An abnormal behavior sensitive student behavior time sequence modeling and academic early warning method

Abstract

The abnormal behavior sensitive student behavior timing modeling and academic warning method belongs to the field of education data mining in big data mining. The method uses the student behavior data collected by the mobile device in school to construct the student campus behavior heterogeneous information network, extracts the meta-path example that can reveal the student campus behavior for coding, and aggregates the representation of each meta-path example through the attention mechanism to learn the student campus behavior pattern representation and improve the discriminability of the student behavior pattern embedding based on the mobile device data. Meanwhile, in order to improve the timeliness of early perception of abnormal behavior, the abnormal behavior sensitive gating module based on the attention mechanism is proposed to effectively integrate the long-term and short-term behaviors of students, establish a more semantic information student campus timing behavior representation, and improve the accuracy of student academic grade prediction.

Description

Technical Field

[0001] This invention belongs to the field of educational data mining within big data mining, specifically involving a method for time-series modeling of student behavior and academic early warning that is sensitive to abnormal behavior. Background Technology

[0002] As academic subjects become more challenging, exams become more difficult, and university enrollment expands, the academic pressure on students is gradually increasing, leading to frequent occurrences of absenteeism, failing grades, and delayed graduation. For student affairs administrators, early detection and intervention of academic problems, along with providing assistance, can prevent delayed graduation and expulsion, ensuring students' healthy development. Currently, most schools rely heavily on regular exams to monitor student academic progress, but these exams have time constraints. For example, student assessments are typically conducted at fixed times, such as mid-term or end-of-term. Student affairs administrators can only rely on exam results to help students who have failed or are close to failing, and cannot provide early warnings for students at risk of academic difficulties.

[0003] From the perspective of educational psychology, students' behavioral performance is closely related to their academic performance. Therefore, analyzing student behavior can help monitor changes in students' academic status in a timely manner, and researchers have already conducted relevant studies in this area. Using behavioral analysis to monitor student academic performance compensates for many shortcomings of conventional examination methods. In particular, with the widespread use of mobile devices, a large amount of individual activity information, such as sleep duration and exercise data, has been collected and accumulated. How to use mobile sensor data to analyze behavioral characteristics, predict student academic performance, and provide a basis for process management for student administrators is beginning to attract more attention from researchers.

[0004] In recent years, with the rapid development of cutting-edge computer science technologies such as artificial intelligence and machine learning, AI technology has provided more accurate and efficient solutions for big data mining and analysis. In this environment, driven by national policies, AI is gradually penetrating various traditional industries such as education, finance, and healthcare, forming emerging cross-application fields. Using AI technology, the massive amounts of student data accumulated by universities and student behavior data collected by mobile devices can be used to analyze and perceive student behavior on campus, explore the impact of students' academic performance at different times on their behavior, monitor students' learning enthusiasm in a timely manner, and provide theoretical support and practical basis for campus student academic management. Summary of the Invention

[0005] The purpose of this invention is to provide a method for sensitive student behavior temporal modeling and academic early warning. This method utilizes student behavior data collected from mobile devices to construct a heterogeneous information network of student campus behavior. Meta-path examples revealing student campus behavior are extracted and encoded, and the representations of these meta-path examples are aggregated through an attention mechanism to learn student campus behavior pattern representations, thereby improving the discriminability of student behavior pattern embedding based on mobile device data. Simultaneously, to improve the timeliness of early detection of abnormal behavior, an attention-based abnormal behavior sensitive gating module is proposed, effectively integrating students' short- and long-term behaviors to establish a more semantically informative temporal representation of student campus behavior, thus improving the accuracy of student academic level prediction.

[0006] To achieve the above objectives, this invention adopts the following technical solution: a method for time-series modeling of student behavior and academic early warning sensitive to abnormal behavior. First, student behavior data is collected using mobile devices. Daily behavior is used as short-term behavior to construct a heterogeneous information network of short-term student campus behavior. Meta-paths that reveal student campus behavior are designed to guide instance extraction on this network. A meta-path instance encoder and attention mechanism are used to learn a representation of student short-term campus behavior patterns with node attribute semantics. Second, student behavior throughout the entire semester is used as long-term behavior. An attention mechanism is used to weight important behavioral patterns in the time-series of academic performance-sensitive behaviors, thus overcoming the shortcoming of current research that does not focus on abnormal behaviors in historical behavior data. Finally, a gating mechanism is used to fuse short-term and long-term student behavior information. Historical and current behaviors are used synergistically to learn more discriminative student time-series behavior representations, effectively reducing the loss of academically relevant behavioral features during behavior modeling and improving the predictive performance of student academic performance.

[0007] A method for time-series modeling of student behavior and academic early warning sensitive to abnormal behavior, the method includes the following steps:

[0008] Step 1: Preprocess the student campus behavior data collected by mobile devices and input it into the model.

[0009] Step 2: Construct a heterogeneous information network of short-term student campus behavior based on daily campus behavior data.

[0010] Step 2.1: Use word embedding algorithm to represent the attributes of each node in the heterogeneous information network as the initial representation of the node.

[0011] Step 2.2: Construct a heterogeneous information network of students' short-term campus behaviors.

[0012] Step 3: Learn the representation of students' short-term campus behavior patterns based on meta-paths.

[0013] Step 3.1: Design metapaths that can reveal students' campus trajectory behaviors.

[0014] Step 3.2: Extract instances from the heterogeneous information network of students' short-term campus behaviors based on metapaths, and encode the behavior instances using a metapath instance encoder.

[0015] Step 3.3: Aggregate the representations of each meta-path using the attention mechanism to obtain the representation of students' short-term campus behavior patterns.

[0016] Step 4: Using semester-long behavior as a model of students' long-term behavior, construct a long-term behavior modeling module based on attention mechanisms. Utilize attention mechanisms to extract the most relevant behavioral patterns to students' academic performance from temporal behavior data, focusing on anomalous behaviors in historical behavioral data.

[0017] Step 5: Construct a fusion module based on a gating mechanism. Utilize the gating mechanism to fuse features from students' short-term and long-term behavioral information, learning a temporal behavioral representation of students who are fully aware of abnormal behavior.

[0018] Step 6: Input the student's temporal behavior representation into the fully connected layer to predict the student's academic performance.

[0019] Compared with the prior art, the present invention has the following significant advantages:

[0020] Compared to other methods previously proposed in the field, this invention can accurately, discriminatively, and predictably model students' temporal behavior on campus based on mobile device data. It designs a heterogeneous information network of students' short-term daily behavior on campus to fully model the semantic information in student behavior. By designing a semester-long behavior modeling module based on an attention mechanism, it perceives abnormal behavior in the early stages of academic anomalies. Through a fusion module based on a gating mechanism, it collaboratively utilizes historical and current behaviors to enhance the predictability of students' temporal behavior on campus. This invention can assist university campus administrators in the early detection of students with academic abnormalities, enabling timely identification and intervention of academic problems, providing assistance to these students, and promptly preventing failures, delayed graduation, and expulsion. Attached Figure Description

[0021] Figure 1 This is a diagram of the overall model structure of the present invention;

[0022] Figure 2 This is a flowchart of the method of the present invention;

[0023] Figure 3 A schematic diagram of the long-term behavior modeling module;

[0024] Figure 4 This is a schematic diagram of the fusion module. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to specific embodiments and the accompanying drawings.

[0026] The overall structural diagram of the present invention is as follows: Figure 1 As shown, the method flow is as follows: Figure 2 As shown. Specifically, it includes the following steps:

[0027] Step 1 involves preprocessing the student accelerometer, sound sensor, and Wi-Fi data collected from mobile devices. This includes removing student samples with numerous missing values, adding seven time-related semantics (early morning, breakfast, morning, lunch, afternoon, dinner, and late night) to the time of student behavior, and replacing latitude and longitude coordinates with specific location names, such as cafeteria, sports field, and teaching building. A sliding window is then used to segment the student behavior data, with the window size representing one day. This segmented daily student behavior data is used as input to the model, and single-day behaviors are defined as short-term student behaviors.

[0028] Step 2: Construct a heterogeneous information network of short-term student campus behavior based on the preprocessed daily student campus behavior data.

[0029] Step 2.1: Use word embedding algorithm to represent the attributes of each node in the heterogeneous information network as the initial representation of the node.

[0030] The heterogeneous information network of short-term student campus behavior includes four types of nodes: student nodes, accelerometer data nodes, sound sensor data nodes, and Wi-Fi data nodes. Each type of node has its corresponding basic attributes. For example, student nodes include attributes such as the student's gender, major, and grade, while Wi-Fi data nodes include the time and location of the occurrence. Therefore, this invention utilizes the GloVe algorithm from the word embedding field to represent the attributes of each type of node as the initial representation of the node.

[0031] Step 2.2: Construct a heterogeneous information network of students' short-term campus behaviors.

[0032] The heterogeneous information network of short-term student behavior on campus includes four types of nodes and three types of links. The four types of nodes are the student nodes mentioned in step 2.1, accelerometer data nodes, sound sensor data nodes, and Wi-Fi data nodes. The three types of links are the connection between student nodes and accelerometer data through student motion records, the connection between student nodes and sound sensor data through records of the student's surrounding sound environment, and the connection between student nodes and Wi-Fi data through records of student internet access.

[0033] Step 3: Learn the representation of students' short-term campus behavior patterns based on meta-paths.

[0034] Step 3.1: Design metapaths that can reveal students' campus trajectory behaviors.

[0035] Different meta-paths express different semantic relationships. Based on the scenario of students' activities on campus, this invention defines three meta-paths revealing students' campus behavior on a heterogeneous information network of short-term campus behavior: SAS, SMS, and SWS. Here, S represents a student-type node, A represents an accelerometer-type node, M represents a microphone-type node, and W represents a Wi-Fi-type node. The semantics of meta-path SAS are that two students have the same actions within the same time period; the semantics of meta-path SMS are that two students have the same sound environment within the same time period; and the semantics of meta-path SWS are that two students used the same Wi-Fi at the same location within the same time period.

[0036] Step 3.2: Extract instances from the heterogeneous information network of students' short-term campus behaviors based on meta-paths, and encode the behavior instances using a meta-path instance encoder.

[0037] The student campus behavior meta-path designed in step 3.1 is used to extract instances from the heterogeneous information network, resulting in multiple instances reflecting student campus behavior. These instances are then encoded using a meta-path instance encoder, as shown in the following equation:

[0038]

[0039] Where MeanEncoder is the mean encoder, P(v,u) is an instance extracted based on the metapath with a starting node of v and an ending node of u, and x P(v,u) For the obtained meta-path instance representation, x v Let x be the starting node. u To indicate termination, m P(v,u) In a metapath instance, t is an intermediate node excluding the start and end nodes, and x is one of the intermediate nodes. t This represents an intermediate node.

[0040] Step 3.3: Calculate the impact of each meta-path instance representation on the target student node using the attention mechanism, calculate its attention score, and weight and aggregate the meta-path instance representations to obtain the short-term behavior pattern representation of the target student node.

[0041] Step 4: Construct a long-term behavior modeling module based on attention mechanisms. Long-term behavior refers to the student's overall behavior throughout a semester. The attention mechanism is used to extract the most relevant behavioral patterns to the student's academic performance from daily behavior, and to focus on anomalous behaviors in historical behavioral data, thereby aggregating these patterns to obtain a representation of the student's behavior throughout the entire semester.

[0042] Through academic early warning tasks, this invention obtains a representation of long-term student behavior patterns that perceive early abnormal behaviors by training the weights of the attention mechanism, such as... Figure 3 As shown. The attention score for short-term behavior on different dates is calculated based on the importance of the generated behavioral pattern representation to the prediction results, using the following formula:

[0043]

[0044] in, Let W represent the short-term behavior of student u on day k, where W is a learnable parameter. The attention scores for each short-term behavior are calculated. The attention scores are then normalized using the Softmax function. The normalized result represents the attention distribution across each short-term behavior input during long-term behavior pattern aggregation, where each value corresponds to the original input. The formula is as follows:

[0045]

[0046] in, Here, n represents the normalized attention weight for each short-term behavior, and n is the total number of short-term behaviors of the student. The student's short-term behavioral attention score on day j is calculated. The attention weights are then weighted and summed with each short-term behavioral representation to obtain the student's long-term behavioral representation L. u The formula is shown below:

[0047]

[0048] Where n is the total number of short-term behaviors of student u. The normalized short-term attentional weights for student u on day j. This represents the short-term behavior of student u on day j.

[0049] Step 5: Construct a fusion module based on a gating mechanism. Utilize the gating mechanism to fuse features from students' short-term and long-term behavioral information, learning a temporal behavioral representation of students who are fully aware of abnormal behavior.

[0050] In the fusion module, this invention introduces a gating mechanism to balance the influence of students' current behavior and past behavior on behavioral representation, such as... Figure 4 As shown. Specifically, this invention uses a learnable gating system to calculate the importance of short-term and long-term behaviors to the overall student behavior ultimately used for academic prediction. The mathematical formula is as follows:

[0051] G u =σ(W S S u +W L Lu )

[0052] Where σ represents the activation function, and this invention uses the sigmoid function. W S and W L S are trainable parameters. u Let L represent the current short-term behavior of student u. u This represents students' long-term behavioral patterns. (G) u The learned gating weights are used to generate the student's short-term and long-term behavioral representations. Using these learned gating weights, this invention fuses the student's short-term and long-term behavioral representations to obtain the final time-series behavioral representation of the student, as shown in the following formula:

[0053] P u =G u S u +(1-G u )L u

[0054] Step 6: Input the student's temporal behavior representation into the fully connected layer to predict the student's academic performance.

[0055] This invention categorizes students' academic performance into four levels: excellent, good, average, and poor, and uses a fully connected layer to predict these levels. During training, a cross-entropy loss function is used to optimize the model, improve the imbalance in the distribution of student academic levels, and accelerate the convergence speed of training. The mathematical formula for the loss function is as follows:

[0056]

[0057] Where N is the total number of samples used for training, T is the number of classes, and y ic p is the sign function (1 if the true class of sample i is equal to c, 0 otherwise). ic Let be the predicted probability that sample i belongs to category c.

[0058] By minimizing the loss function, W and W' in the model can be determined. S W L The optimal value of the parameter.

[0059] Finally, during the testing phase, the student samples are input into the trained model, and the predicted student academic level is output through a fully connected layer.

[0060] This concludes the description of the specific implementation process of the present invention.

Claims

1. A method for time-series modeling of student behavior and academic early warning sensitive to abnormal behavior, characterized in that: Includes the following steps: Step 1: Preprocess the student campus behavior data collected by mobile devices and input it into the model; Step 2: Construct a heterogeneous information network of short-term student campus behavior based on daily campus behavior data; Step 2.1: Use word embedding algorithm to represent the attributes of each node in the heterogeneous information network as the initial representation of the node; Step 2.2: Construct a heterogeneous information network of students' short-term campus behaviors; Step 3: Learn student short-term campus behavior patterns based on meta-paths; Step 3.1: Design metapaths that can reveal students' campus trajectory behaviors; Step 3.2: Extract instances from the heterogeneous information network of students' short-term campus behaviors based on metapaths, and encode the behavior instances using a metapath instance encoder; Step 3.3: Aggregate the representations of each meta-path using the attention mechanism to obtain the representation of students' short-term campus behavior patterns; Step 4: Using semester-long behavior as students' long-term behavior, construct a long-term behavior modeling module based on attention mechanism; use attention mechanism to extract the most important behavioral patterns most relevant to students' academic performance in time-series behavior in order to focus on abnormal behaviors in historical behavior data. Step 5: Construct a fusion module based on a gating mechanism; use the gating mechanism to perform feature fusion on students' short-term and long-term behavioral information, and learn the temporal behavioral representation of students who are fully aware of abnormal behavior; Step 6: Input the student's temporal behavior representation into the fully connected layer to predict the student's academic performance; Specifically, the following steps are included: Step 1: Preprocess the student accelerometer data, sound sensor data, and Wi-Fi data collected by mobile devices. Remove student samples with many missing values. Add seven time semantics to the time of student behavior: early morning, breakfast, morning, lunch, afternoon, dinner, and late night. Replace the latitude and longitude coordinates of student behavior with specific location names, and add at least specific location semantics such as canteen, sports field, and teaching building. Use a sliding window to divide the student behavior data into segments of one day. Use the student behavior data divided by day as the input to the model, and set the single-day behavior as the short-term behavior of students. Step 2: Construct a heterogeneous information network of short-term student campus behavior based on the preprocessed daily student campus behavior data; Step 2.1: Use word embedding algorithm to represent the attributes of each node in the heterogeneous information network as the initial representation of the node; The heterogeneous information network of short-term student campus behavior includes four types of nodes: student nodes, accelerometer data nodes, sound sensor data nodes, and Wi-Fi data nodes. Each type of node has its corresponding basic attributes. Student nodes include the student's gender, major, and grade. Wi-Fi data nodes include the time and location of the occurrence. The GloVe algorithm in the field of word embedding is used to represent the attributes of each type of node as the initial representation of the node. Step 2.2: Construct a heterogeneous information network of students' short-term campus behaviors; The heterogeneous information network of short-term student behavior on campus includes four types of nodes and three types of links. The four types of nodes are the student nodes, accelerometer data nodes, sound sensor data nodes, and Wi-Fi data nodes mentioned in step 2.

1. The three types of links include the connection between student nodes and accelerometer data through student motion records, the connection between student nodes and sound sensor data through records of the student's surrounding sound environment, and the connection between student nodes and Wi-Fi data through records of student internet access. Step 3: Learn student short-term campus behavior patterns based on meta-paths; Step 3.1: Design metapaths that can reveal students' campus trajectory behaviors; Different meta-paths express different semantic relationships. Based on the scenario of students' activities on campus, three meta-paths revealing students' campus behavior are defined on the heterogeneous information network of students' short-term campus behavior: SAS, SMS, and SWS. Here, S represents a student-type node, A represents an accelerometer-type node, M represents a microphone-type node, and W represents a Wi-Fi-type node. The semantics of meta-path SAS is that two students have the same actions within the same time period, the semantics of meta-path SMS is that two students have the same sound environment within the same time period, and the semantics of meta-path SWS is that two students used the same Wi-Fi in the same location within the same time period. Step 3.2: Extract instances from the heterogeneous information network of students' short-term campus behaviors based on metapaths, and encode the behavior instances using a metapath instance encoder; The student campus behavior meta-path designed in step 3.1 is used to extract instances from the heterogeneous information network, resulting in multiple instances reflecting student campus behavior. These instances are then encoded using a meta-path instance encoder, as shown in the following equation: in, For average encoder, This is an instance extracted based on the metapath, with a starting node of v and an ending node of u. The resulting meta-path instance representation, Represented by the starting node, To indicate termination, t represents an intermediate node in the metapath instance, excluding the start and end nodes. t is one of the intermediate nodes. Represented as intermediate nodes; Step 3.3: Calculate the impact of each meta-path instance representation on the target student node using the attention mechanism, calculate its attention score, and weight and aggregate the meta-path instance representations to obtain the short-term behavior pattern representation of the target student node. ; Step 4: Construct a long-term behavior modeling module based on attention mechanism; long-term behavior refers to the overall behavior of students in a semester; use attention mechanism to extract the most important behavior patterns that are most relevant to students' academic performance in daily behavior, and pay attention to abnormal behaviors in historical behavior data, so as to aggregate and obtain the behavior representation of students in the whole semester. Through an academic early warning task, the long-term behavioral pattern representation of students who perceive early abnormal behavior is obtained by training the weights of the attention mechanism. That is, the attention score of short-term behavior on different dates is calculated based on the importance of the generated behavioral pattern representation to the prediction result, as shown in the following formula: in, Let W represent the short-term behavior of student u on day k, where W is a learnable parameter. The short-term behavioral attention score for student u on day k is calculated. The attention score is then normalized using the Softmax function. The normalized result represents the attention distribution across various short-term behavioral inputs during long-term behavioral pattern aggregation, where each value corresponds to the original input. The formula is as follows: in, Here, n represents the normalized attention weight for each short-term behavior, and n is the total number of short-term behaviors of the student. The student's short-term behavioral attention score on day j is calculated; the attention weights are then weighted and summed with each short-term behavioral representation to obtain the student's long-term behavioral representation. The formula is shown below: Where n is the total number of short-term behaviors of student u. The normalized short-term attentional weights for student u on day j. This represents the short-term behavior of student u on day j. Step 5: Construct a fusion module based on a gating mechanism; use the gating mechanism to perform feature fusion on students' short-term and long-term behavioral information, and learn the temporal behavioral representation of students who are fully aware of abnormal behavior; In the fusion module, a gating mechanism is introduced to balance the influence of students' current and past behaviors on behavioral representation. A learnable gating system is used to calculate the importance of short-term and long-term behaviors to the overall student behavior ultimately used for academic prediction; the mathematical formula is as follows: in, This represents the activation function, using the sigmoid function. and For trainable parameters, Let the current short-term behavior of student u be represented. Representation of students' long-term behavioral patterns; The learned gating weights are used to merge the short-term and long-term student behavior representations to obtain the final time-series student behavior representation, as shown in the following formula: Step 6: Input the student's temporal behavior representation into the fully connected layer to predict the student's academic performance; Students' academic performance is categorized into four levels: excellent, good, average, and poor. A fully connected layer is used to predict these levels. During training, a cross-entropy loss function is employed to optimize the model, mitigating the imbalance in the distribution of student academic levels and accelerating training convergence. The mathematical formula for the loss function is as follows: Where N is the total number of samples used for training, and T is the number of classes. The sign function is set to 1 if the true class of sample i is equal to c, and 0 otherwise. Let i be the predicted probability that sample i belongs to category c; By minimizing the loss function, the model can be determined. , , The optimal value of the parameter; Finally, during the testing phase, the student samples are input into the trained model, and the predicted student academic level is output through a fully connected layer.

Images