. Summary of the Invention

[0004] To address this, the present invention provides an AI-based online learning behavior monitoring system to overcome the problems in existing technologies that cannot guarantee recommendation efficiency when there are many users, nor can they achieve adaptive matching of recommendation methods and dynamic determination and optimization of recommendation range according to actual application scenarios.

[0005] To achieve the above objectives, the present invention provides an AI-based online learning behavior monitoring system, comprising: The data acquisition module is used to obtain online class monitoring data for several users. The behavior monitoring module, which is connected to the data acquisition module, is used to input the online class monitoring data of each user into the AI ​​model to identify users with abnormal behavior and corresponding abnormal learning periods; The recommendation analysis module, which is connected to the behavior monitoring module, is used to determine the online course recommendation status based on the number of users with abnormal behavior and the course differences of users with abnormal behavior, and to determine the recommendation method and recommendation items based on the online course recommendation status. The recommendation method is either a combination recommendation based on learning isomorphism or a separate recommendation. The learning isomorphism is determined based on the operation-oriented correlation and knowledge correlation during abnormal learning periods; The range determination module, which is connected to the recommendation analysis module, is used to determine the selection priority coefficient of each online course to be selected based on the graph correlation degree and the group mapping correlation degree, and to determine the online course recommendation range based on the number of online courses recommended for each recommendation item and the selection priority coefficient; wherein, the number of online courses recommended is determined based on the state representation value and topological dispersion of the recommendation item; The range optimization module, which is connected to the range determination module, is used to determine whether to perform range optimization based on the freshness of the recommended range and the negative feedback coefficient of the online course recommendation range. When performing range optimization, it determines whether to add online courses based on the relevance of the recommended items or adjust the mapping weight coefficient based on the relevance of the online courses. The online course recommendation module is connected to the range determination module and the range optimization module respectively, and is used to recommend online courses for each recommendation item based on the online course recommendation range.

[0006] Furthermore, the recommendation analysis module determines the recommendation method as learning isomorphism-based combination recommendation for online course recommendation status where the number of users with abnormal behavior is greater than or equal to the preset number of users with abnormal behavior and the course difference degree of users with abnormal behavior is greater than or equal to the preset course difference degree. In the recommendation based on learning isomorphism, users with abnormal behavior within the current behavior monitoring period are clustered and grouped, and each group is recorded as a recommendation item. In this context, the learning isomorphism between any two users exhibiting abnormal behavior in a single group is greater than or equal to a preset behavior isomorphism, and for any user exhibiting abnormal behavior outside the group, the learning isomorphism between that user and at least one user exhibiting abnormal behavior within the group is less than the preset behavior isomorphism.

[0007] Furthermore, the recommendation analysis module determines that the recommendation method should be to make individual recommendations when the number of users with abnormal behavior is less than the preset number of users with abnormal behavior or the course difference between users with abnormal behavior is less than the preset course difference. In individual recommendations, each user exhibiting abnormal behavior within the current behavior monitoring period is recorded as a separate recommendation item.

[0008] Furthermore, the recommendation analysis module determines the learning isomorphism based on the operation-oriented relevance and knowledge relevance during abnormal learning periods; Among them, the learning isomorphism degree is positively correlated with both the operation direction correlation degree and the knowledge correlation degree. The operation direction correlation degree is determined based on the mouse feature values ​​corresponding to two abnormal learning users. The mouse feature values ​​are determined based on the ratio of the number of directional change angles corresponding to the user's mouse movement trajectory to the total number of line segment pairs during each abnormal learning period.

[0009] Furthermore, the range determination module determines the number of online courses recommended based on the state representation value and topological dispersion of the recommended items; Among them, the number of online courses recommended for a single recommendation item is positively correlated with the state representation value and topological dispersion of that recommendation item.

[0010] Furthermore, the range determination module determines the selection priority coefficient of each online course to be selected based on the graph correlation degree and the group mapping correlation degree; Among them, the selection priority coefficient of a single online course to be selected is positively correlated with both the graph correlation degree and the group mapping correlation degree.

[0011] Furthermore, the process by which the range determination module determines the group mapping correlation degree includes: For a single online course to be selected, identify the associated user corresponding to that online course; A scatter plot of each associated user is drawn based on the chapter jump dispersion of each associated user and the behavioral similarity between each associated user and a single recommendation item. The group mapping correlation degree of the online course to be selected is determined by the ratio of the clustering quality value of the scatter plot to the preset distance quality value.

[0012] Furthermore, the range optimization module determines to perform range optimization when the recommended range freshness is less than the preset recommended range freshness or the negative feedback coefficient is greater than or equal to the preset negative feedback coefficient.

[0013] Furthermore, the range optimization module responds to the condition that the correlation of the recommended item is greater than or equal to the preset correlation of the recommended item, and determines to add online courses based on the correlation of the online courses. In the process of adding online courses based on their relevance, online courses that meet the preset conditions are selected for each recommendation item in descending order of their relevance, until the required number of online courses to be added is reached. The number of additional online courses is determined based on the freshness of the recommendation range and the negative feedback coefficient. The preset condition is that none of the users with abnormal behavior in the recommendation item have studied the online course, and the online course is not within the initial online course recommendation range.

[0014] Furthermore, the range optimization module responds to the condition that the correlation tightness of the recommended item is less than the preset correlation tightness of the recommended item, and determines to increase the mapping weight coefficient. The increase in the mapping weight coefficient is negatively correlated with the correlation between the recommended item and the weight.

[0015] Compared with the prior art, the beneficial effects of the present invention are that, in the technical solution of the present invention, the online class monitoring data is input into the AI ​​model through the behavior monitoring module to identify users with abnormal behavior and corresponding abnormal learning periods. This is conducive to accurately locating the abnormal state window of the user during the learning process, providing a reliable temporal basis for the differentiated matching of subsequent recommendation strategies, and avoiding the problem of misalignment between recommendations and the user's true state caused by ignoring abnormal periods in traditional solutions.

[0016] Furthermore, this invention quantifies the number of abnormal behaviors and content dispersion of the learning group in real time by using the number of users exhibiting abnormal behavior and the degree of course discrepancy. This allows for dynamic adaptation of combined or individual recommendation methods, which is beneficial for quickly responding to changes in the abnormal state of the group, reducing the delay in recommendation decisions, achieving efficient matching between recommendation strategies and real-time teaching scenarios, and avoiding the risk of delay and mismatch in complex dynamic environments caused by a single recommendation mode.

[0017] Furthermore, this invention adaptively determines the number of recommendations by using state representation values ​​and topological dispersion, allowing groups with more severe anomalies and more dispersed knowledge to receive a wider range of recommendations. This achieves a positive match between recommendation strength and the risk of deviation, overcoming the drawbacks of a fixed number of recommendations. Simultaneously, a selection priority coefficient is constructed based on graph correlation and group mapping correlation. Candidate online courses are comprehensively evaluated from two dimensions: semantic knowledge relevance and group behavioral consensus. This ensures that the recommended content is highly relevant at the knowledge level, while also using the chapter jump dispersion and behavioral similarity of historical groups for stability verification. This significantly reduces the risk of semantically relevant but practically mismatched recommendations, thereby constructing an accurate and reliable range of online course recommendations.

[0018] Furthermore, this invention effectively reflects the novelty of the current recommended content and the user's rejection response level to the recommendation results by using the freshness and negative feedback coefficient of the online course recommendation range. Based on the freshness and negative feedback coefficient of the online course recommendation range, it determines whether to perform range optimization. This helps to trigger the optimization mechanism in a timely manner when the recommendation range becomes outdated or user negative feedback is too high, avoiding user churn and decreased learning efficiency caused by the continuous accumulation of low-quality recommendations. This enables adaptive iterative updates of the recommendation range, ensuring the long-term effectiveness of the recommendation system.

[0019] Furthermore, this invention effectively reflects the degree of topological clustering of courses learned by users within a recommended item in the knowledge graph through the association tightness of the recommended items. When the tightness is high, the group's knowledge structure converges, and adding online courses along existing knowledge associations can effectively expand the common boundaries. When the tightness is low, the knowledge distribution is discrete, and knowledge graph recommendations are prone to bias. Therefore, the mapping weight coefficient is increased, making the recommendations more reliant on the convergence pattern of historical behavior groups. This mechanism facilitates the adaptive selection of optimization paths based on the semantic distribution characteristics of the group, thereby ensuring that the recommendation system has stable adaptability and recommendation accuracy in both high and low homogeneity scenarios. Attached Figure Description

[0020] Figure 1 This is a module connection diagram of the AI-based online learning behavior monitoring system of the present invention; Figure 2 This is a flowchart illustrating how the recommendation method is determined based on the online course recommendation status in this invention. Figure 3 This is a flowchart illustrating how the present invention determines whether to perform range optimization based on the freshness of the online course recommendation range and the negative feedback coefficient. Figure 4 This is a flowchart illustrating how the present invention determines whether to add online courses or adjust the mapping weight coefficient based on the relevance of online courses according to the relevance of recommended items. Detailed Implementation

[0021] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0022] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] Please see Figures 1 to 4 As shown, this invention provides an AI-based online learning behavior monitoring system, comprising: The data acquisition module is used to obtain online class monitoring data for several users. The behavior monitoring module, which is connected to the data acquisition module, is used to input the online class monitoring data of each user into the AI ​​model to identify users with abnormal behavior and corresponding abnormal learning periods; The recommendation analysis module, which is connected to the behavior monitoring module, is used to determine the online course recommendation status based on the number of users with abnormal behavior and the course differences of users with abnormal behavior, and to determine the recommendation method and recommendation items based on the online course recommendation status. The recommendation method is either a combination recommendation based on learning isomorphism or a separate recommendation. The learning isomorphism is determined based on the operation-oriented correlation and knowledge correlation during abnormal learning periods; The range determination module, which is connected to the recommendation analysis module, is used to determine the selection priority coefficient of each online course to be selected based on the graph correlation degree and the group mapping correlation degree, and to determine the online course recommendation range based on the number of online courses recommended for each recommendation item and the selection priority coefficient; wherein, the number of online courses recommended is determined based on the state representation value and topological dispersion of the recommendation item; The range optimization module, which is connected to the range determination module, is used to determine whether to perform range optimization based on the freshness of the recommended range and the negative feedback coefficient of the online course recommendation range. When performing range optimization, it determines whether to add online courses based on the relevance of the recommended items or adjust the mapping weight coefficient based on the relevance of the online courses. The online course recommendation module is connected to the range determination module and the range optimization module respectively, and is used to recommend online courses for each recommendation item based on the online course recommendation range.

[0024] The application scenario of this invention is online learning behavior monitoring and intelligent recommendation. This invention includes several users, who are users who study on the target platform and whose study time for a single online course exceeds three-quarters of the total time of a single behavior monitoring cycle; the target platform is an online education platform, and the online courses to be selected are available online course resources stored on the platform.

[0025] The online class monitoring data consists of facial images collected in real time during the learning process. The data acquisition module uses the image acquisition device built into the user terminal, such as a computer camera, as the acquisition unit. Each user terminal corresponds to one acquisition unit. The acquisition unit is bound to the user's online class learning client. After the user starts the online class learning client and completes the identity verification, the acquisition unit automatically starts and begins to collect the user's facial images in real time.

[0026] The acquisition parameters can be adaptively configured according to the terminal's computing power, network environment, and monitoring accuracy requirements. For example, the acquisition frame rate is 1 frame / second, the image resolution is 1280×720 pixels, and the image format is JPG. Simultaneously, the data acquisition module has a built-in preprocessing unit that performs real-time preprocessing on the acquired facial images. The preprocessing steps include: image grayscale processing to reduce image noise; face detection and cropping, locating the face region using a Haar feature classifier, cropping the background region, and retaining the key facial regions, which are from the forehead to the chin and from the left and right sides to the cheeks; image normalization processing, scaling the cropped face image to a uniform size, such as 224×224 pixels, and normalizing the pixel values ​​to the [0,1] range to eliminate the impact of image size and brightness differences on subsequent recognition; and time-series marking, adding a timestamp and a unique user identifier to each frame of the facial image, including the user's student ID and account ID, forming a time-series facial image sequence as input data for subsequent AI models.

[0027] The training process of an AI model consists of the following four steps: Step 1: Construct a real-world facial behavior dataset for online classes. Collect normal and abnormal behavior samples covering different scenarios, including varying lighting conditions, user postures, and degrees of facial occlusion. Abnormal behavior samples include, but are not limited to: gaze deviating from the screen, eyes closed while dozing off, excessive head tilting, and abnormal facial expressions. Manually label the collected samples, marking normal learning behavior as 0 and abnormal learning behavior as 1. After labeling, clean the data, removing invalid samples such as blurry, overexposed, or completely occluded samples. Finally, divide the dataset into training, validation, and test sets in a 7:2:1 ratio.

[0028] Step 2: Build a lightweight fusion recognition network. Use MobileNetV3 as the backbone network to build a lightweight fusion recognition model.

[0029] Step 3: Iterative Training. A hierarchical learning rate strategy is used to set different initial learning rates for different layers of the network. The initial learning rate for the backbone layer is set to 0.0001, and the initial learning rate for newly added branch layers is also set to 0.001. The optimizer is Adam, the batch size is set to 32, and the training epochs are set to 50. Training stops when the validation set loss does not decrease for 5 consecutive epochs. Simultaneously, when the validation set accuracy does not improve for 3 consecutive epochs, the learning rate is decayed to 0.5 times the current value.

[0030] Step 4: Accuracy Verification and Model Deployment. The trained model is then validated using a test set. Performance requirements are: accuracy ≥ 95% for normal learning behavior recognition and ≥ 92% for abnormal behavior recognition. If these requirements are not met, additional samples are added for retraining until the accuracy metrics are satisfied.

[0031] The behavior monitoring module calls the deployed AI model to infer the facial image sequence of each user. For each time window, the window length is 60 frames, corresponding to 1 minute, and the model outputs anomaly probability p. When p ≥ 0.7, the current window is determined to be in an abnormal state; when more than 3 consecutive windows are determined to be in an abnormal state, the continuous period is marked as an abnormal learning period, and the corresponding user is marked as an abnormal behavior user.

[0032] This invention sets up a continuous cyclical behavior monitoring cycle. At the end of each behavior monitoring cycle, an AI model is used to identify users with abnormal behavior and corresponding abnormal learning periods by analyzing the online learning monitoring data of users who are currently taking online classes. A value for the duration of the behavior monitoring cycle is provided, which is 20 minutes in this embodiment.

[0033] This invention utilizes an online course knowledge graph, which is constructed by extracting the knowledge content of all online courses on a target platform. For each online course on the target platform, its course outline, chapter divisions, knowledge point tags, and teaching video subtitles are analyzed one by one, along with multi-source heterogeneous information, and text cleaning and normalization preprocessing are performed. Then, a large language model is used to identify and extract the core knowledge points involved in each course as entities from the preprocessed text. Next, logical relationships between knowledge points are defined, including sequential preconditions, parent-child inclusion relationships, and associative reference relationships. Afterwards, the extracted entities and relationships are merged and aligned, eliminating duplication and conflicts, and then stored in a graph database. Knowledge points are used as nodes, and the semantic relationships between knowledge points are used as edges to obtain the online course knowledge graph. This is a common technique used by those skilled in the art, and will not be elaborated further.

[0034] Specifically, the recommendation analysis module determines the recommendation method as a combination recommendation based on learning isomorphism when the number of users with abnormal behavior is greater than or equal to the preset number of users with abnormal behavior and the course difference between users with abnormal behavior is greater than or equal to the preset course difference. In the recommendation based on learning isomorphism, users with abnormal behavior within the current behavior monitoring period are clustered and grouped, and each group is recorded as a recommendation item. In this context, the learning isomorphism between any two users exhibiting abnormal behavior in a single group is greater than or equal to a preset behavior isomorphism, and for any user exhibiting abnormal behavior outside the group, the learning isomorphism between that user and at least one user exhibiting abnormal behavior within the group is less than the preset behavior isomorphism.

[0035] The number of users exhibiting abnormal behavior refers to the number of users exhibiting abnormal behavior during the current behavior monitoring period. The method for confirming course differences is as follows: The online courses that users with abnormal behavior are learning during the current behavior monitoring period are recorded as the analyzed online courses, and the average of the difference coefficients corresponding to each analyzed online course is recorded as the course difference degree. The difference coefficient corresponding to a single online course is the average of the sub-difference coefficients corresponding to that online course and the other online courses. For any two online courses, the first entity set and the second entity set are determined respectively, consisting of all the knowledge point entities involved in the online course knowledge graph. The sub-difference coefficient = 1 - the number of entities contained in the intersection of the two sets / the number of entities contained in the union of the two sets.

[0036] The preset values ​​for the number of users exhibiting abnormal behavior and the preset course difference can be determined based on the platform's current computing resources and recommendation response time requirements. The number of users exhibiting abnormal behavior and the course difference effectively reflect the necessity and urgency of triggering the homogeneity combination recommendation mode. The greater the user's need for rapid convergence of learning behavior and reduction of recommendation complexity, the smaller the preset values ​​for the number of users exhibiting abnormal behavior and the preset course difference. In this embodiment, the preset number of users exhibiting abnormal behavior is 50, and the preset course difference is 0.7.

[0037] The preset behavioral isomorphism value can be determined based on the platform's requirements for recommendation accuracy and the degree of consistency of user group behavior. Behavioral isomorphism effectively reflects the similarity of learning patterns among users with abnormal behavior. The greater the demand for interpretability of recommendation results and the tightness of group aggregation, the larger the preset behavioral isomorphism value. In this embodiment, the preset behavioral isomorphism value is 0.75.

[0038] Specifically, the recommendation analysis module determines that the recommendation method should be to make individual recommendations when the number of users with abnormal behavior is less than the preset number of users with abnormal behavior or the course difference between users with abnormal behavior is less than the preset course difference. In individual recommendations, each user exhibiting abnormal behavior within the current behavior monitoring period is recorded as a separate recommendation item.

[0039] Specifically, the recommendation analysis module determines the learning isomorphism based on the operation-oriented relevance and knowledge relevance during abnormal learning periods; Among them, the learning isomorphism degree is positively correlated with both the operation direction correlation degree and the knowledge correlation degree. The operation direction correlation degree is determined based on the mouse feature values ​​corresponding to two abnormal learning users. The mouse feature values ​​are determined based on the ratio of the number of directional change angles corresponding to the user's mouse movement trajectory to the total number of line segment pairs during each abnormal learning period.

[0040] Learning isomorphism = operation orientation correlation × first weight coefficient + knowledge correlation × second weight coefficient; in this embodiment, the first weight coefficient and the second weight coefficient are preferably 0.5, but the above values ​​are not limited to this. Users can adaptively adjust the first weight coefficient and the second weight coefficient according to the actual application scenario, course type characteristics, credibility of behavioral data, or the system's emphasis on behavioral and knowledge dimensions.

[0041] The correlation degree of operation points corresponding to any two abnormal learning users = 1 - the absolute value of the difference between the mouse feature values ​​corresponding to the two abnormal learning users / (the larger value of the mouse feature values ​​corresponding to the two abnormal learning users + ε), where ε is a minimum value to avoid the denominator being 0, and ε is 0.001; The mouse feature value corresponding to a single abnormal learning user is the average of the sub-feature values ​​of each abnormal learning period corresponding to that abnormal learning user; For each abnormal learning period, the user's mouse movement trajectory is collected during that period. Each sampling point in the mouse movement trajectory is recorded as a trajectory point. The sampling points are continuously collected at a fixed sampling frequency, preferably 30 times per second. Each trajectory point includes at least a timestamp, a horizontal coordinate of the screen, and a vertical coordinate of the screen. The line segment pairs formed by every three consecutive points in the trajectory point sequence are traversed sequentially, and the turning angle between the previous line segment and the next line segment is calculated. When the absolute value of the turning angle exceeds a preset threshold, the turning angle is recorded as a direction change angle. The preset threshold is preferably 30°. The sub-feature value corresponding to a single abnormal learning period is equal to the number of direction change angles divided by the total number of line segment pairs. The total number of line segment pairs is the total number of trajectory points minus 2.

[0042] The method for confirming knowledge relevance includes: for any two abnormal learning users, based on the start and end times of each abnormal learning period, locating them on the video timeline of the online courses they are learning, extracting the corresponding teaching video subtitle text within that period, using a pre-trained large language model to extract knowledge point entities from the extracted text, obtaining all knowledge point entities involved by each user within the corresponding abnormal learning period, and recording them as abnormal knowledge point entities, and determining two entity sets formed by the abnormal knowledge point entities corresponding to the two abnormal learning users, and recording the ratio of the number of entities contained in the intersection of the two sets to the number of entities contained in the union of the two sets as the knowledge relevance.

[0043] Specifically, the range determination module determines the number of online courses recommended based on the state representation value and topological dispersion of the recommended items; Among them, the number of online courses recommended for a single recommendation item is positively correlated with the state representation value and topological dispersion of that recommendation item.

[0044] Among them, the state representation value is used to represent the severity of the overall abnormal learning of users within the recommended item; the state representation value corresponding to a single recommended item is the average of the abnormal duration percentages of each abnormal behavior user in that recommended item, and the abnormal duration percentage of a single abnormal behavior user is the ratio of the total duration of all abnormal learning periods of that abnormal behavior user in the current behavior monitoring period to the total duration of that abnormal behavior user learning the online course in the current behavior monitoring period. Topological dispersion is used to characterize the degree of dispersion of knowledge points of abnormal user groups within the same recommendation item. The larger the value, the greater the difference in the knowledge structure of the group. The topological dispersion is determined as follows: for a single recommendation item, the abnormal knowledge point entities corresponding to each abnormal user of the recommendation item are recorded as reference entities. The topological dispersion of the recommendation item is 1 - the number of duplicate abnormal knowledge point entities corresponding to each abnormal user of the recommendation item / the total number of entities in the minimum subgraph. The minimum subgraph is the connected subgraph that can contain all abnormal knowledge point entities corresponding to each abnormal user of the recommendation item and has the fewest edges.

[0045] The number of online courses recommended for a single recommendation item = [(state representation value / preset state representation value) × third weight coefficient + (topological dispersion / preset topological dispersion) × fourth weight coefficient] × online course recommendation number threshold, where the online course recommendation number threshold is 5, and both the third and fourth weight coefficients are 0.5; The preset state representation value and preset topological dispersion can be determined according to the actual application scenario. These values ​​effectively reflect the severity of abnormal learning and the dispersion of knowledge structure among the user group within the current recommendation item. The greater the user's demand for precise and intensive recommendations, the smaller the preset state representation value and preset topological dispersion should be. In this embodiment, the preset state representation value is 0.4, and the preset topological dispersion is 0.6.

[0046] It should be noted that the minimum number of online course recommendations for a single recommendation item is one. When the calculated result is less than 1, the value is set to 1.

[0047] Specifically, the range determination module determines the selection priority coefficient of each online course to be selected based on the graph correlation degree and the group mapping correlation degree; Among them, the selection priority coefficient of a single online course to be selected is positively correlated with both the graph correlation degree and the group mapping correlation degree.

[0048] For a single recommended item and a single online course to be selected, the method for confirming the graph correlation is as follows: the abnormal knowledge point entities corresponding to each abnormal behavior user of the recommended item are recorded as Class I entities, and the knowledge point entities of the online course to be selected are recorded as Class II entities. Graph correlation degree = Number of entities in type I and type II that are the same / Number of type I entities × Fifth weight coefficient + Number of type II entities that are different from type I entities but have direct relationship edges / (Number of type II entities that are different from type I entities + ε) × Sixth weight coefficient, where the fifth weight coefficient is 0.7 and the sixth weight coefficient is 0.3. The selection priority coefficient of a single online course to be selected = graph correlation degree / preset graph correlation degree × graph weight coefficient + group mapping correlation degree / preset group mapping correlation degree × mapping weight coefficient, graph weight coefficient + mapping weight coefficient = 1, in this embodiment the graph weight coefficient is 0.6 and the mapping weight coefficient is 0.4; The values ​​of the preset graph correlation degree and the preset group mapping correlation degree can be determined according to the actual application scenario. The graph correlation degree effectively reflects the overlap density between the online courses to be selected and the recommended items at the knowledge semantic level, while the group mapping correlation degree effectively reflects the convergence degree between the online courses to be selected and the recommended items at the group behavior level. The greater the user's demand for low false positives and high reliability recommendations, the higher the threshold for judging strong correlation is required, that is, the larger the values ​​of the preset graph correlation degree and the preset group mapping correlation degree. In this embodiment, the preset graph correlation degree is 0.6 and the preset group mapping correlation degree is 0.5.

[0049] Online courses are selected as recommended courses in descending order of priority coefficient until the required number of recommended courses is reached.

[0050] Specifically, the process by which the range determination module determines the group mapping correlation degree includes: For a single online course to be selected, identify the associated user corresponding to that online course; A scatter plot of each associated user is drawn based on the chapter jump dispersion of each associated user and the behavioral similarity between each associated user and a single recommendation item. The group mapping correlation degree of the online course to be selected is determined by the ratio of the clustering quality value of the scatter plot to the preset distance quality value.

[0051] The associated users corresponding to the online courses to be selected are users who have studied the online courses to be selected for more than three-quarters of the total time of the corresponding behavior monitoring period in the historical behavior monitoring period. The historical behavior monitoring period is the 10 behavior monitoring periods before the current behavior monitoring period. The formula for calculating the chapter jump dispersion D of a single associated user is:

[0052] Where T represents the total number of times the user actively switches between or clicks to jump to different chapters during the online course; t represents the order of the jump actions, where t is 1, 2, 3, ..., T. Each time a chapter jump is completed, the final chapter number where the user stays to study is recorded, with Chapter 1 recorded as 1, Chapter 5 as 5, etc. This represents the chapter number reached by the user during their t-th learning session. This is the average number of all chapter numbers accessed by this user. The method for confirming behavioral similarity is as follows: For a single associated user and a single recommendation item, calculate the mouse feature values ​​of each abnormal learning user in each normal learning period, as well as the mouse feature values ​​of the associated user in the corresponding historical behavior monitoring period; the normal learning period is other periods outside the abnormal learning period. The formula for calculating behavioral similarity R is:

[0053] Where N is the number of associated users corresponding to the recommendation item, and i is 1, 2, ..., N. Let a1 be the average normal pointing correlation between the i-th associated user and each abnormal learning user in the recommendation item. Let a2 be the mouse feature value of any abnormal learning user during the normal learning period, and let a2 be the mouse feature value of the i-th associated user during the total learning period in the corresponding historical behavior monitoring cycle. The normal pointing correlation between a single abnormal learning user and the i-th associated user in the recommendation item is 1-|a1-a2| / (the larger of a1 and a2 + ε).

[0054] When plotting a scatter plot of each associated user based on the chapter jump dispersion of each associated user and the behavioral similarity between each associated user and the recommendation item, the behavioral similarity is used as the horizontal axis and the chapter jump dispersion is used as the vertical axis to map each associated user to a coordinate point in the scatter plot.

[0055] Cluster quality value = 1 - the arithmetic mean of the Euclidean straight-line distances from the coordinate points corresponding to each associated user to the reference point (1,0); for any coordinate point (R, D), the Euclidean straight-line distance L between the two points is: The baseline (1,0) represents the ideal, normal learning user with the highest behavioral similarity and the most stable chapter navigation.

[0056] Group mapping correlation = cluster quality value / preset cluster quality value; The preset clustering quality value can be determined according to the actual application scenario. The clustering quality value effectively reflects the degree to which the group behavior pattern between the online courses to be selected and the recommended items converges to the ideal behavior point (1,0). The greater the user's need for the credibility and risk control of the recommendation results, the larger the preset clustering quality value. In this embodiment, the preset clustering quality value is 0.6.

[0057] Specifically, the range optimization module determines to perform range optimization when the recommended range freshness is less than the preset recommended range freshness or the negative feedback coefficient is greater than or equal to the preset negative feedback coefficient.

[0058] The freshness of the recommendation range corresponding to a single recommendation item is the average of the sub-freshness corresponding to each abnormal learning user in that recommendation item. The sub-freshness corresponding to a single abnormal learning user = 1 - the number of online courses that the abnormal learning user has already studied in the online course recommendation range corresponding to that recommendation item / the number of online courses in the online course recommendation range corresponding to that recommendation item. Negative feedback coefficient = 1 - number of abnormal learning users who clicked and studied recommended online courses in adjacent behavior monitoring periods / (number of abnormal behavior users in adjacent behavior monitoring periods + ε); adjacent monitoring periods are behavior monitoring periods that are earlier than the current behavior monitoring period and adjacent to the current behavior monitoring period; The values ​​of the preset recommendation range freshness and the preset negative feedback coefficient can be determined based on the user's historical learning behavior data. The recommendation range freshness and the negative feedback coefficient effectively reflect the user's tolerance for repetitive content and their sensitivity to unsuitable content. The greater the user's demand for improved recommendation accuracy, the greater the value of the preset recommendation range freshness and the smaller the value of the preset negative feedback coefficient. The greater the user's demand for novel content, the higher the value of the preset recommendation range freshness and the lower the value of the preset negative feedback coefficient. In this embodiment, the preset recommendation range freshness is 0.8 and the preset negative feedback coefficient is 0.45.

[0059] Specifically, the range optimization module responds to the condition that the correlation of recommended items is greater than or equal to the correlation of preset recommended items, and determines to add online courses based on the correlation of online courses; In the process of adding online courses based on their relevance, online courses that meet the preset conditions are selected for each recommendation item in descending order of their relevance, until the required number of online courses to be added is reached. The number of additional online courses is determined based on the freshness of the recommendation range and the negative feedback coefficient. The preset condition is that none of the users with abnormal behavior in the recommendation item have studied the online course, and the online course is not within the initial online course recommendation range.

[0060] The correlation tightness of a single recommendation item is the average of the sub-correlation mean values ​​of each abnormal learning user in that recommendation item. For a single abnormal learning user, each online course studied by the abnormal learning user is recorded as a reference online course. For a single reference online course, the average of the sub-correlation mean values ​​of the single reference online course and other reference online courses is recorded as the sub-correlation mean value. The sub-correlation mean value of any two online courses = 1 - sub-difference coefficient. The preset recommendation item association tightness value can be determined based on the platform's tolerance for recommendation diversity and the distribution characteristics of user group knowledge consistency. The recommendation item association tightness effectively reflects the degree of consistency of knowledge structure among the courses learned by users within the recommendation item. The greater the user's demand for the strength of group consensus of the recommendation results, the greater the preset recommendation item association tightness value. In this embodiment, the preset recommendation item association tightness is 0.75.

[0061] For a single online course to be recommended and a single recommendation item, the course correlation is the average of the sub-correlation between the online courses corresponding to each user with abnormal behavior and the online course to be recommended. If the abnormal assessment coefficient is greater than or equal to the preset abnormal assessment coefficient, then the number of online courses is set to = the first proportional coefficient × the threshold for the number of recommended online courses. If the abnormal assessment coefficient is less than the preset abnormal assessment coefficient, then the number of online courses is set to = the second proportional coefficient × the threshold for the number of recommended online courses. The first proportionality coefficient is greater than the second proportionality coefficient. The value range of the first proportionality coefficient is [1.0, 1.3], and the value range of the second proportionality coefficient is [0.5, 0.8]. In this embodiment, the first proportionality coefficient is preferably 1.2, and the second proportionality coefficient is preferably 0.65.

[0062] Anomaly assessment coefficient = (1 - Freshness of recommended range / Freshness of preset recommended range) × Freshness weight coefficient + (Negative feedback coefficient / Preset negative feedback coefficient) × Negative feedback weight coefficient, where both the freshness weight coefficient and the negative feedback weight coefficient are 0.5; The value of the preset anomaly evaluation coefficient can be determined based on the user's historical learning behavior data and learning stability indicators. The anomaly evaluation coefficient effectively reflects the comprehensive fit between the current recommendation range and the user's learning status. The greater the user's demand for the accuracy of online course recommendations, the smaller the value of the preset anomaly evaluation coefficient. In this embodiment, the preset anomaly evaluation coefficient is 0.6.

[0063] The newly added online courses to be recommended, along with the initial online courses selected based on the number of recommended courses corresponding to the recommendation items and the selection priority coefficient, will be used together as the scope of online course recommendations.

[0064] Specifically, the range optimization module responds to the condition that the correlation tightness of the recommended item is less than the preset correlation tightness of the recommended item, and determines to increase the mapping weight coefficient. The increase in the mapping weight coefficient is negatively correlated with the correlation between the recommended item and the weight.

[0065] The increase in the mapping weight coefficient = initial mapping weight coefficient × (preset association density of recommended items - association density of recommended items) / preset association density of recommended items. The initial mapping weight coefficient is 0.4. It should be noted that after the mapping weight coefficient is increased, the spectral weight coefficient decreases synchronously, and the sum of the spectral weight coefficient and the mapping weight coefficient equals 1; the maximum value of the mapping weight coefficient after the increase is 0.6, and the minimum value of the corresponding spectral weight coefficient is 0.4.

[0066] The adjusted mapping weight coefficient = the initial mapping weight coefficient + the increase value, and 0.6 is used when it exceeds 0.6.

[0067] After increasing the mapping weight coefficient, the scope of online course recommendations is redefined based on the number of online courses recommended for each recommended item and the selection priority coefficient.

[0068] Understandably, the correlation strength of recommended items effectively reflects the consistency of knowledge structure among the courses learned by users within that recommended item. When the correlation strength is greater than or equal to the preset correlation strength, it indicates that the learning content of users within that recommended item is highly similar at the knowledge level. In this case, the existing recommendation scope may be too limited to the repetition or fine-tuning of existing knowledge, lacking the expansion of the common knowledge boundaries of the group. Therefore, online courses are added based on the correlation strength to introduce new online courses that are highly relevant to the existing knowledge graph of the group but not yet covered. When the correlation strength of recommended items is less than the preset correlation strength, it indicates that the learning content of users within that recommended item is relatively scattered and the knowledge structure is highly heterogeneous. In this case, recommendations based on knowledge graphs are prone to large biases or low-confidence matching. Therefore, it is necessary to rely more on the commonalities of group behavior patterns to guide recommendations. Thus, the mapping weight coefficient is increased to improve the proportion of "group mapping correlation" in the selection priority coefficient, thereby replacing content homogeneity with behavioral convergence to ensure the reliability and relevance of recommendations.

[0069] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An AI-based online learning behavior monitoring system, characterized in that, include: The data acquisition module is used to obtain online class monitoring data for several users. The behavior monitoring module, which is connected to the data acquisition module, is used to input the online class monitoring data of each user into the AI ​​model to identify users with abnormal behavior and corresponding abnormal learning periods; The recommendation analysis module, which is connected to the behavior monitoring module, is used to determine the online course recommendation status based on the number of users with abnormal behavior and the course differences of users with abnormal behavior, and to determine the recommendation method and recommendation items based on the online course recommendation status. The recommendation method is either a combination recommendation based on learning isomorphism or a separate recommendation. The learning isomorphism is determined based on the operation-oriented correlation and knowledge correlation during abnormal learning periods; The range determination module, which is connected to the recommendation analysis module, is used to determine the selection priority coefficient of each online course to be selected based on the graph correlation degree and the group mapping correlation degree, and to determine the online course recommendation range based on the number of online courses recommended for each recommendation item and the selection priority coefficient; wherein, the number of online courses recommended is determined based on the state representation value and topological dispersion of the recommendation item; The range optimization module, which is connected to the range determination module, is used to determine whether to perform range optimization based on the freshness of the recommended range and the negative feedback coefficient of the online course recommendation range. When performing range optimization, it determines whether to add online courses based on the relevance of the recommended items or adjust the mapping weight coefficient based on the relevance of the online courses. The online course recommendation module is connected to the range determination module and the range optimization module respectively, and is used to recommend online courses for each recommendation item based on the online course recommendation range.

2. The AI-based online learning behavior monitoring system according to claim 1, characterized in that, The recommendation analysis module determines the recommendation method as learning isomorphism-based combination recommendation when the number of users with abnormal behavior is greater than or equal to the preset number of users with abnormal behavior and the course difference of users with abnormal behavior is greater than or equal to the preset course difference. In the recommendation based on learning isomorphism, users with abnormal behavior within the current behavior monitoring period are clustered and grouped, and each group is recorded as a recommendation item. In this context, the learning isomorphism between any two users exhibiting abnormal behavior in a single group is greater than or equal to a preset behavior isomorphism, and for any user exhibiting abnormal behavior outside the group, the learning isomorphism between that user and at least one user exhibiting abnormal behavior within the group is less than the preset behavior isomorphism.

3. The AI-based online learning behavior monitoring system according to claim 2, characterized in that, The recommendation analysis module determines that the recommendation method should be to make individual recommendations when the number of users with abnormal behavior is less than the preset number of users with abnormal behavior or the course difference between users with abnormal behavior is less than the preset course difference. In individual recommendations, each user exhibiting abnormal behavior within the current behavior monitoring period is recorded as a separate recommendation item.

4. The AI-based online learning behavior monitoring system according to claim 2, characterized in that, The recommendation analysis module determines the learning isomorphism based on the operation-oriented relevance and knowledge relevance during abnormal learning periods; Among them, the learning isomorphism degree is positively correlated with both the operation direction correlation degree and the knowledge correlation degree. The operation direction correlation degree is determined based on the mouse feature values ​​corresponding to two abnormal learning users. The mouse feature values ​​are determined based on the ratio of the number of directional change angles corresponding to the user's mouse movement trajectory to the total number of line segment pairs during each abnormal learning period.

5. The AI-based online learning behavior monitoring system according to claim 4, characterized in that, The range determination module determines the number of online courses recommended based on the state representation value and topological discreteness of the recommended items. Among them, the number of online courses recommended for a single recommendation item is positively correlated with the state representation value and topological dispersion of that recommendation item.

6. The AI-based online learning behavior monitoring system according to claim 5, characterized in that, The range determination module determines the selection priority coefficient of each online course to be selected based on the graph correlation degree and the group mapping correlation degree; Among them, the selection priority coefficient of a single online course to be selected is positively correlated with both the graph correlation degree and the group mapping correlation degree.

7. The AI-based online learning behavior monitoring system according to claim 6, characterized in that, The process by which the range determination module determines the group mapping correlation degree includes: For a single online course to be selected, identify the associated user corresponding to that online course; A scatter plot of each associated user is drawn based on the chapter jump dispersion of each associated user and the behavioral similarity between each associated user and a single recommendation item. The group mapping correlation degree of the online course to be selected is determined by the ratio of the clustering quality value of the scatter plot to the preset distance quality value.

8. The AI-based online learning behavior monitoring system according to claim 7, characterized in that, The range optimization module determines to perform range optimization when the recommended range freshness is less than the preset recommended range freshness or the negative feedback coefficient is greater than or equal to the preset negative feedback coefficient.

9. The AI-based online learning behavior monitoring system according to claim 8, characterized in that, The range optimization module responds to the condition that the correlation of the recommended item is greater than or equal to the correlation of the preset recommended item, and determines to add online courses based on the correlation of the online courses. In the process of adding online courses based on their relevance, online courses that meet the preset conditions are selected for each recommendation item in descending order of their relevance, until the required number of online courses to be added is reached. The number of additional online courses is determined based on the freshness of the recommendation range and the negative feedback coefficient. The preset condition is that none of the users with abnormal behavior in the recommendation item have studied the online course, and the online course is not within the initial online course recommendation range.

10. The AI-based online learning behavior monitoring system according to claim 9, characterized in that, The range optimization module responds to the condition that the correlation tightness of the recommended item is less than the preset correlation tightness of the recommended item, and determines to increase the mapping weight coefficient. The increase in the mapping weight coefficient is negatively correlated with the correlation between the recommended item and the weight.

Images