Student academic performance prediction method and system based on learning behaviors and reflection texts

Abstract

The invention belongs to the technical field of deep learning and education data mining, discloses a student academic performance prediction method based on learning behaviors and reflection texts, and provides a representation framework CE-BERT for the reflection texts of learners to obtain reflection text representation. Through extraction and verification of learning behaviors and reflective text features, the study shows that the learning behaviors and the reflective text features both influence academic performance prediction of students. According to the research, after the two feature sets are integrated, the accuracy of predicting the academic performance of the student is further improved. The discoveries show that the combination of the learning behavior features, the reflective text features and the two feature sets is very important for predicting the academic records, and the method is an important method for predicting the academic records of the students in current mixed learning.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and educational data mining, and in particular relates to a method and system for predicting student academic performance based on learning behavior and reflection text. Background Art

[0002] In the context of blended learning, the methods for predicting students' academic performance are becoming more and more diverse, and have become an important task and challenge. Based on a blended teaching scenario, the present invention proposes an academic performance prediction method that integrates learning behavior and reflective text. Currently, student performance prediction mainly relies on two types of feature information: one is based on students' online learning behavior characteristics, and the other is based on the emotional and cognitive characteristics of students' reflective texts. In existing studies, learning behavior characteristics are often extracted through sequence information such as the number of clicks, learning time, and access frequency. At the same time, some researchers have introduced time series modeling methods such as LSTM and decision trees to model them; while reflective text features mainly use emotional dictionaries, keyword extraction or Bert, Word2Vec and other methods to obtain the emotional tendencies and cognitive states contained in the text.

[0003] However, these methods have certain limitations in practical applications. In terms of learning behavior features, traditional sequence modeling methods are prone to problems such as sparse behavioral data and one-sided feature extraction when faced with hybrid learning, a teaching model that alternates between offline and online learning. They ignore the spatial and temporal dimensions of the interaction between different learning materials and are unable to fully characterize students' investment and participation in aspects such as video learning. As for reflective text features, dictionary methods have limitations in semantic understanding and are difficult to adapt to learning texts with changing contexts. Although the traditional BERT model can capture contextual semantic information, it has anisotropy problems in feature representation, which easily makes it difficult to distinguish different semantic and emotional features, reducing the generalization ability of the prediction model.

[0004] Through the above analysis, the problems and defects of the existing technology are as follows:

[0005] (1) When faced with hybrid learning, a teaching mode that alternates between offline and online learning, traditional sequence modeling methods are prone to problems such as sparse behavioral data and one-sided feature extraction. They ignore the spatial and temporal dimensions of the interaction process between different learning materials and are unable to fully characterize the students' investment and participation in video learning and other aspects.

[0006] (2) For reflective text features, the dictionary method has semantic understanding limitations and is difficult to adapt to learning texts with changing contexts. Although the traditional BERT model can capture contextual semantic information, it has anisotropy problems in feature representation, which easily makes it difficult to distinguish different semantic and emotional features, reducing the generalization ability of the prediction model. Summary of the Invention

[0007] In response to the problems existing in the prior art, the present invention provides a method for predicting student academic performance based on learning behavior and reflection text.

[0008] The present invention is implemented as follows: a method for predicting student academic performance based on learning behavior and reflection text includes:

[0009] Step 1: Review multiple educational data mining related literature to determine the extraction steps of learning behavior characteristics and reflection text features;

[0010] Step 2: Platform construction;

[0011] Step 3, learning behavior feature extraction;

[0012] Step 4: CE-BERT model construction and reflection text feature extraction;

[0013] Step 5, feature fusion;

[0014] Step 6: Verify the authenticity and accuracy of the model results.

[0015] Furthermore, the platform is constructed:

[0016] For hybrid professional courses in colleges and universities, an online learning platform was designed and developed that provides pre-class video previews and guided post-class reflection writing. Combined with explicit data such as learner reflection texts and test answers collected by the backend business system, it provides multimodal data that can be used for modeling for academic performance prediction tasks. Finally, the students' learning behavior data and reflection text data in hybrid course learning are stored in a database to facilitate subsequent feature extraction.

[0017] Furthermore, the learning behavior feature extraction:

[0018] A data preprocessing stage is performed; the Z-score method is used to detect and eliminate outliers; three feature sets are utilized from the perspectives of participation, control, and alignment, and they are combined with the course cycle to extract multidimensional spatiotemporal learning behavior characteristics from video and quiz content.

[0019] Furthermore, the CE-BERT model is constructed and extracted from reflection text features:

[0020] Fully explore the semantic and emotional cognitive information in the reflective text to build a more comprehensive learner state model;

[0021] From the perspective of introducing external domain knowledge, a representation framework CE-BERT is proposed for learners' reflective texts. The K-BERT algorithm is used to enhance the original text at the input level. The implicit psychological characteristics of learners in reflective texts are extracted through the LIWC dictionary and fused with the BERT knowledge semantic representation at the feature level. Finally, the multi-view features in the reflective texts are learned through the linear layer nonlinear mapping capability, and binary academic performance prediction is performed to extract the features of reflective texts in hybrid courses.

[0022] Furthermore, the feature fusion:

[0023] First, timestamp information is used to dynamically align time series and synchronize the timelines of data of different modalities. Second, pooling technology is used for high-dimensional text sequence features to align them with behavioral features on the time scale. Finally, the LinearSVC model is selected as the prediction model. The sensitivity to the fusion features is improved by adjusting the parameters, and visualization is used to reveal the contribution of different modal data to the prediction results.

[0024] Furthermore, the authenticity and accuracy of the verification model results are:

[0025] 100 students used the learning platform in this study during the learning process, of which 41 were from the Database Principles and Applications course and 59 were from the Instructional Technology and Media course; the Database Principles and Applications course was represented by Course 1, and the Instructional Technology and Media course was represented by Course 2; the behavioral data included videos and quizzes, with a total of 23,832 records, and 859 reflective texts were collected; the schedule of this course includes 12 weeks, and this study analyzed the data from weeks 2 to 12; after extracting students' learning behavior characteristics and reflective text characteristics, the effectiveness of the learning behavior feature set and the reflective text feature set in predicting students' academic performance were verified respectively, as well as the further effectiveness of the fusion of the two in predicting students' academic performance.

[0026] Another object of the present invention is to provide a student academic performance prediction system based on learning behavior and reflection text, comprising:

[0027] The determination module is used to review multiple educational data mining related documents and determine the extraction steps of learning behavior characteristics and reflection text features;

[0028] Building modules for platform construction;

[0029] Extraction module, used for learning behavior feature extraction;

[0030] Building modules for CE-BERT model construction and reflection text feature extraction;

[0031] Fusion module, used for feature fusion;

[0032] Verification module, used to verify the authenticity and accuracy of model results.

[0033] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the student academic performance prediction method based on learning behavior and reflective text.

[0034] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the method for predicting student academic performance based on learning behavior and reflective text.

[0035] Another object of the present invention is to provide an information data processing terminal, which is used to implement the student academic performance prediction system based on learning behavior and reflection text.

[0036] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0037] First, the present invention proposes a method for predicting student academic performance in hybrid learning courses, which integrates learning behavior characteristics and reflective text features, making up for the shortcomings of existing methods. In terms of learning behavior feature extraction, the present invention constructs a multi-granularity learning behavior feature set based on the three dimensions of student engagement, control, and alignment, effectively capturing the detailed behavioral characteristics of students in aspects such as video viewing and test completion; in terms of reflective text feature extraction, the present invention combines the K-BERT model and the LIWC sentiment dictionary to capture the emotional cognitive characteristics in student texts, and further proposes a CE-BERT model that integrates contrastive learning to alleviate the semantic aggregation problem of the traditional BERT model in feature expression and improve the accuracy and diversity of text emotional semantic expression.

[0038] In order to achieve unified fusion of features, the present invention adopts the maximum pooling strategy to reduce the dimensionality of high-dimensional features in the feature processing stage. Finally, considering the advantages of LinearSVC in processing high-dimensional data and providing model interpretability, the present invention uses it as a model to verify the effectiveness of the academic performance prediction method in the present invention.

[0039] This paper first investigates the problem of predicting academic performance in small sample sizes and long prediction time windows by focusing on the representation, alignment, and fusion of multi-source heterogeneous data from online learning behaviors and reflection texts in hybrid courses. Specifically, the paper designs and develops a SpringBoot-based online learning platform that centrally collects data on learners' fine-grained interactions with videos, quiz completion, and reflection text content. Next, a student interaction clickstream representation framework for hybrid courses is proposed. Leveraging three external feature sets—engagement, control, and alignment—multidimensional spatiotemporal features of fine-grained video interactions and quiz answering behaviors are extracted. Furthermore, based on the learning engagement ternary framework, the K-BERT algorithm is used to introduce a knowledge graph to expand key knowledge points and abstract concepts in reflection texts. The LIWC dictionary is used to extract emotional cognitive features, and a single linear layer is used to learn multi-view text representations. Furthermore, a weekly time-granularity alignment strategy is explored to perform feature-level fusion of reflection text and learning behavior representations. Using the extracted feature set, the LinearSVC model is used to further verify the accuracy of student academic performance prediction.

[0040] This approach has the following advantages: First, by designing and developing a specialized online learning platform, it can accurately collect learners' multi-dimensional learning behaviors and reflective text data, providing a rich and high-quality data foundation for subsequent academic performance prediction. Second, the student interaction clickstream representation framework can comprehensively and multi-dimensionally extract the spatiotemporal characteristics of learning behaviors based on multiple key processes such as learning content modeling and behavior modeling, drawing on a variety of external feature sets. This effectively compensates for the previous problem of insufficient spatiotemporal feature extraction of learning behavior data and improves the accuracy of academic performance prediction. Third, based on the learning input ternary framework, the reflective text is processed. Not only does it use the K-BERT algorithm to introduce knowledge graphs to expand key knowledge points and abstract concepts, but it also uses the LIWC dictionary to extract emotional cognitive features. This text representation framework, which integrates multiple technologies and methods, can fully explore the cognitive and emotional information contained in the reflective text, solve the anisotropy problem in the original BERT model's text representation, and further improve the effect of text-based academic performance prediction. Finally, the weekly time granularity alignment strategy achieves a good balance between preserving intra-modal sequence information and controlling sequence length. Through feature-level fusion, it combines learning behavior and reflective text features, fully leveraging the advantages of multimodal data and improving the ability to predict learners' academic performance. It provides a new and effective approach for learner modeling and academic performance prediction based on multi-source heterogeneous data in the field of education.

[0041] Second, multimodal data fusion for academic performance prediction in blended learning courses: Previous research on academic performance prediction has mostly focused on single-modal data (such as learning behavior data or text data), while this invention is the first to deeply fuse the learning behavior characteristics and reflective text characteristics in blended learning courses. Through feature-level fusion strategies, it fully utilizes the advantages of multimodal data, significantly improves the accuracy and reliability of academic performance prediction, and fills the technical gap in multimodal data fusion for predicting academic performance in blended learning scenarios.

[0042] Fine-grained learning behavior feature extraction for blended learning: Existing learning behavior feature extraction methods are mostly based on macroscopic page browsing data, making it difficult to capture the fine-grained interactive behaviors at the video level in blended learning. This invention, by designing a specialized online learning platform, accurately collects fine-grained interaction data between learners and video content, and extracts multidimensional spatiotemporal features from the three dimensions of engagement, control, and alignment. This addresses the problem of insufficient fine-grained learning behavior feature extraction in blended learning, providing a new technical means for in-depth analysis of blended learning behaviors and prediction of academic performance.

[0043] Improved Reflective Text Feature Extraction Model (CE-BERT): The traditional BERT model has anisotropy problems in sentiment and cognitive feature extraction, making it difficult to effectively capture sentiment polarity differences and domain characteristics in text. The CE-BERT model proposed in this paper combines knowledge graphs and sentiment dictionaries, introduces external knowledge through the K-BERT algorithm to expand key knowledge points and abstract concepts, and uses the LIWC dictionary to extract sentiment cognitive features. This addresses the shortcomings of the original BERT model in sentiment cognitive representation, provides an innovative technical solution for deep mining of reflective text features in hybrid learning, and fills the technical gap in the field of reflective text feature extraction in hybrid learning.

[0044] In the field of education, learning behavior data and reflective text data come from different modalities and have different characteristics and structures. Traditional methods often have difficulty effectively fusing these two types of data, resulting in inaccurate or biased prediction results. The present invention designs a special feature fusion strategy, uses timestamp information to dynamically align time series, and uses pooling technology to process high-dimensional text sequence features so that they are aligned with behavioral features on a temporal scale, thereby achieving effective fusion of multi-source heterogeneous data. This fusion method not only solves the heterogeneity problem between data modalities, but also fully utilizes the advantages of multimodal data, improving the accuracy of academic performance prediction.

[0045] Previous methods for processing learning behavior data have focused primarily on macro-level behaviors such as page browsing, lacking in-depth exploration of fine-grained interactions at the video level. This has resulted in inadequate extraction of spatiotemporal features from learning behavior data. This invention, by extending the clickstream encoding scheme, incorporates fine-grained interactions with video content and test scores into the clickstream interaction sequence. Furthermore, it extracts multidimensional spatiotemporal learning behavior features from the three dimensions of engagement, control, and alignment. This effectively addresses the shortcomings of previous methods and provides more comprehensive and accurate behavioral data support for academic performance prediction.

[0046] In blended learning courses, the present invention focuses on the additional information generated by the interaction between learners and learning materials, such as fine-grained interactions within the video, the video content itself, and test scores, enriching the spatiotemporal contextual characteristics of learning behavior, thereby improving the ability to predict academic performance based on learning behavior.

[0047] 3. Reflection on the Insufficient Emotional Cognitive Representation in Text Feature Extraction

[0048] The original BERT model suffers from anisotropy in its emotional and cognitive feature extraction, making it difficult to capture differences in domain features across texts. It also ignores the sentiment of words themselves, leading to similar word vectors with different sentiment polarities in context. To address this issue, the present paper proposes the CE-BERT model. By introducing a knowledge graph to expand key knowledge points and abstract concepts in reflective texts, and extracting emotional cognitive features with the help of the LIWC dictionary, the model is then integrated with the BERT knowledge semantic representation at the feature level. This model addresses the shortcomings of the original BERT model in emotional cognitive representation and provides richer information for text-based academic performance prediction.

[0049] In the field of education, especially in blended learning courses, small sample sizes and long prediction windows are often challenging. Traditional methods are prone to overfitting or insufficient prediction accuracy in such situations. This paper successfully addresses the problem of predicting academic performance in small sample sizes and long prediction windows by designing and developing a specialized online learning platform that accurately collects multi-dimensional learning behaviors and reflection text data from learners. This paper combines feature fusion strategies with effective prediction models to successfully address the problem of predicting academic performance in such situations with small sample sizes and long prediction windows.

[0050] In the experiment of the present invention, the behavioral data and reflective text data of 100 students in blended course learning were analyzed to verify the predictive effectiveness of the technical solution in different courses and different stages, indicating that this method can predict students' academic performance under the conditions of small samples and long prediction time windows. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a method for predicting student academic performance based on learning behavior and reflection text provided by an embodiment of the present invention.

[0052] Figure 2 This is a structural block diagram of a student academic performance prediction system based on learning behavior and reflection text provided by an embodiment of the present invention.

[0053] Figure 3 This is a schematic diagram of the overall process of a method for predicting student academic performance based on learning behavior and reflective text in a blended learning course provided by an embodiment of the present invention.

[0054] Figure 4 This is a schematic diagram of an online learning platform developed by the inventor according to an embodiment of the present invention.

[0055] Figure 5 Schematic diagram of the extraction process of the learning behavior feature set provided by an embodiment of the present invention.

[0056] Figure 6 2 is a schematic diagram of a reflective text representation framework of the CE-BERT model provided in an embodiment of the present invention.

[0057] Figure 7 This is a schematic diagram of multi-feature fusion prediction records for weeks 2 to 12 of the course provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0059] This method first draws on research in educational data mining (EDM) and learning analytics (LA) to systematically analyze and summarize learning behavior characteristics (such as clickstream sequences, access frequency, and homework submission intervals) and reflective text characteristics (such as emotional polarity, cognitive hierarchy, and motivational tendencies) that influence academic performance. A standardized feature extraction process was then developed. To ensure the theoretical completeness and practical feasibility of the feature system, bibliometric analysis was used to cross-validate the scientificity and relevance of the feature set.

[0060] During the platform development phase, the learning management system (LMS) behavioral log collection module and the reflection text input interface were integrated, and data cleaning and preprocessing (including anomaly detection, missing value filling, and standardization) processes were deployed. The system architecture adopts a modular design, and the data flow adopts an event-driven architecture (EDA) to ensure low latency and high robustness for large-scale asynchronous data collection, laying the data foundation for subsequent feature extraction and modeling.

[0061] During the feature extraction phase, learning behavior data is encoded into fixed-dimensional feature vectors using sequence modeling methods. These include statistical features (such as mean, standard deviation, and maximum / minimum values) and dynamic features (such as behavioral change rate and cyclical patterns). Reflective text is deeply represented using the improved CE-BERT (Contrastive Enhanced Bidirectional Encoder Representations from Transformers) model. This model employs contrastive learning to enhance the discriminability and semantic richness of text encoding, extracting affective-cognitive embeddings as feature input.

[0062] During the feature fusion and predictive modeling phase, learning behavior features and reflection text features are weighted and fused using an attention mechanism. A multimodal fusion network is employed to preserve the discriminative information of each modality. Finally, academic performance prediction is performed using models such as XGBoost, LightGBM, or an ensemble neural network. Cross-validation (k-Fold Cross Validation) and multi-metric evaluation (such as Accuracy, F1-score, and AUC) are incorporated into the training process to test the model's robustness and generalization performance. An ablation study system is used to verify the contribution of each feature module to overall performance, ensuring statistical significance and interpretability of the prediction results.

[0063] like Figure 1 As shown, the embodiment of the present invention provides a method for predicting student academic performance based on learning behavior and reflection text, which includes the following steps:

[0064] 1. Learning Behavior Feature Extraction Model: By analyzing students' video learning and test-taking behaviors in blended learning courses, we extract multidimensional spatiotemporal features based on engagement, control, and alignment. This model effectively captures students' interactions with learning content, providing data support for subsequent academic performance prediction.

[0065] 2. Reflection Text Feature Extraction Model (CE-BERT): This model, combining knowledge graphs and sentiment lexicons, extracts semantic, sentimental, and cognitive features from reflection texts through an improved BERT model. This model effectively addresses the shortcomings of traditional BERT models in extracting sentiment and cognitive features, providing richer textual information for academic performance prediction.

[0066] The prediction technology described above achieves accurate prediction of students' academic performance by integrating learning behavior characteristics and reflection text characteristics. Specifically, it includes the following steps:

[0067] S101, review multiple educational data mining related literature to determine the extraction steps of learning behavior characteristics and reflection text features;

[0068] Learning behavior has traditionally been used as a sequential feature for predicting academic performance in online learning. For example, Zhao et al. used LSTM to extract temporal patterns in learner behavior and compared the predictive performance of five classifiers. However, existing methods are mostly targeted at MOOCs, which differ significantly from blended courses and present biases. Therefore, in blended courses, attention should be paid to additional information generated by learners' interactions with learning materials, such as fine-grained interactions within videos, the video content itself, and test scores, to enrich the spatiotemporal contextual features of learning behavior. Regarding reflective text, learners' reflective text can depict their cognitive level and underlying emotional state. With the development of natural language processing, pre-trained language models such as Bert (BERT) have been widely used for text feature extraction. For example, Zou et al. used Bert to obtain embedding representations of MOOC forum text, and Liu et al. integrated Bert with a sentiment lexicon to achieve more fine-grained cognitive emotion recognition. However, the Bert model suffers from anisotropy, making it difficult to capture domain feature differences between texts. It also ignores the sentiment information of words themselves, resulting in similar word vectors with different sentiment polarities in context. Therefore, to address the shortcomings of the BERT model, an enhanced deep learning model is needed to better extract multidimensional emotional and cognitive features from the reflection texts of blended courses. In summary, traditional academic performance prediction models rely on static data, which is subject to lags and biases. This study identified the following extraction steps: starting with the learning behaviors and reflection texts in blended learning videos, extracting spatiotemporal features and multidimensional emotional and cognitive features; focusing on additional information about the learner's interaction with the learning materials in the video learning behaviors to avoid data sparsity; and designing an enhanced deep learning model for the reflection texts to improve the BERT model. After feature extraction, an effective blended learning academic performance prediction model is constructed.

[0069] S102, platform construction;

[0070] Combining behavioral and content data can better model learners and predict their academic performance. However, relevant datasets are currently scarce in the education field. Moodle, an open-source MOOC learning platform, collects learner behavior data through its server logs, but this data consists of macro-level clickthrough data such as page views and does not capture fine-grained, controlled interactions at the video level. Based on this, this study designed and developed an online learning platform for hybrid university courses that provides pre-class video previews and guided post-class reflection writing. This platform, combined with explicit data such as learner reflection text and quiz answers collected through backend business systems, provides multimodal data for modeling academic performance prediction. Finally, the learning behavior data and reflection text data from students in hybrid courses are stored in a database to facilitate subsequent feature extraction.

[0071] S103, learning behavior feature extraction;

[0072] In order to effectively capture the spatiotemporal characteristics of learning behavior data, this paper extends the clickstream encoding scheme through content and behavior modeling, incorporating fine-grained interactions with video content and quiz scores into the clickstream interaction sequence. In order to improve the stability and robustness of the model, this study conducted a data preprocessing stage. The Z-score method is used to detect and eliminate outliers to avoid the interference of outlier data on training results. This study utilizes three feature sets from the perspectives of participation, control, and alignment, combines them with the course cycle, and extracts multidimensional spatiotemporal learning behavior characteristics from video and quiz content.

[0073] S104, CE-BERT model construction and reflection text feature extraction;

[0074] Fully mining the semantic and emotional cognitive information in reflective texts and building a more comprehensive learner state model can improve the performance of downstream academic performance prediction tasks. In response to the insufficient emotional cognitive representation and anisotropy problems in the original BERT model, this paper proposes a representation framework CE-BERT for learner reflective texts from the perspective of introducing external domain knowledge. This framework uses the K-BERT algorithm to enhance the original text at the input level and solve the noise problem solved by the introduction of external knowledge. In addition, the learner's implicit psychological characteristics in the reflective text are extracted through the LIWC dictionary and fused with the BERT knowledge semantic representation at the feature level. Finally, the multi-view features in the reflective text are learned through the linear layer nonlinear mapping capability, and binary academic performance prediction is performed to extract the reflective text features in hybrid courses.

[0075] S105, feature fusion;

[0076] To solve the problem of fusing behavioral features and text features in academic performance prediction, the present invention adopts the following feature fusion method: first, timestamp information is used to dynamically align time series and synchronize the timelines of data of different modalities; second, pooling technology is used for high-dimensional text sequence features to align them with behavioral features on the time scale; finally, the LinearSVC model is selected as the prediction model, and the sensitivity to the fused features is improved by adjusting parameters. The contribution of different modal data to the prediction results is revealed with the help of visualization methods.

[0077] S106, verifying the authenticity and accuracy of the model results;

[0078] 100 students used the learning platform of this study during the learning process, of which 41 were from the Database Principles and Applications course and 59 were from the Instructional Technology and Media course. The Database Principles and Applications course was represented by Course 1, and the Instructional Technology and Media course was represented by Course 2. The behavioral data included videos and quizzes, with a total of 23,832 records, and 859 reflective texts were collected. The schedule of this course includes 12 weeks, but due to the scarcity of data in the early stages of the course, this study analyzed the data from weeks 2 to 12. After extracting the students' learning behavior characteristics and reflective text characteristics, the effectiveness of the learning behavior feature set in predicting students' academic performance and the effectiveness of the reflective text feature set in predicting students' academic performance were verified respectively, as well as the further effectiveness of the fusion of the two in predicting students' academic performance. In these experiments, the invention also analyzed the temporal changes in the predictive effectiveness of learning behavior characteristics and reflective text characteristics at different stages of the course.

[0079] like Figure 2 As shown, an embodiment of the present invention provides a student academic performance prediction system based on learning behavior and reflection text, including:

[0080] The determination module is used to review multiple educational data mining related documents and determine the extraction steps of learning behavior characteristics and reflection text features;

[0081] Building modules for platform construction;

[0082] Extraction module, used for learning behavior feature extraction;

[0083] Building modules for CE-BERT model construction and reflection text feature extraction;

[0084] Fusion module, used for feature fusion;

[0085] Verification module, used to verify the authenticity and accuracy of model results.

[0086] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the student academic performance prediction method based on learning behavior and reflective text.

[0087] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the steps of the method for predicting student academic performance based on learning behavior and reflective text.

[0088] Another object of the present invention is to provide an information data processing terminal, which is used to implement the student academic performance prediction system based on learning behavior and reflection text.

[0089] The present invention is specifically implemented:

[0090] Example: 100 students used the learning platform in this study, 41 from the Database Principles and Applications course and 59 from the Instructional Technology and Media course. The Database Principles and Applications course is represented by Course 1, while the Instructional Technology and Media course is represented by Course 2. Behavioral data, including videos and quizzes, totaled 23,832 records, and 859 reflective texts were collected. The course schedule consisted of 12 weeks, but due to the scarcity of data in the early stages of the course, this study analyzed data from Weeks 2–12. Figure 3 This is a schematic diagram of the overall flow of a method for predicting student academic performance based on learning behavior and reflective text in a blended learning course provided by an embodiment of the present invention. Based on this overall flow diagram, the implementation steps of the embodiment are introduced in detail as follows.

[0091] Step 1: Review multiple educational data mining publications to identify the extraction steps for learning behavior features and reflective text features. Learning behavior has traditionally been used as a sequential feature for predicting academic performance in online learning. For example, Zhao et al. used LSTM to extract temporal patterns in learner behavior and compared the predictive performance of five classifiers. However, existing methods are mostly targeted at MOOCs, which differ significantly from blended courses and may exhibit bias. Therefore, in blended courses, attention should be paid to additional information generated by learners' interactions with learning materials, such as fine-grained interactions within videos, the video content itself, and test scores, to enrich the spatiotemporal contextual features of learning behavior. Regarding reflective text, learners' reflective text can depict their cognitive level and underlying emotional state. With the development of natural language processing, pre-trained language models such as BERT have been widely used for text feature extraction. However, the BERT model suffers from anisotropy, making it difficult to capture domain feature differences between texts. It also ignores the sentiment information of words themselves, resulting in similar word vectors with different sentiment polarities within the context. Therefore, to address the shortcomings of the BERT model, an enhanced deep learning model is needed to better extract multidimensional emotional and cognitive features from the reflection texts of blended courses. In summary, traditional academic performance prediction models rely on static data, which is subject to lags and biases. This study identified the following extraction steps: starting with the learning behavior and reflection texts in blended learning videos, spatiotemporal features and multidimensional emotional and cognitive features are extracted; for the learning behavior in the videos, additional information about the learner's interaction with the learning materials is focused on to avoid data sparsity; for the reflection texts, an enhanced deep learning model is designed to improve the BERT model (CE-BERT). After feature extraction, an effective blended learning academic performance prediction model is constructed.

[0092] Step 2: Platform Construction. Combining behavioral and content data can better model learners and predict their academic performance. However, there are currently few relevant datasets in the education field: Moodle is an open-source MOOC learning platform, and its server logs can collect learner behavior data, but this data is macro-click data such as page views, and does not involve fine-grained controlled interactive behaviors at the video level. Based on this, this study designed and developed an online learning platform for hybrid professional courses in colleges and universities that provides pre-class video previews and guided post-class reflection writing. It also uses tracking technology to track learners' implicit behavior data. Combined with explicit data such as learner reflection texts and test answers collected by the backend business system, it provides multimodal data that can be used for modeling for academic performance prediction tasks. In the development of this platform, we adopt the server / browser architecture mode (Browser / Server, B / S) and the front-end and back-end separation construction mode. We choose Bootstrap and jQuery as the front-end framework and SpringBoot as the back-end framework. In terms of deployment, we use cloud server products that are easy for users to access. Finally, the learning behavior data and reflection text data of students in the hybrid course learning are stored in the database for subsequent related feature extraction. The specific platform system interface is as follows: Figure 4 shown.

[0093] Step 3, learning behavior feature extraction. In order to effectively capture the spatiotemporal characteristics of clickstream data, the present invention extends the clickstream encoding scheme through content and behavior modeling, and incorporates fine-grained interactions with video content and quiz scores into the clickstream interaction sequence. In order to improve the stability and robustness of the model, this study conducted a data preprocessing stage. The Z-score method is used to detect and eliminate outliers to avoid the interference of outlier data on the training results. This study uses three feature sets from the perspectives of participation, control, and alignment, combines them with the course cycle, and extracts multidimensional spatiotemporal features from video and quiz content. The combination diagram of the various feature sets that need to be extracted is shown in the figure below. Figure 5 shown.

[0094] a) Engagement feature: This feature set measures students’ interaction with various learning contents from the perspective of duration and frequency, and extracts more fine-grained conversation features within a given time. This study extracted 11 features and obtained the feature matrix

[0095] b) Controllability feature: This feature set focuses on video content and extracts behavioral features based on two granularities: video and in-video operations. This study extracted 16 features and obtained the feature matrix

[0096] c) Alignment feature: This feature focuses on the video content and test content, and focuses on measuring students' knowledge level, learning patterns, and advance planning ability. This study extracted 5 features and obtained the feature matrix

[0097] For each given student u, the clickstream Iu generated in the first w weeks of course c is used to extract features based on the above feature set. The feature representation of each student is denoted as hubehavio. The feature matrix of all students U enrolled in course c is defined as The final feature matrix Hbehavior can be obtained by the following formula:

[0098] Hbehavior=[H1·H2·H3]

[0099] Step 4, CE-BER model construction and reflection text feature extraction. In response to the insufficient emotional cognitive representation and anisotropy problems in the original BERT model, this paper proposes a representation framework CE-BERT for learner reflection texts from the perspective of introducing external domain knowledge. This study introduced the LIWC dictionary and the CN DBpedia knowledge graph, and enhanced the general semantics of BERT from different perspectives to capture the emotion- and cognition-related information in student reflection texts. This approach ensures a more reasonable distribution of the reflection text representation vector in the vector space, which is helpful for academic performance prediction tasks. Therefore, this study proposed the CE-BERT model, such as Figure 6 As shown, this study emphasizes the representation of reflective text. The following sections describe in detail the semantic representation module, sentiment and cognitive representation module of the CE-BERT model, and the process of extracting and integrating reflective text features. CE-BERT consists of three modules:

[0100] (1) Original text enhancement module: This module normalizes and filters noise on the original text to improve the stability and robustness of text processing. In the reasoning stage, the text is expanded into a tree structure using knowledge triples, which is divided into two steps: knowledge retrieval and knowledge injection:

[0101] Knowledge retrieval: From the input text sequence x=(x1,…,x n ) extracts the entity sequence e=(e1,…,e h ), where e i =(x1,…,x p ). Retrieve triples with the entity as the head entity. If there are multiple candidate triples, select the one with the closest semantics to the head entity.

[0102] Knowledge injection: Fusion of entities and triples to generate sentence tree structure x tree Replace the original text x. For the entity sequence e=(e1,...,e h ), each entity ei is the root node, and its tail entity e i1 ,...,e ik is a leaf node, relation r ij It is an edge and does not expand to multiple layers, but only forms a subtree of one layer. The final structure is as follows:

[0103] ·xtree=(x1,...,ei[(ri1,ei1),...,(rik,eik)],...,xn)

[0104] (2) Knowledge semantic representation module: This module is based on the K-BERT algorithm, integrates knowledge tree structure text, and introduces two mechanisms to reduce noise:

[0105] Soft position encoding: Flatten the triple into a "head entity-relationship-tail entity" sequence, and use soft position encoding to distinguish knowledge from the original token.

[0106] Mask matrix: limits the information propagation between different branches in the attention mechanism. The mask matrix formula is as follows:

[0107]

[0108] Among them, i and j are the positions of the input token, M ij is the value in the mask matrix. This matrix is added to the Transformer attention mechanism to shield the interference of different branches. The formula is as follows:

[0109]

[0110] In order to better represent the overall information of the reflected text, this paper uses the last hidden state of the [CLS] tag as the universal semantic representation of the reflected text. Specifically, in the 12-layer TransformerBERT model, the output of the reflected text is a 768-dimensional vector. In particular, the reflected text x = (x1, x2, ..., xn) generates a universal semantic representation vector after passing through the BERT model.

[0111] (3) Emotional Cognitive Representation Module: In this study, the SC-LIWC (Simplified Chinese LIWC) dictionary in LIWC 2015 was used to extract features from the reflection text, and 92 features were finally obtained.

[0112] Finally, the features extracted by the BERT model are fused with the features extracted by LIWC to complete the final extraction of the reflection text features. This study adopts feature-level fusion. Specifically, for the reflection text x=(x1,x2,...,xn), its semantic representation and emotional cognitive representation Performing reflected text representation The resulting representation of the reflected text completes the CE-BERT model approach and the feature extraction of the reflected text.

[0113] Step 5, feature fusion. Regarding data alignment strategies, this study explored the prediction of academic performance at different stages to evaluate the early prediction capabilities of integrating behavioral and textual data in hybrid courses. For the reflective text data generated by students in the first few weeks of the course, this study used an improved BERT model to extract general semantic features of 768 dimensions. 92-dimensional emotion and cognitive features were extracted using the LIWC dictionary Connect the two to form 860-dimensional reflected text content features This study applied global maximum pooling to each student’s reflection text representation in both the time-week dimension and the reflection text sequence dimension. Finally, the text features of all students are as follows:

[0114] For the clickstream behavior data generated by student u in the first w weeks of course c, this study extracts 32 spatiotemporal behavior features This is extracted from clickstreams across content types using three feature sets (including 11 engagement features, 16 control features, and 5 alignment features). Inspired by max pooling, the maximum value is taken along the time (week) dimension to obtain clickstream behavioral features.

[0115] This study believes that the behavioral features and text features derived from the feature set and CE-BERT text representation framework have good inter-modal consistency and complementarity. In feature fusion, a cascade method is used to achieve feature-level fusion. The behavioral features of student U are taken w weeks before course c. Reflective text features Connect them together to get the final multimodal learner fusion features:

[0116]

[0117] Finally, considering the advantages of the LinearSVC model in processing high-dimensional data and providing model interpretability, this study uses it as a model to verify the effectiveness of the academic performance prediction method in this study.

[0118] Step 6: Verify the authenticity and accuracy of the model results. A total of 100 students used the learning platform in this study during their studies, 41 of whom were from the Database Applications course and 59 from the Instructional Technology and Media course. Course 1 and Course 2 represent the Database Principles and Applications and Instructional Technology and Media courses, respectively. Behavioral data, including videos and quizzes, comprised 23,832 records, and 859 reflective texts were collected. The course schedule consisted of 12 weeks, but due to the scarcity of data in the early stages of the course, this study analyzed data from weeks 2–12.

[0119] To record the predictive performance of learning behavior features, the dataset was collected from behavioral data of two courses: Course 1 and Course 2. The experiment classified students according to the binary labels of successful and unsuccessful. The experiment compared the predictive performance of various feature sets and their combinations over 12 weeks. The input, alignment, and control feature sets were labeled ch, ma, and la, respectively. The three feature sets were arranged and combined to verify their effectiveness, and the experimental results are shown in Table 1. Table 1 shows the combination of all three feature sets that produced the best area under the curve (AUC) and accuracy performance, indicating that each feature set identified relevant features of high-risk students from different perspectives.

[0120] Table 1 Predictive performance records of behavioral characteristics at week 12

[0121]

[0122] Table 2 Comparative experimental results of different text representation methods

[0123]

[0124]

[0125] For the comparison of different text representation methods, this experiment evaluated the effectiveness of the CE-BERT model in extracting reflective text features. The resulting binary labels were used to classify students, with a success rate of 47.4% and an unsuccessful rate of 52.6%. The present invention used three models, BERT, K-BERT, and BERT-LIWC, for experiments. The study used the precision, recall, and F1 scores of the classification task to evaluate the effectiveness of text representation in downstream academic performance prediction tasks. Table 2 shows the final experimental results. As shown in Table 2, the accuracy and F1 scores of the CE-BERT representation framework were 76.9% and 76.0%, respectively. The framework integrated sentiment and cognitive features and performed best. It represents the effectiveness of advanced multidimensional features in reflective text, and is 3.3% higher than the BERT model in F1 score.

[0126] Finally, when using multi-feature fusion prediction, this experiment studies the effectiveness of combining behavior and text content data to predict learners' learning performance. This paper uses the dataset of Course 2, which includes data from 59 students and has 12,148 behavior records involving videos and tests; 418 reflection texts, and selects the linear SVC algorithm as the prediction model. The results are as follows Figure 7 As shown. Figure 7 As can be clearly seen in the figure, behavioral features demonstrate stronger early-stage predictive power than text features. However, in later stages, the predictive power of text features significantly outperforms behavioral features. As the number of weeks increases, the accuracy of the fused feature set of behavioral and reflective text features gradually improves. After integrating both behavioral and text features, overall predictive performance is further improved. The linear SVC model, combining behavioral and text features, outperforms the single feature set in all four classification evaluation metrics. This model achieves the highest AUC score of 80% and an accuracy of 77%.

[0127] 1. Specific application fields or related products of the present invention.

[0128] The specific application areas of the present invention include teaching management and learning effect evaluation of mixed-major courses in colleges and universities.

[0129] The relevant product is an online learning platform based on SpringBoot. This study aims at hybrid professional courses in colleges and universities, and designs and develops an online learning platform that provides pre-class video previews and post-class guided reflection writing. It also uses tracking technology to track learners' implicit behavior data, and combines learners' reflection texts and test answers collected by the background business system with explicit data to provide multimodal data that can be used for modeling for academic performance prediction tasks. In the development of this platform, the server / browser architecture mode (Browser / Server, B / S) and the front-end and back-end separation construction mode are adopted. Bootstrap and jQuery are selected for the front-end framework, and SpringBoot is selected for the back-end framework. In terms of deployment, a cloud server product that is easy for users to access is used. Finally, the students' learning behavior data and reflection text data in the hybrid course learning are saved in the database for subsequent related feature extraction. The specific platform system interface is as follows: Figure 4 shown.

[0130] 2. Relevant evidence of the technical effects obtained by the embodiments of the present invention.

[0131] A total of 100 students used the learning platform during their studies, 41 from the Database Applications course and 59 from the Instructional Technology and Media course. Course 1 and Course 2 represent the Database Principles and Applications and Instructional Technology and Media courses, respectively. Behavioral data, including videos and quizzes, comprised 23,832 records, and 859 reflective texts were collected. The course schedule consisted of 12 weeks, but due to the scarcity of data from the early stages of the course, this study analyzed data from Weeks 2–12.

[0132] In the experiment, the effectiveness of the learning behavior feature set and the reflection text feature set in predicting students' academic performance were verified respectively, as well as the further effectiveness of the fusion of the two in predicting students' academic performance.

[0133] To record the predictive performance of learning behavior features, the dataset was collected from behavioral data of two courses: Course 1 and Course 2. The experiment classified students according to the binary labels of successful and unsuccessful. The experiment compared the predictive performance of various feature sets and their combinations over 12 weeks. The input, alignment, and control feature sets were labeled ch, ma, and la, respectively. The three feature sets were arranged and combined to verify their effectiveness, and the experimental results are shown in Table 1. Table 1 shows the combination of all three feature sets that produced the best area under the curve (AUC) and accuracy performance, indicating that each feature set identified relevant features of high-risk students from different perspectives.

[0134] Table 1 Predictive performance records of behavioral characteristics at week 12

[0135]

[0136]

[0137] For the comparison of different text representation methods, this experiment evaluated the effectiveness of the CE-BERT model in extracting reflective text features. The resulting binary labels were used to classify students, with a success rate of 47.4% and an unsuccessful rate of 52.6%. The present invention used three models, BERT, K-BERT, and BERT-LIWC, for experiments. The study used the precision, recall, and F1 scores of the classification task to evaluate the effectiveness of text representation in downstream academic performance prediction tasks. Table 2 shows the final experimental results. As shown in Table 2, the accuracy and F1 scores of the CE-BERT representation framework were 76.9% and 76.0%, respectively. The framework integrated sentiment and cognitive features and performed best. It represents the effectiveness of advanced multidimensional features in reflective text, and is 3.3% higher than the BERT model in F1 score.

[0138] 2 Comparative experimental results of different text representation methods

[0139]

[0140] Finally, when using multi-feature fusion prediction, this experiment studies the effectiveness of combining behavior and text content data to predict learners' learning performance. This paper uses the dataset of Course 2, which includes data from 59 students and has 12,148 behavior records involving videos and tests; 418 reflection texts, and selects the linear SVC algorithm as the prediction model. The results are as follows Figure 7 As shown. Figure 7 As can be clearly seen in the figure, behavioral features demonstrate stronger early-stage predictive power than text features. However, in later stages, the predictive power of text features significantly outperforms behavioral features. As the number of weeks increases, the accuracy of the fused feature set of behavioral and reflective text features gradually improves. After integrating both behavioral and text features, overall predictive performance is further improved. The linear SVC model, combining behavioral and text features, outperforms the single feature set in all four classification evaluation metrics. This model achieves the highest AUC score of 80% and an accuracy of 77%.

[0141] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0142] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for predicting student academic performance based on learning behavior and reflection text, characterized in that: The following steps are involved: Step 1: Based on the research results in the field of educational data mining and learning analytics, we reviewed multiple relevant papers to identify the learning behavior characteristics and reflective text features that affect academic performance and developed a standardized feature extraction process; Step 2: Build a platform that integrates the learning management system's behavior log collection module with the reflection text input interface. The platform uses an event-driven architecture and configures data cleaning and preprocessing processes, including anomaly detection, missing value filling, and normalization. Step 3: Use the time series modeling method to extract features from the learning behavior data, including statistical features and dynamic features; Step 4: Build a CE-BERT model optimized based on contrastive learning mechanism to perform feature encoding on the reflection text and extract emotional cognitive embedding as text representation features; Step 5: Use the attention mechanism to perform weighted fusion of learning behavior features and reflection text features, and build a multimodal fusion network to retain the discriminative information of each modality feature; In step 6, based on the fusion features, an integrated learning model is used to predict academic performance, and the robustness and prediction accuracy of the model are verified through cross-validation and multi-indicator evaluation.

2. The method for predicting student academic performance based on learning behavior and reflection text as claimed in claim 1, characterized in that: The platform is constructed as follows: For hybrid professional courses in colleges and universities, an online learning platform was designed and developed that provides pre-class video previews and guided post-class reflection writing. Combined with explicit data such as learner reflection texts and test answers collected by the backend business system, it provides multimodal data that can be used for modeling for academic performance prediction tasks. Finally, the students' learning behavior data and reflection text data in hybrid course learning are stored in a database to facilitate subsequent feature extraction.

3. The method for predicting student academic performance based on learning behavior and reflection text as claimed in claim 1, characterized in that: The learning behavior feature extraction: A data preprocessing stage is performed; the Z-score method is used to detect and eliminate outliers; three feature sets are utilized from the perspectives of participation, control, and alignment, and they are combined with the course cycle to extract multidimensional spatiotemporal learning behavior characteristics from video and quiz content.

4. The method for predicting student academic performance based on learning behavior and reflection text as claimed in claim 1, characterized in that: The CE-BERT model is constructed and reflected on text feature extraction: Fully explore the semantic and emotional cognitive information in the reflective text to build a more comprehensive learner state model; From the perspective of introducing external domain knowledge, a representation framework CE-BERT is proposed for learners' reflective texts. The K-BERT algorithm is used to enhance the original text at the input level. The implicit psychological characteristics of learners in reflective texts are extracted through the LIWC dictionary and fused with the BERT knowledge semantic representation at the feature level. Finally, the multi-view features in the reflective texts are learned through the linear layer nonlinear mapping capability, and binary academic performance prediction is performed to extract the features of reflective texts in hybrid courses.

5. The method for predicting student academic performance based on learning behavior and reflection text as claimed in claim 1, characterized in that: The feature fusion: First, timestamp information is used to dynamically align time series and synchronize the timelines of data of different modalities. Second, pooling technology is used for high-dimensional text sequence features to align them with behavioral features on the time scale. Finally, the LinearSVC model is selected as the prediction model. The sensitivity to the fusion features is improved by adjusting the parameters, and visualization is used to reveal the contribution of different modal data to the prediction results.

6. The method for predicting student academic performance based on learning behavior and reflection text as claimed in claim 1, characterized in that: The authenticity and accuracy of the verification model results: 100 students used the learning platform in this study during the learning process, of which 41 were from the Database Principles and Applications course and 59 were from the Instructional Technology and Media course; the Database Principles and Applications course was represented by Course 1, and the Instructional Technology and Media course was represented by Course 2; the behavioral data included videos and quizzes, with a total of 23,832 records, and 859 reflective texts were collected; the schedule of this course includes 12 weeks, and this study analyzed the data from weeks 2 to 12; after extracting students' learning behavior characteristics and reflective text characteristics, the effectiveness of the learning behavior feature set and the reflective text feature set in predicting students' academic performance were verified respectively, as well as the further effectiveness of the fusion of the two in predicting students' academic performance.

7. A student academic performance prediction system based on learning behavior and reflection texts, which implements the student academic performance prediction method based on learning behavior and reflection texts as described in any one of claims 1 to 6, characterized in that: The student academic performance prediction system based on learning behavior and reflection text includes: The determination module is used to review multiple educational data mining related documents and determine the extraction steps of learning behavior characteristics and reflection text features; Building modules for platform construction; Extraction module, used for learning behavior feature extraction; Building modules for CE-BERT model construction and reflection text feature extraction; Fusion module, used for feature fusion; Verification module, used to verify the authenticity and accuracy of model results.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the student academic performance prediction method based on learning behavior and reflection text as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the student academic performance prediction method based on learning behavior and reflection text as described in any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the student academic performance prediction system based on learning behavior and reflection text as described in claim 7.

Images