Course comment anomaly detection method, system and product based on differential privacy protection

Abstract

The invention relates to the field of online comment anomaly detection, in particular to a course comment anomaly detection method, system and product based on differential privacy protection, and the method comprises the steps: S1, obtaining an original data set of an online education platform, and storing the original data set in a database in a CSV structured format; s2, performing data cleaning on the original data set in the database to obtain a preprocessed original data set; s3, constructing a privacy protection model by adopting a differential privacy technology, wherein the privacy protection model comprises a gradient cutting module and a noise adding module; s4, a DP-TextCNN model is constructed; s5, the DP-TextCNN model is trained, and a classification result is obtained; through combination of a differential privacy technology and a text convolutional neural network model, efficient and accurate identification of abnormal scores in course comments is realized.

Description

Technical Field

[0001] The present invention relates to the field of online comment anomaly detection, and in particular to a course comment anomaly detection method, system and product based on differential privacy protection. Background Art

[0002] With the rapid development of information technology, especially the rise of big data and artificial intelligence technology, online education platforms such as MOOC (Massive Open Online Courses) have become an important channel for global learners to acquire knowledge. MOOC platforms break through the time and space limitations of traditional education by providing rich course resources and flexible learning methods, making knowledge dissemination more extensive and convenient.

[0003] However, with the increase in the number of users and interactions, MOOC platforms face challenges in the quality of course reviews. As an important part of course feedback, user reviews play an important role in the continuous optimization of course quality. However, malicious, false or erroneous reviews may pose a threat to the fairness of the course evaluation system, affecting the accurate evaluation of teaching quality and students' learning experience. Summary of the invention

[0004] The purpose of the present invention is to provide a course review anomaly detection method, system and product based on differential privacy protection, which can efficiently and accurately identify abnormal scores in course reviews.

[0005] In order to achieve the above object, the present invention adopts such technical solution: 1. The course review anomaly detection method based on differential privacy protection includes the following steps performed in sequence: S1: Obtain an original data set of an online education platform, the original data set including scoring data and evaluation data, and store the original data set in a database in a CSV structured format; S2: Clean the original data set in the database, remove irrelevant content and content with incorrect formats, obtain the preprocessed original data set, and divide the preprocessed original data set into a training data set and a verification data set according to a preset ratio; S3: Use differential privacy technology to build a privacy protection model, which includes a gradient clipping module and a noise addition module; S3-1: The gradient clipping module uses the following formula (1) to perform gradient clipping on the gradient vector during model training: (1); in, Represents the gradient vector calculated during model training. l2_norm_clip is a scalar that represents the threshold for gradient clipping. S3-2: This noise adding module adds noise to the gradient after gradient clipping. The following formula (2) is used to add Gaussian noise to the gradient after gradient clipping: (2); in, is the noise standard deviation, which is a constant related to the privacy budget calculation. Noisy_gradient is a vector representing the gradient with Gaussian noise added. Noise_multiplier is the noise multiplier, which is used to control the amount of noise added to the gradient. Indicates that the mean is 0 and the covariance matrix is Gaussian noise; The specific operation of adding noise in step S3-2 is as follows: S3-2-1: Determine the key differential privacy parameters that need to be optimized, including the gradient clipping threshold l2_norm_clip and the noise standard deviation σ; S3-2-2: Generate all possible parameter combinations through Cartesian product; S3-2-3: For each parameter combination in step S3-2-2, use the training dataset to train the privacy-preserving model, and evaluate the performance of the trained privacy-preserving model on the validation dataset; S3-2-4: For each parameter combination, record the performance indicators of the privacy-preserving model on the validation dataset; S3-2-5: By comparing the performance indicators of different parameter combinations, select the parameter combination with the best performance; S3-2-6: Use parameter optimization score graph to visualize the performance scores of different parameter combinations; S3-2-7: Use the selected optimal parameter combination to train the privacy-preserving model again on the complete training dataset for final privacy-preserving model evaluation or deployment; S4: Construct a DP-TextCNN model, which includes a sequentially connected embedding layer, several serially connected one-dimensional convolutional layers with different convolution kernel sizes, a global maximum pooling layer, a concatenation layer, and a fully connected layer; The embedding layer is used to convert the integer index of the text sequence into a high-dimensional dense vector representation; Each of the one-dimensional convolutional layers extracts a corresponding local feature vector from the training data set; The global maximum pooling layer is used to pool the local feature vectors output by each one-dimensional convolutional layer; The concatenation layer is used to concatenate the local feature vectors pooled by different one-dimensional convolutional layers into a unified feature vector; The fully connected layer is used to further process the unified feature vector; S5: Training and performance evaluation of the DP-TextCNN model constructed in step S4, including the following steps: S5-1: Randomly initialize the weights of the DP-TextCNN model to provide a starting point for learning; S5-2: Input the preprocessed original data set, and the DP-TextCNN model performs forward propagation, that is, feature extraction and classification; S5-3: Use the loss function to calculate the difference between the predicted value and the true label; S5-4: Use the loss function of step S5-3 to calculate the gradient of the model parameters, perform back propagation on the DP-TextCNN model, introduce the privacy protection model constructed in step S3, update the gradient clipping and noise parameters according to the model training results, and protect data privacy; S5-5: Regularly evaluate the loss and accuracy on the training set and validation set to prevent overfitting; S5-6: Generate the confusion matrix and ROC curve of the DP-TextCNN model, along the effectiveness of the model in distinguishing normal from anomalous ratings.

[0007] Preferably, in step S1, a web crawler is used to obtain the original data set of the MOOC platform, and an HTML parsing library is used to parse the web page content to extract the rating data and the evaluation data.

[0008] Preferably, step S2 includes the following steps performed in sequence: S2-1: Import the original data set into the data processing system, check the null value data in the original data set, and delete the null value data; S2-2: Use conditional filtering to delete invalid records, including those that do not conform to the format or contain illegal characters; S2-3: Use sentiment analysis tools to determine the sentiment tendency of the evaluation data, and compare the sentiment analysis score with the corresponding score data. If the sentiment tendency of the comment is positive but the score data is low, or the comment is negative but the score data is high, it is considered that there is inconsistency, and the inconsistent records are manually reviewed, automatically corrected, or data cleaned; S2-4: Use text analysis tools to identify evaluation data that is too brief, repetitive, or has no substantive content, and delete the identified evaluation data; S2-5: Build a stop word list and use text processing tools to remove stop words from the evaluation data; S2-6: performing word segmentation processing on the evaluation data processed in step S2-5, mapping each word in the evaluation data to a vector space of a preset dimension, ensuring that the format of all evaluation data is consistent for subsequent processing; S2-7: Standardize the scoring data processed in step S2-5; S2-8: Check the consistency and logic of the scoring data and evaluation data again to ensure that the cleaning process does not introduce new errors; S2-9: Verify the integrity of the original data set to ensure that all necessary data fields are not missing.

[0009] Preferably, step S4 includes two fully connected layers, the first fully connected layer uses the ReLU activation function to increase the nonlinear ability of the model, and the second fully connected layer uses the softmax activation function to output category probabilities for final classification.

[0010] A course review anomaly detection system includes a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, any of the above methods is implemented.

[0011] A course review anomaly detection product includes a computer program, and when the computer program is executed by a processor, it implements any of the above methods.

[0012] By adopting the aforementioned design scheme, the beneficial effects of the present invention are as follows: the present application realizes efficient and accurate identification of abnormal ratings in course reviews by combining differential privacy technology and text convolutional neural network model. First, the original text data is converted into a vector form that can be processed by the model through data preprocessing and text segmentation; then, by integrating differential privacy technology, the data is processed for privacy protection before model training, including gradient clipping and adding noise; then, a DP-TextCNN model is constructed, and text features are extracted through convolutional layers and pooling layers, and classified through fully connected layers; finally, during the model training process, a differential privacy SGD optimizer is used to update parameters and evaluate model performance at the same time, ensuring that high-accuracy anomaly detection is achieved while protecting user privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of raw data set collection and preprocessing of the present invention; Figure 2 A parameter optimization score graph for the differential privacy technology of the present invention; Figure 3 It is the DP-TextCNN model structure diagram of the present invention; Figure 4 The loss function and accuracy of different training epochs of the present invention; Figure 5 is the confusion matrix of the DP-TextCNN model of the present invention; Figure 6It is the ROC curve of the DP-TextCNN model of the present invention; Figure 7 This is a comparison chart of recognition rates of different models of the present invention. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0015] The course review anomaly detection method based on differential privacy protection includes the following steps performed in sequence: like Figure 1 As shown, S1: obtain the original data set of the online education platform, the original data set includes scoring data and evaluation data, and store the original data set in the database in a CSV structured format; In this embodiment, a web crawler is used to obtain the original data set of the MOOC (Massive Open Online Course) platform, and an HTML parsing library (such as BeautifulSoup, lxml, etc.) is used to parse the web page content to extract the rating data and evaluation data.

[0016] The MOOC platform provides rich original data sets for research due to its wide course coverage and large number of user participation.

[0017] Scoring data is quantitative data, and evaluation data is qualitative data. The combination of the two can provide a more comprehensive perspective on course evaluation.

[0018] Subject coverage: Data collection spans multiple disciplines, including law, engineering, computer science, education and teaching, economics and management, science, foreign languages, literature, history and philosophy, psychology, medicine and health, and art and design. This interdisciplinary data collection strategy helps ensure the wide applicability and depth of the research results.

[0019] S2: Perform data cleaning on the original data set in the database, remove irrelevant content and content with incorrect format, obtain the preprocessed original data set, and divide the preprocessed original data set into a training data set and a verification data set according to a preset ratio; in this embodiment, the preprocessed original data set is divided into a training data set and a verification data set according to a preset ratio of 4:1.

[0020] S2-1: Import the original data set into a conventional data processing system, such as a Python environment, use Python's pd.isnull() to check the null value data in the original data set, and delete the null value data; S2-2: Use conditional filtering, such as df.drop() in Python to delete invalid records. Invalid records here refer to records that do not conform to the format or contain illegal characters.

[0021] S2-3: Use sentiment analysis tools to determine the sentiment tendency of the evaluation data, using trained models or pre-trained sentiment analysis tools to achieve this. In this embodiment, a pre-trained BERT (Bidirectional Encoder Representations from Transformers) model or other conventional sentiment analysis tools such as VADER, TextBlob, etc. are used to analyze the sentiment tendency of each comment; the result of the sentiment analysis is usually a score representing the sentiment polarity. Rating comparison: Compare the sentiment analysis score with the corresponding rating data. For example, if the sentiment tendency of the comment is positive, but the rating data is low, or the comment is negative, but the rating data is high, it is considered that there is inconsistency.

[0022] By setting a threshold or rule to identify inconsistent records, for example, if the difference between the sentiment analysis score and the rating data exceeds a certain preset threshold, the record is considered inconsistent.

[0023] For inconsistent records, the following correction strategies can be adopted: Manual review: Inconsistent records are submitted for manual review, and professionals will decide whether the rating data or evaluation data needs to be corrected.

[0024] Automatic correction: Automatically adjust the rating data or evaluation data according to preset rules. For example, if the evaluation data is obviously positive, but the rating data is low, the rating data can be automatically raised to the preset value.

[0025] Data cleaning: If the inconsistency cannot be resolved by correction, you can choose to delete these records to avoid affecting the model training and analysis results.

[0026] S2-4: Use text analysis tools, such as the nltk library in Python, to identify evaluation data that is too brief, repetitive, or has no substantive content, and delete the identified evaluation data; S2-5: construct a stop word list, and use text processing tools, such as jieba word segmentation in Python, to remove stop words in the evaluation data; the stop word list constructed in this embodiment includes 'of', 'of', 'in', 'is', 'I', 'and', 'have', 'all', 'just', 'also', 'not', 'still', 'very', 'all', 'want', 'wear', 'to', 'up', 'one', 'go', 'say', 'big', 'see', 'but', 'will', 'over', 'no', 'right', 'just', 'use', 'only', 'come', 'with', '……', ', '. ', '? ', '! '; S2-6: performing word segmentation processing on the evaluation data processed in step S2-5, mapping each word in the evaluation data to a vector space of a preset dimension. In this embodiment, each word is mapped to a 300-dimensional vector space to ensure that the format of all evaluation data is consistent, which is convenient for subsequent processing; S2-7: Standardize the score data processed in step S2-5, such as scaling to the [0,1] interval or performing Z-score standardization. S2-8: Check the consistency and logic of the scoring data and evaluation data again to ensure that the cleaning process does not introduce new errors. S2-9: Verify the integrity of the original data set to ensure that all necessary data fields are not missing.

[0027] S3: Use differential privacy technology to build a privacy protection model. The privacy protection model processing includes a gradient clipping module and a noise addition module. The specific operations of step S3 are: S3-1: The gradient clipping module uses the following formula (1) to perform gradient clipping on the gradient vector during model training: (1); in, Represents the gradient vector calculated during model training. l2_norm_clip is a scalar that represents the threshold for gradient clipping. Clipping the L2 norm of the gradient before model training helps limit the potential impact of a single gradient update on model parameters, thereby reducing the impact of a single training sample on the final model.

[0028] Gradient clipping is performed on the gradient vector during model training, rather than directly on the original vectorized data. In deep learning models, the gradient is calculated by the back-propagation algorithm, which represents the sensitivity of the loss function to the model parameters.

[0029] S3-2: The noise adding module adds noise to the gradient after gradient clipping. In this embodiment, the following formula (2) is used to add Gaussian noise to the gradient after gradient clipping: (2); in, is the noise standard deviation, which is a constant related to the privacy budget calculation. Noisy_gradient is a vector representing the gradient with Gaussian noise added. Noise_multiplier is the noise multiplier, which is used to control the amount of noise added to the gradient. Indicates that the mean is 0 and the covariance matrix is Gaussian noise; The addition of noise further obscures the association between gradient updates and specific training samples. The gradient is modified by adding Gaussian noise of appropriate size, so that the model training can be performed without leaking too much information about a single data point. The noise_multiplier, l2_norm_clip and The value of can find an appropriate balance between the privacy protection level and model performance.

[0031] The specific operation of adding noise in step S3-2 is as follows: S3-2-1: Determine the key differential privacy parameters that need to be optimized, including the gradient clipping threshold l2_norm_clip and the noise standard deviation σ. In this embodiment, the gradient clipping threshold l2_norm_clip is in the range of [0.01, 0.1], and the noise_multiplier is in the range of [0.01, 0.1].

[0032] S3-2-2: Generate all possible parameter combinations through Cartesian product, where l2_norm_clip includes 10 candidate values ​​and noise_multiplier includes 10 candidate values.

[0033] S3-2-3: For each parameter combination in step S3-2-2, train the privacy-preserving model using the training dataset, and evaluate the performance of the trained privacy-preserving model on the validation dataset, which usually involves cross-validation or using an independent validation set; S3-2-4: For each parameter combination, record the performance indicators of the privacy-preserving model on the validation dataset, such as accuracy, loss value, or other relevant indicators.

[0034] S3-2-5: By comparing the performance indicators of different parameter combinations, select the parameter combination with the best performance. This is usually the parameter combination that makes the model perform best on the validation set.

[0035] S3-2-6: If Figure 2 As shown in the figure, a parameter optimization score graph is used to visualize the performance scores of different parameter combinations, which helps to intuitively understand the impact of parameters on model performance.

[0036] S3-2-7: Use the selected optimal parameter combination to train the privacy-preserving model again on the complete training dataset for final privacy-preserving model evaluation or deployment.

[0037] In the context of differential privacy, the key to confirming the differential privacy parameter combination is to balance privacy protection and model performance, ensuring that the added noise is sufficient to meet the requirements of differential privacy and protect personal information in the training data from being leaked; while satisfying privacy protection, choose as small noise as possible to maintain the model's prediction accuracy and generalization ability.

[0038] S4: Construct Figure 3 The DP-TextCNN model shown in the figure includes an embedding layer connected in sequence, several one-dimensional convolutional layers with different convolution kernel sizes connected in series, a global maximum pooling layer, a splicing layer, and a fully connected layer; The embedding layer is used to convert the integer index of the text sequence into a high-dimensional dense vector representation. The "high-dimensional" here means that the vector space to which each word is mapped has a higher dimension, which usually means that each word can be represented by a vector containing multiple elements. For example, if the embedding dimension is 300, then each word will be represented as a point in a 300-dimensional space. Each dimension of this space may capture some semantic or grammatical features of the word.

[0039] The DP-TextCNN model includes multiple one-dimensional convolutional layers with different convolution kernel sizes, and each one-dimensional convolutional layer is used to extract local feature vectors. In this embodiment, one-dimensional convolutional layers with convolution kernel sizes of 3, 4, and 5 are used, which means that for each convolution kernel size, there will be a corresponding convolution layer, and each convolution layer has 128 convolution kernels; the convolutional layers with convolution kernels of 3, 4, and 5 are connected in sequence, which helps the DP-TextCNN model to extract features from different granularities, from smaller local features to larger contextual features.

[0040] The global maximum pooling layer is used to pool the local feature vectors output by each convolutional layer, and select the maximum value from the feature map generated by each convolutional layer to form a new feature vector. This process reduces the dimension of the feature while retaining the most important information in each feature map.

[0041] The concatenation layer is used to concatenate the local feature vectors of different convolutional layers after pooling into a unified feature vector. The local feature vectors output by different convolutional layers are concatenated in the order in which they are processed after global max pooling. That is, the feature vectors with a convolution kernel size of 3 are concatenated first, then 4, and finally 5. This concatenation method retains the features captured by each convolutional layer and provides a comprehensive feature representation for the subsequent fully connected layer.

[0042] The fully connected layer is used to further process the unified feature vector. The first fully connected layer uses the ReLU activation function to increase the nonlinearity of the model, while the second fully connected layer uses the softmax activation function to output the category probability for the final classification.

[0043] S5: Training and performance evaluation of the DP-TextCNN model constructed in step S4, including the following steps: S5-1: Randomly initialize the weights of the DP-TextCNN model to provide a starting point for learning; S5-2: Input the preprocessed original data set, and the DP-TextCNN model performs forward propagation, that is, feature extraction and classification; S5-3: Use the loss function to calculate the difference between the predicted value and the true label, such as Figure 4 As shown; S5-4: Use the loss function of step S5-3 to calculate the gradient of the model parameters, perform back propagation on the DP-TextCNN model, introduce the privacy protection model constructed in step S3, update the gradient clipping and noise parameters according to the model training results, and protect data privacy; S5-5: Regularly evaluate the loss and accuracy on the training set and validation set to prevent overfitting; S5-6: Generate the confusion matrix and ROC curve of the DP-TextCNN model, along the effectiveness of the model in distinguishing normal from anomalous ratings.

[0044] Figure 5 The confusion matrix of the DP-TextCNN model is shown. Figure 6 The ROC curve of the DP-TextCNN model is shown. Figure 7 A comparison chart of recognition rates of different models is shown, verifying the effectiveness of the DP-TextCNN model in distinguishing normal scores from abnormal scores.

[0045] This embodiment also provides a detection system that can implement the above method.

[0046] A course review anomaly detection system includes a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, any of the above methods is implemented.

[0047] This embodiment also provides a detection product that can implement the above method.

[0048] A course review anomaly detection product includes a computer program, and when the computer program is executed by a processor, it implements any of the above methods.

[0049] In summary, this application combines differential privacy technology and text convolutional neural network models to achieve efficient and accurate identification of abnormal ratings in course reviews. First, through data preprocessing and text segmentation, the original text data is converted into a vector form that can be processed by the model; then, by integrating differential privacy technology, the data is processed for privacy protection before model training, including gradient clipping and adding noise; then, the DP-TextCNN model is constructed, and text features are extracted through convolutional layers and pooling layers, and classified through fully connected layers; finally, during the model training process, the differential privacy SGD optimizer is used to update parameters and evaluate model performance at the same time to ensure that high-accuracy anomaly detection is achieved while protecting user privacy.

[0050] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A course review anomaly detection method based on differential privacy protection includes the following steps performed in sequence: S1: Obtain an original data set of an online education platform, the original data set including scoring data and evaluation data, and store the original data set in a database in a CSV structured format; S2: Clean the original data set in the database, remove irrelevant content and content with incorrect formats, obtain the preprocessed original data set, and divide the preprocessed original data set into a training data set and a verification data set according to a preset ratio; S3: Use differential privacy technology to build a privacy protection model, which includes a gradient clipping module and a noise addition module; S3-1: The gradient clipping module uses the following formula (1) to perform gradient clipping on the gradient vector during model training: (1); in, Represents the gradient vector calculated during model training. l2_norm_clip is a scalar that represents the threshold for gradient clipping. S3-2: This noise adding module adds noise to the gradient after gradient clipping. The following formula (2) is used to add Gaussian noise to the gradient after gradient clipping: （2）； in, is the noise standard deviation, which is a constant related to the privacy budget calculation. Noisy_gradient is a vector representing the gradient with Gaussian noise added. Noise_multiplier is the noise multiplier, which is used to control the amount of noise added to the gradient. Indicates that the mean is 0 and the covariance matrix is Gaussian noise; The specific operation of adding noise in step S3-2 is as follows: S3-2-1: Determine the key differential privacy parameters that need to be optimized, including the gradient clipping threshold l2_norm_clip and the noise standard deviation σ; S3-2-2: Generate all possible parameter combinations through Cartesian product; S3-2-3: For each parameter combination in step S3-2-2, use the training dataset to train the privacy-preserving model, and evaluate the performance of the trained privacy-preserving model on the validation dataset; S3-2-4: For each parameter combination, record the performance indicators of the privacy-preserving model on the validation dataset; S3-2-5: By comparing the performance indicators of different parameter combinations, select the parameter combination with the best performance; S3-2-6: Use parameter optimization score graph to visualize the performance scores of different parameter combinations; S3-2-7: Use the selected optimal parameter combination to train the privacy-preserving model again on the complete training dataset for final privacy-preserving model evaluation or deployment; S4: Construct a DP-TextCNN model, which includes a sequentially connected embedding layer, several serially connected one-dimensional convolutional layers with different convolution kernel sizes, a global maximum pooling layer, a concatenation layer, and a fully connected layer; The embedding layer is used to convert the integer index of the text sequence into a high-dimensional dense vector representation; Each of the one-dimensional convolutional layers extracts a corresponding local feature vector from the training data set; The global maximum pooling layer is used to pool the local feature vectors output by each one-dimensional convolutional layer; The concatenation layer is used to concatenate the local feature vectors pooled by different one-dimensional convolutional layers into a unified feature vector; The fully connected layer is used to further process the unified feature vector; S5: Training and performance evaluation of the DP-TextCNN model constructed in step S4, including the following steps: S5-1: Randomly initialize the weights of the DP-TextCNN model to provide a starting point for learning; S5-2: Input the preprocessed original data set, and the DP-TextCNN model performs forward propagation, that is, feature extraction and classification; S5-3: Use the loss function to calculate the difference between the predicted value and the true label; S5-4: Use the loss function of step S5-3 to calculate the gradient of the model parameters, perform back propagation on the DP-TextCNN model, introduce the privacy protection model constructed in step S3, update the gradient clipping and noise parameters according to the model training results, and protect data privacy; S5-5: Regularly evaluate the loss and accuracy on the training set and validation set to prevent overfitting; S5-6: Generate the confusion matrix and ROC curve of the DP-TextCNN model, along the effectiveness of the model in distinguishing normal from anomalous ratings.

2. The method for detecting anomalies in course comments based on differential privacy protection as claimed in claim 1, characterized in that: In step S1, a web crawler is used to obtain the original data set of the MOOC platform, and the HTML parsing library is used to parse the web page content to extract the rating data and evaluation data.

3. The course review anomaly detection method based on differential privacy protection as claimed in claim 1, characterized in that: S2-1: Import the original data set into the data processing system, check the null value data in the original data set, and delete the null value data; S2-2: Use conditional filtering to delete invalid records, including those that do not conform to the format or contain illegal characters; S2-3: Use sentiment analysis tools to determine the sentiment tendency of the evaluation data, and compare the sentiment analysis score with the corresponding score data. If the sentiment tendency of the comment is positive but the score data is low, or the comment is negative but the score data is high, it is considered that there is inconsistency, and the inconsistent records are manually reviewed, automatically corrected, or data cleaned; S2-4: Use text analysis tools to identify evaluation data that is too brief, repetitive, or has no substantive content, and delete the identified evaluation data; S2-5: Build a stop word list and use text processing tools to remove stop words from the evaluation data; S2-6: performing word segmentation processing on the evaluation data processed in step S2-5, mapping each word in the evaluation data to a vector space of a preset dimension, ensuring that the format of all evaluation data is consistent for subsequent processing; S2-7: Standardize the scoring data processed in step S2-5; S2-8: Check the consistency and logic of the scoring data and evaluation data again to ensure that the cleaning process does not introduce new errors; S2-9: Verify the integrity of the original data set to ensure that all necessary data fields are not missing.

4. The course review anomaly detection method based on differential privacy protection as claimed in claim 1, characterized in that: In step S4, two fully connected layers are included. The first fully connected layer uses the ReLU activation function to increase the nonlinear ability of the model, while the second fully connected layer uses the softmax activation function to output category probabilities for final classification.

5. A course review anomaly detection system, comprising a memory and a processor, wherein a computer program is stored in the memory, characterized in that: When the computer program is executed by the processor, the method according to any one of claims 1 to 4 is implemented.

6. A course review anomaly detection product, comprising a computer program, characterized in that: When the computer program is executed by a processor, the method described in any one of claims 1 to 4 is implemented.

Images