A Method and System for Sentiment Analysis of Film Reviews Based on Classifiers and Feature Ensemble

Abstract

This invention relates to a method and system for film review sentiment analysis based on classifiers and feature ensemble. The method includes the following steps: S1, converting film reviews into recognizable numerical vectors; S2, inputting the recognizable numerical vectors into a Feature Selection Ensemble (FSE) model to identify key film review features; S3, inputting the key film review features into a Dynamic Ensemble Selection (DES) model based on the META-DES framework for sentiment analysis and prediction. This invention employs feature selection ensemble for film review features and dynamic ensemble selection fusion meta-learning in the classification stage, achieving higher classification accuracy compared to other common machine learning sentiment analysis techniques.

Description

Technical Field

[0001] This invention belongs to the field of data analysis technology, specifically relating to a method and system for film review sentiment analysis based on classifiers and feature integration. Background Technology

[0002] In today's information-rich era with over 4.9 billion internet users, a vast amount of online commentary on social issues is generated daily—including product or service reviews, blog posts, and movie reviews. This online text is valuable for improving decision-making. Analyzing online reviews of products or services can provide consumers with the ability to extract key information that can be used to improve those products or services. Among various forms of entertainment, movies are arguably the best, with people watching them and expressing their opinions on social media. Filmmakers can leverage film review analysis and sentiment analysis to enhance the quality of their work.

[0003] Film review sentiment classification, a process of identifying the sentiment categories of film reviews, plays a crucial role in the film industry. In related research, two main methods are widely used in sentiment analysis: lexical-based classification and machine learning-based classification. Lexical-based methods use a dictionary to determine the polarity of each film review for classification. Unlike lexical-based methods, machine learning methods convert film reviews into numerical data after several preprocessing operations and then use classification algorithms to categorize the data accordingly. Regardless of the method used, two key issues need to be considered: vectorization and the construction of the classification model. To address the data silo problem, federated learning, a distributed machine learning framework, has emerged. This technology allows data owners to form a federation, where participants can collaboratively train a global model without disclosing their local plaintext data, meeting the requirements of user privacy and data confidentiality. Because participants in a federated learning system do not share local data but rather collaboratively train the federated model based on their own data, the performance of the final global model obtained by the federation depends on the data quality of each participant.

[0004] To determine the sentiment category of film reviews, it is necessary to find suitable numerical vectors to represent them. Several popular methods exist for vectorizing film reviews, such as Single Word Model (SOW), Bag-of-Words (BOW), Term Frequency-Inverse Document Frequency (TF-IDF), n-grams, and one-hot encoding. Among these methods, TF-IDF is a mainstream approach. However, after converting film reviews into numerical vectors using TF-IDF, the enormous dimensionality of these vectors implies the presence of noisy and irrelevant features. Furthermore, many studies primarily employ single feature selection algorithms, but their application is not yet comprehensive. Therefore, to improve the efficiency and accuracy of film review sentiment classification, an appropriate feature selection method must be considered. In the classification process, with further research in recent years, classification methods can generally be divided into two categories: static and dynamic. These two types can be further divided into two different methods: single classifier selection and ensemble selection. In the construction of classification models, many static classification methods, and even basic dynamic methods, fail to yield statistically significant experimental results.

[0005] Therefore, given the shortcomings of existing technologies, it is necessary to propose a technical solution to address the problem of sentiment analysis in film reviews. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention aims to provide a film review sentiment analysis method and system based on classifiers and feature ensembles. On the one hand, it uses reasonable text vectorization and feature selection methods to solve the problems of feature redundancy and the inadequacy of single feature selection algorithms. On the other hand, it develops an improved dynamic ensemble selection method for more effective prediction.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0008] A sentiment analysis method for film reviews based on classifiers and feature ensemble includes the following steps:

[0009] S1. Convert film reviews into recognizable digital vectors;

[0010] S2. Integrate the identifiable digital vector input features into the FSE model to identify key film review features;

[0011] S3. Input key film review features into a dynamic ensemble selection DES model based on the META-DES framework for sentiment analysis and prediction.

[0012] As a preferred embodiment, step S1 includes the following steps:

[0013] S11. The BOW model is used to extract textual features from the film reviews as the initial feature set, and each film review is converted into a vector of 0 or 1 based on whether the feature exists in the film review.

[0014] S12. Using the TF-IDF model, calculate the TF-IDF value of each feature in each film review, and convert each film review into a feature vector composed of TF-IDF values.

[0015] As a preferred embodiment, the tf-idf value is calculated as follows:

[0016]

[0017] Among them, f l Representing the l-th film review characteristic, tf l Representative characteristics of film reviews f l The frequency of occurrence in film review texts, where H represents the total number of film review texts, h l Represents all features f l The number of training texts.

[0018] As a preferred embodiment, step S2 includes the following steps:

[0019] S21. The Feature Selection Ensemble (FSE) model is applied to reduce the dimensionality of the high-dimensional feature vectors. The training film reviews are randomly divided into A parts, and B feature selection algorithms are combined with the Feature Selection Ensemble (FSE) model. These algorithms are Information Gain (IG), Gain Ratio (GR), Chi-Square Distribution (CHI), Mutual Information (MI), Relief-F algorithm, and F-test. The calculation methods are as follows:

[0020] IG:

[0021]

[0022] GR:

[0023] Where m represents the total number of film review categories, cl i Let f represent the i-th film review. l Representing the l-th film review characteristic, P(cl) i ) indicates that a film review belongs to cl i The probability of the class, P(f) l ) represents containing f l The probability of a film review appearing in all collections of film reviews. This means that f is not present. l The probability of a film review appearing in all film review collections;

[0024] CHI: Where n represents the total number of film review features, and a represents the number of features included in the review. l And belongs to category cl i The number of film reviews, b indicates that it includes f. l And does not belong to category cli The number of film reviews, where c indicates that f is not included. l And belongs to category cl i The number of film reviews, where d indicates that f is not included. l And does not belong to category cl i The number of film reviews;

[0025] Relief-F:

[0026]

[0027] The Relief-F algorithm requires three stages of computation: First, the initial weights of all features are set to 0. Second, the number of sampling rounds is set to m. In each round, a sample S is randomly selected from the training set, and then the k nearest neighbor sample H is found from samples of the same class as S. j (j = 1, 2, ..., k), find the k nearest neighbor sample M from samples of different classes from S. j (S); The third step is to update the feature weights according to the formula in each round, where W(f l () represents the weight of each feature, initially set to 0. Let represent other classes that do not belong to class S, P(cl) represent the probability that a sample belongs to class cl, class(S) represent the class of a randomly selected sample S, and diff(f) represent the probability that a sample belongs to class cl. l R1, R2) represent feature f l The difference between sample R1 and sample R2, R1[f] l ] represents the feature f of sample R1 l Corresponding value, R2[f l Similarly, max(f) l ) represents all film review characteristics f l The maximum value among the corresponding values, min(f) l ) is the minimum value;

[0028] S22. After B feature selection algorithms iteratively calculate A datasets respectively, each feature selection algorithm will produce A different weight vectors for the A datasets. Using Vec... a,b This represents the weight vector obtained by the b-th feature selection algorithm after calculating the weight vector for the a-th data point. Next, the data in all calculated weight vectors are sorted and selected, using the following calculation method:

[0029]

[0030] Where, τ a,b Indicates targeting Vec a,bThe threshold obtained using the CDF-g algorithm is then applied to the weight vector Vec. a,b In, all values ​​less than the threshold τ a,b The features corresponding to the weights are removed, FS a,b This represents the film review features selected by the b-th feature selection algorithm after processing the a-th dataset. Then, the intersection of these A threshold-filtered feature vector sets yields the final feature selection result FS of the b-th feature selection algorithm across all training data. b Finally, the final key film review features, Fial-features, are obtained by performing a union operation on the results of all B feature selection algorithms.

[0031] As a preferred embodiment, step S3 includes the following steps:

[0032] S31. Establish the base classifier pool: Train the base classifier pool using the feature-selected training dataset T. λ With validation dataset D SEL The data is divided into two parts; an initial base classifier pool P is established, and the base classifiers uniformly use the SVM algorithm, with the training dataset T as the input. λ The base classifiers are trained by sampling different datasets and assigning them to each base classifier. The heterogeneity between the base classifiers is reflected by the differences in the datasets, and thus the initial base classifier pool P is obtained.

[0033] S32. Training Data Filtering: Filtering the training dataset T λ Chinese data x j，train The process iterates, using all base classifiers in the initial base classifier pool P to predict the data; if the percentage of correctly classified classifiers in the base classifier pool exceeds a set threshold h... c If the film reviews are selected, the training data is chosen; otherwise, they are discarded. This process is repeated until all film reviews in the training data have been filtered out. The filtered training data is then T′. λ ;

[0034] S33: Computational domain θ and output file domain Using the k-nn algorithm, each classifier C in the initial base classifier pool P is used to train the dataset T′. λ Chinese data x j，train Calculate its domain θ and the domain of the output file First, iterate through the classifiers in the classifier pool P, and for a given classifier C... i Then for T′ λ The classifier iterates through all the data and calculates x. j，train The domain θ; θ is derived from the validation dataset D SEL The data is composed of the K closest samples, θ = {x1, x2, ..., x...} kNext, x j，train Convert to output file Similarly, all data in the validation set are transformed and set to... Then, the k-nn algorithm is used for calculation. Depend on Zhongyu Nearest K p Composed of 1 sample,

[0035] S34. Establish Meta-features: Classifier C i For training data x j，train This yields the domain θ = {x1, x2, ..., x}. k} and output file fields Then, 15 meta-features are calculated and stored in a vector v. i,j ={f1,f2,f3,…,f 15 In the process, if the classifier predicts correctly, the class attribute α = 1; otherwise, it is 0. Then, the base classifier P in the classifier pool and the data in the training data T′ are iterated continuously, resulting in a vector v consisting of all 15 meta-features. i,j The data is aggregated into meta-training data with dimensions Q×N, representing the number of base classifiers. As training data for the meta classifier needed for subsequent base classification, the label of each training data is the corresponding α;

[0036] The 15 meta-features are as follows:

[0037] neighbors'hard classification,f Hard Using a base classifier C for θ = {x1, x2, ..., x} K Classify the data; if the classification is correct, return 1; otherwise, return 0, generating a K-dimensional binary vector.

[0038] Posterior probability, f Prob : Posterior probability, P(class|x) k Generate a K-dimensional binary vector;

[0039] Overall local accuracy, f Overall : Throughout the entire θ = {x1, x2, ..., x k In the context of classification, the accuracy of the base classifier C is considered.

[0040] Output profiles classification:f Op Using base classifier C Classify the data; if the classification is correct, return 1; otherwise, return 0, generating a K-dimensional binary vector.

[0041] Classifier's confidence, f Conf : The vertical distance between the sample and the decision boundary of C, which is the largest value among the predictions made by the classifier for that sample;

[0042] Conditional local accuracy,f Cond : Calculate local precision, that is, the number of correctly classified items in each category compared to the total number of items in the class θ = {x1, x2, ..., x}. k The ratio of the number of correctly classified items in};

[0043] Ambiguity, f Amb The difference between the highest probability and the second highest probability after classification;

[0044] Logarithmic, f Log : Where S lk (X k ) is the classifier C for θ={x1,x2,…,x k The support value of the correctly classified samples in the classification results of};

[0045] Entropy, f Ent : For θ = {x1, x2, ..., x k The samples in} are calculated separately, and finally a K-dimensional vector is formed; where S l (X k This refers to the support value for misclassification;

[0046] Minimal difference, f MD :

[0047] Kullback-Leibler divergence, f KL :

[0048] Exponential, f Exp :

[0049] Randomized reference classifier,f RRC For θ = {x1, x2, ..., x k For the samples in}, the conditional probability of correct classification is calculated using RRC, resulting in a K-dimensional vector;

[0050] Simplified classifier rank,f Rank The validation set data are sorted from nearest to farthest from the samples, and classifier C is used to classify the validation set data until the first classification error occurs. The number of correctly classified data is then f. Rank ;

[0051] classifier rank OP, f RankOP : Implemented in the decision space, that is, the data of the sample and validation sets are transformed to form the output file domain, and then compared with f. Rank The same processing is used in the middle;

[0052] S35, Feature Selection: In calculation Then, the FSE model from step S2 is used to select important features;

[0053] S36. Training Phase: Using the features selected in step 35 Recorded as First of all The dataset is divided into 75% training data and 25% validation data. The 75% training data is used as input to train the meta classifier. During training, a prediction is made on the validation set after each training round. If the performance on the validation set decreases or fails to improve compared to the previous prediction result at a certain point, the training stops.

[0054] S37. Final Prediction: Predict the film review using the trained meta-classifier. First, use each classifier C from the classifier pool. i Calculate the film review x test Obtain its meta-features vector v test Input metaclassifier, and if the output is α test If the value is 1, then the classifier is selected as a classifier with classification ability and placed into the classifier pool P. test This process is repeated until the final classifier pool P for the film review is obtained. test And by adjusting the final classifier pool P test The prediction results of the middle classifier are used to obtain the final prediction result of the film review by majority voting.

[0055] This invention also provides a film review sentiment analysis system based on classifiers and feature ensemble, comprising:

[0056] The conversion module is used to convert film reviews into recognizable digital vectors.

[0057] The recognition module is used to select and integrate identifiable digital vector input features into the FSE model to identify key film review features;

[0058] The prediction module is used to input key film review features into a dynamic ensemble selection DES model based on the META-DES framework for sentiment analysis prediction.

[0059] As a preferred embodiment, the process of converting film reviews into recognizable digital vectors includes the following steps:

[0060] (1) The text features in the film reviews are extracted using the BOW model as the initial feature set, and each film review is converted into a vector of 0 or 1 based on whether the feature exists in the film review;

[0061] (2) Using the TF-IDF model, calculate the tf-idf value of each feature in each film review, and convert each film review into a feature vector composed of tf-idf values.

[0062] As a preferred embodiment, the tf-idf value is calculated as follows:

[0063]

[0064] Among them, f l Representing the l-th film review characteristic, tf l Representative characteristics of film reviews f l The frequency of occurrence in film review texts, where H represents the total number of film review texts, h l Represents all features f l The number of training texts.

[0065] As a preferred embodiment, the identification process of the identification module includes:

[0066] (1) The feature selection ensemble FSE model is applied to reduce the dimensionality of the huge feature vector. The training film reviews are randomly divided into part A, and the feature selection algorithms B are combined with the feature selection ensemble FSE model. The algorithms are information gain IG, gain ratio GR, chi-square distribution CHI, mutual information MI, Relief-F algorithm, and F test. The calculation methods are as follows:

[0067] IG:

[0068]

[0069] GR:

[0070] Where m represents the total number of film review categories, cl i Let f represent the i-th film review. l Representing the l-th film review characteristic, P(cl) i ) indicates that a film review belongs to cl i The probability of the class, P(f) l ) represents containing f l The probability of a film review appearing in all collections of film reviews. This means that f is not present. l The probability of a film review appearing in all film review collections;

[0071] CHI: Where n represents the total number of film review features, and a represents the number of features included in the review. l And belongs to category cl i The number of film reviews, b indicates that it includes f. l And does not belong to category cl i The number of film reviews, where c indicates that f is not included. l And belongs to category cl i The number of film reviews, where d indicates that f is not included. l And does not belong to category cl i The number of film reviews;

[0072] Relief-F:

[0073]

[0074] The Relief-F algorithm requires three stages of computation: First, the initial weights of all features are set to 0. Second, the number of sampling rounds is set to m. In each round, a sample S is randomly selected from the training set, and then the k nearest neighbor sample H is found from samples of the same class as S. j (j = 1, 2, ..., k), find the k nearest neighbor sample M from samples of different classes from S. j (S); The third step is to update the feature weights according to the formula in each round, where W(f l () represents the weight of each feature, initially set to 0. Let represent other classes that do not belong to class S, P(cl) represent the probability that a sample belongs to class cl, class(S) represent the class of a randomly selected sample S, and diff(f) represent the probability that a sample belongs to class cl. l R1, R2) represent feature f l The difference between sample R1 and sample R2, R1[f] l ] represents the feature f of sample R1 l Corresponding value, R2[f lSimilarly, max(f) l ) represents all film review characteristics f l The maximum value among the corresponding values, min(f) l ) is the minimum value;

[0075] (2) After B feature selection algorithms iteratively calculate A datasets respectively, each feature selection algorithm will produce A different weight vectors for the A datasets. (Using Vec...) a,b This represents the weight vector obtained by the b-th feature selection algorithm after calculating the weight vector for the a-th data point. Next, the data in all calculated weight vectors are sorted and selected, using the following calculation method:

[0076]

[0077] Where, τ a,b Indicates targeting Vec a,b The threshold obtained using the CDF-g algorithm is then applied to the weight vector Vec. a,b In, all values ​​less than the threshold τ a,b The features corresponding to the weights are removed, FS a,b This represents the film review features selected by the b-th feature selection algorithm after processing the a-th dataset. Then, the intersection of these A threshold-filtered feature vector sets yields the final feature selection result FS of the b-th feature selection algorithm across all training data. b Finally, the final key film review features are obtained by performing a union operation on the results of all B feature selection algorithms.

[0078] As a preferred embodiment, the prediction process of the prediction module includes:

[0079] (1) Establish a base classifier pool: Train the training dataset T after feature selection. λ With validation dataset D SEL The data is divided into two parts; an initial base classifier pool P is established, and the base classifiers uniformly use the SVM algorithm, with the training dataset T as the input. λ The base classifiers are trained by sampling different datasets and assigning them to each base classifier. The heterogeneity between the base classifiers is reflected by the differences in the datasets, and thus the initial base classifier pool P is obtained.

[0080] (2) Training data selection: For the training dataset T λ Chinese data x j，train The process iterates, using all base classifiers in the initial base classifier pool P to predict the data; if the percentage of correctly classified classifiers in the base classifier pool exceeds a set threshold h... cIf the film reviews are selected, the training data is chosen; otherwise, they are discarded. This process is repeated until all film reviews in the training data have been filtered out. The filtered training data is then T′. λ ;

[0081] (3) Calculation domain θ and output file domain Using the k-nn algorithm, each classifier C in the initial base classifier pool P is used to train the dataset T′. λ Chinese data x j，train Calculate its domain θ and the domain of the output file First, iterate through the classifiers in the classifier pool P, and for a given classifier C... i Then for T′ λ The classifier iterates through all the data and calculates x. j，train The domain θ; θ is derived from the validation dataset D SEL The data is composed of the K closest samples, θ = {x1, x2, ..., x...} k Next, x j，train Convert to output file Similarly, all data in the validation set are transformed and set to... Then, the k-nn algorithm is used for calculation. Depend on Zhongyu Nearest K p Composed of 1 sample,

[0082] (4) Establish Meta-features: Classifier C i For training data x j，train This yields the domain θ = {x1, x2, ..., x}. k} and output file fields Then, 15 meta-features are calculated and stored in a vector v. i,j ={f1,f2,f3,…,f 15 In the process, if the classifier predicts correctly, the class attribute α = 1; otherwise, it is 0. Then, the base classifier P in the classifier pool is continuously compared with the training data T′. λ Iterate through the data to generate vectors v consisting of all 15 meta-features. i,j The data is aggregated into meta-training data with dimensions Q×N, representing the number of base classifiers. As training data for the meta classifier needed for subsequent base classification, the label of each training data is the corresponding α;

[0083] The 15 meta-features are as follows:

[0084] neighbors'hard classification,f Hard Using a base classifier C for θ = {x1, x2, ..., x} K Classify the data; if the classification is correct, return 1; otherwise, return 0, generating a K-dimensional binary vector.

[0085] Posterior probability, f Prob : Posterior probability, P(class|x) k Generate a K-dimensional binary vector;

[0086] Overall local accuracy, f Overall : Throughout the entire θ = {x1, x2, ..., x k In the context of classification, the accuracy of the base classifier C is considered.

[0087] Output profiles classification: Using base classifier C Classify the data; if the classification is correct, return 1; otherwise, return 0, generating a K-dimensional binary vector.

[0088] Classifier's confidence, f Conf : The vertical distance between the sample and the decision boundary of C, which is the largest value among the predictions made by the classifier for that sample;

[0089] Conditional local accuracy,f Cond : Calculate local precision, that is, the number of correctly classified items in each category compared to the total number of items in the class θ = {x1, x2, ..., x}. k The ratio of the number of correctly classified items in};

[0090] Ambiguity, f Amb The difference between the highest probability and the second highest probability after classification;

[0091] Logarithmic, f Log : Where S lk (X k ) is the classifier C for θ={x1,x2,…,x k The support value of the correctly classified samples in the classification results of};

[0092] Entropy, f Ent : For θ = {x1, x2, ..., x kThe samples in} are calculated separately, and finally a K-dimensional vector is formed; where S l (X k This refers to the support value for misclassification;

[0093] Minimal difference, f MD :

[0094] Kullback-Leibler divergence, f KL :

[0095] Exponential, f Exp :

[0096] Randomized reference classifier,f RRC For θ = {x1, x2, ..., x k For the samples in}, the conditional probability of correct classification is calculated using RRC, resulting in a K-dimensional vector;

[0097] Simplified classifier rank,f Rank The validation set data are sorted from nearest to farthest from the samples, and classifier C is used to classify the validation set data until the first classification error occurs. The number of correctly classified data is then f. Rank ;

[0098] classifier rank OP, f RankOP : Implemented in the decision space, that is, the data of the sample and validation sets are transformed to form the output file domain, and then compared with f. Rank The same processing is used in the middle;

[0099] (5) Feature selection: In calculation Then, the FSE model from step S2 is used to select important features;

[0100] (6) Training phase: Using the features selected in step 35 Recorded as First of all The dataset is divided into 75% training data and 25% validation data. The 75% training data is used as input to train the meta classifier. During training, a prediction is made on the validation set after each training round. If the performance on the validation set decreases or fails to improve compared to the previous prediction result at a certain point, the training stops.

[0101] (7) Final prediction: The film review is predicted using the trained meta classifier. First, each classifier C in the classifier pool is used. i Calculate the film review x test Obtain its meta-features vector v test Input metaclassifier, and if the output is α test If the value is 1, then the classifier is selected as a classifier with classification ability and placed into the classifier pool P. test This process is repeated until the final classifier pool P for the film review is obtained. test And by adjusting the final classifier pool P test The prediction results of the middle classifier are used to obtain the final prediction result of the film review by majority voting.

[0102] Compared with the prior art, the present invention has the following beneficial effects:

[0103] 1. This invention applies a feature selection set framework to film review features and integrates six widely used feature selection algorithms, which effectively improves the final classification accuracy.

[0104] 2. This invention applies dynamic ensemble selection and integrates meta-learning and feature selection ensemble, thereby improving the stability and effectiveness of the film review sentiment analysis system.

[0105] 3. This invention uses Kappa value and accuracy to measure classification performance and uses Wilcoxon signed-rank as a statistical test, demonstrating its advantages in experimental results. Attached Figure Description

[0106] Figure 1 A flowchart illustrating the model method provided by this invention;

[0107] Figure 2 This is a schematic diagram of the feature selection integration method in an embodiment of the present invention;

[0108] Figure 3 This is a schematic diagram of the META-DES classification process in an embodiment of the present invention;

[0109] Figure 4 This is a schematic diagram illustrating the specific content of the 15 meta-features in an embodiment of the present invention. Detailed Implementation

[0110] To more clearly illustrate the embodiments of the present invention, specific implementation methods will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. Those skilled in the art can obtain other drawings and implementation methods based on these drawings without any creative effort.

[0111] The film review sentiment analysis method based on classifier and feature ensemble according to embodiments of the present invention includes the following steps:

[0112] Step S1: Convert the movie review (i.e., film critique) into a digital vector that a computer can recognize, corresponding to... Figure 1 Stage 1 specifically includes the following steps:

[0113] Step S1.1: Use the BOW model to extract textual features from the film reviews as the initial feature set, and convert each film review into a 0 / 1 vector according to whether the feature exists in the film review;

[0114] Step S1.2: Using the TF-IDF model, calculate the TF-IDF value of each feature in each film review, and convert each film review into a feature vector composed of TF-IDF values, such as... Figure 1 As shown in the table in stage1;

[0115] The tf-idf value is calculated as follows:

[0116]

[0117] Among them, f l Representing the l-th film review characteristic, tf l Representative characteristics of film reviews f l The frequency of occurrence in film review texts, where H represents the total number of film review texts, h l Represents all features f l The number of training texts.

[0118] Step S2: Employ the Feature Selection Ensemble (FSE) framework to automatically identify important film review features (i.e., key film review features); corresponding Figure 1 In stage 2, the specific recognition process is as follows: Figure 2 This includes the following steps:

[0119] Step S2.1: Apply Feature Selection Ensemble (FSE) to reduce the dimensionality of the high-dimensional feature vector, such as... Figure 2 As shown, the training film reviews are randomly divided into A parts, namely P1, P2, ..., P... A And the B-type feature selection algorithm (corresponding to Figure 2FM1, FM2, ..., FM B Combined with this framework, the following are the key components: IG (Information Gain), GR (Gain Ratio), CHI (Chi-square Distribution), MI (Mutual Information), Relief-F (Relevant Features), and F_classif (ANOVA F-value), calculated as follows:

[0120] IG:

[0121]

[0122] GR:

[0123] Where m represents the total number of film review categories, cl i Let f represent the i-th film review. l Representing the l-th film review characteristic, P(cl) i ) indicates that a film review belongs to cl i The probability of the class, P(f) l ) represents containing f l The probability of a film review appearing in all collections of film reviews. This means that f is not present. l The probability of a film review appearing in all film review collections.

[0124] CHI:

[0125] Where n represents the total number of film review features, and a represents the number of features included in the review. l And belongs to category cl i The number of film reviews, b indicates that it includes f. l And does not belong to category cl i The number of film reviews, where c indicates that f is not included. l And belongs to category cl i The number of film reviews, where d indicates that f is not included. l And does not belong to category cl i The number of film reviews.

[0126] Relief-F:

[0127]

[0128] The algorithm requires three stages of computation: First, the initial weights of all features are set to 0; second, the number of sampling rounds is set to m, and in each round of sampling, a sample S is randomly selected from the training set, and then the k nearest neighbor sample H is found from the samples of the same class as S. j (j = 1, 2, ..., k), find the k nearest neighbor sample M from samples of different classes from S. j (S); The third step is to update the feature weights according to the formula in each round, where W(f l () represents the weight of each feature, which is initially set to 0. Let represent other classes that do not belong to class S, P(cl) represent the probability that a sample belongs to class cl, class(S) represent the class of a randomly selected sample S, and diff(f) represent the probability that a sample belongs to class cl. l R1, R2) represent feature f l The difference between sample R1 and sample R2, R1[f] l ] represents the feature f of sample R1 l Corresponding value, R2[f l Similarly, max(f) l ) represents all film review characteristics f l The maximum value among the corresponding values, min(f) l ) is the minimum value;

[0129] Step S2.2: Sort and select the data calculated by each algorithm, and use the FSE framework to obtain the final selected feature vector;

[0130] Specifically, after B feature selection algorithms iteratively calculate A datasets respectively, each feature selection algorithm will produce A different weight vectors for the A datasets (e.g., ...). Figure 2 In the middle, FS 1,1 ,...,FS A,1 We use Vec a,b This represents the weight vector obtained by the b-th feature selection algorithm after calculating the weight vector for the a-th data point. Next, the data in all calculated weight vectors are sorted and selected, using the following calculation method:

[0131]

[0132] Where, τ a,b Indicates targeting Vec a,b The threshold obtained using the CDF-g (Cumulative Distribution Function Gradient) algorithm is then applied to the weight vector Vec. a,b In, all values ​​less than the threshold τ a,b The features corresponding to the weights are removed, FS a,bThis represents the film review features selected by the b-th feature selection algorithm after processing the a-th dataset. Then, the intersection of these A sets of feature vectors after threshold filtering is performed. Here, it corresponds to... Figure 2 In the Data Ensemble process, the final result FS of the feature selection algorithm for all training data is obtained by the b-th feature selection algorithm. b Finally, by performing a union operation on the results of all B feature selection algorithms, the result is obtained here. Figure 2 The Union Set process is used to derive the final film review features.

[0133] Step S3: Develop an improved Dynamic Ensemble Selection (DES) framework based on the META-DES framework and apply it to the training and selection stages of the classifier, and perform the final prediction. Figure 1 In stage 3, the specific prediction process is as follows: Figure 3 This includes the following steps:

[0134] Step S3.1: Establish the base classifier pool, such as... Figure 3 The Classifier Generation Process. This involves processing the feature-selected training dataset T. λ With validation dataset D SEL The partitioning is performed. An initial base classifier pool P is established, where the SVM algorithm is uniformly used for the base classifiers, and the training dataset T is used. λ Training the base classifiers involves assigning different datasets to each base classifier through bootstrapping. The differences in datasets reflect the heterogeneity between base classifiers, thus obtaining the initial base classifier pool P, as shown in the figure: C = {C1, C2, ..., C...} Q As shown in the image.

[0135] Step S3.2: Training data selection, such as... Figure 3 Sample Selection Process. For the training dataset T... λ Chinese data x j，train The process is iterated, using all base classifiers in the classifier pool P to predict the data. If the percentage of correctly classified classifiers in the base classifier pool exceeds a set threshold h, the prediction is successful. c If the film reviews are selected, the training data is selected; otherwise, they are discarded. This process is repeated until all film reviews in the training data have been filtered out. The filtered training data is T′λ.

[0136] Step S3.3: Calculate the domain θ and the output file domain First, the k-nn algorithm is used, with each classifier C in the classifier pool P applied to the training dataset T′. λ Chinese data x j，trainCalculate its neighborhood θ. First, iterate through the classifiers in the classifier pool P. For a certain classifier C, then iterate through T′. λ The classifier iterates through all the data and calculates x. j，train The domain θ. θ is derived from the validation dataset D. SEL The data is composed of the K closest samples, θ = {x1, x2, ..., x...} k Next, x j，train Convert to output file Similarly, we transform all the data in the validation set, setting it to... (The n-dimensional feature data is transformed into a vector with the dimension equal to the number of label categories. This vector consists of predicted values. For example, for a three-class classification problem, this classifier C is used to predict the data, resulting in a prediction result of [0.3, 0.4, 0.3]. This predicted result vector is the transformed data. Similarly, this classifier is used to transform all data in the validation set.) Then, the k-nn algorithm is used for computation. Depend on Zhongyu Nearest K p Composed of 1 sample,

[0137] Step S3.4: Create Meta-features, such as Figure 3 The Meta-Feature Extraction process. Based on S3.3, classifier C is used for training data x. j，train This allows us to obtain the domain θ and the output file domain. Then, 15 meta-features are calculated (these 15 meta-features represent 15 different evaluation criteria; even if one criterion fails due to imprecision in local regions of the feature space or due to low confidence, the classification system can still achieve good performance by selecting other meta-features) and stored in a vector v. i,j ={f1,f2,f3,…,f 15 In this process, if the classifier predicts correctly, the class attribute α = 1; otherwise, it is 0. Then, the base classifier P in the classifier pool and the data in the training data T′λ are iterated continuously, resulting in a vector v consisting of all 15 meta-features. i,j The set is composed of the number of base classifiers, Q×N (training data T′). λ Meta-training data (in terms of the amount of data) As training data for the metaclassifier needed for subsequent base classification, the label of each training data is the corresponding α;

[0138] The 15 meta-features are as follows, specifically: Figure 4 :

[0139] F1(neighbors'hard classification:f Hard ): Use base classifier C for θ={x1,x2,…,x K The vector is classified, and if the classification is correct, it is 1; otherwise, it is 0, generating a K-dimensional binary vector.

[0140] F2(Posterior probability:f Prob : Posterior probability, P(class|x) k This generates a K-dimensional binary vector. (This is the support vector for the correctly classified class; for example, if the prediction result for a three-class classification is [0.2, 0.3, 0.5], then this value is 0.5.)

[0141] F3 (Overall local accuracy: f) Overall ): Throughout the entire θ = {x1, x2, ..., x k In the equation, F3 represents the accuracy of the base classifier C. (F3 is a probability value.)

[0142] F4(Output profiles classification:f Op The logic is the same as F1, but for... Conduct training.

[0143] F5 (Classifier's confidence:f) Conf ): The vertical distance between the sample and the decision boundary of C, which is the largest value in the classifier's prediction results for that sample.

[0144] F6(Conditional local accuracy:f Cond ): Calculate local precision, i.e., the number of correctly classified items in each category compared to the total number of items in the class θ = {x1, x2, ..., x}. k The ratio of the number of correctly classified items in}.

[0145] F7(Ambiguity:f Amb Assuming it's a three-class classification, with class 1: 0.65, class 2: 0.30, and class 3: 0.05, then f Amb =0.65 - 0.3 = 0.35 (the difference between the highest probability and the second highest probability).

[0146] F8(Logarithmic:f Log ): Where S lk (X k ) is the classifier C for this sample (θ={x1,x2,…,x... k The support value for correctly classified samples in the classification results of the sample in the table. (Assuming it is a three-class classification, and the sample belongs to class 1, the prediction results are class 1: 0.65, class 2: 0.30, class 3: 0.05, then S lk (X k The value is 0.65, where L is the total number of categories, and 3 categories is 3. Finally, a K-dimensional vector is obtained.

[0147] F9(Entropy:f Ent ): The calculations are performed separately, ultimately resulting in a K-dimensional vector. Here, S... l (X k ) refers to the support value for misclassification. (θ={x1,x2,…,x k (samples in})

[0148] F10 (Minimal difference:f) MD ):

[0149] F11(Kullback-Leibler divergence:f KL ):

[0150] F12(Exponential:f Exp ):

[0151] F13(Randomized reference classifier:f RRC For θ = {x1, x2, ..., x} k The samples in the array are used to calculate the conditional probability of correct classification through RRC (Randomized Reference Classification), resulting in a K-dimensional vector.

[0152] F14(Simplified classifier rank:f Rank The validation set data is sorted from nearest to farthest from the samples, and classifier C is used to classify the validation set data until the first classification error occurs. The number of correctly classified data is then f. Rank .

[0153] F15 (classifier rank OP:f) RankOP ): The logic is the same as F14, but it is implemented in the decision space, that is, the data of the sample and validation sets are transformed to form the output file domain, and then the same processing as in F14 is performed.

[0154] Step S3.5: Feature selection. Calculated in S3.3. Subsequently, to address the issue that the large number of data features may significantly impact operational efficiency and classification performance, we again use FSE from step S2 to select important features.

[0155] Step S3.6: Training phase, such as Figure 3 Meta-Training Process. Using feature selection. This is recorded as First of all The dataset is divided into 75% training data and 25% validation data. The 75% training data is used as input to train the meta classifier (Multi-Layer Perceptron neural network). During the training process, a prediction is made on the validation set after each training round. If the performance on the validation set decreases or fails to improve compared to the previous prediction result at a certain point, the training stops.

[0156] Step S3.7: Predict movie reviews using the trained meta-classifier. First, use each classifier C from the classifier pool. i Calculate the film review x test Obtain its meta-features vector v test Input the meta classifier, and if the output is α test If the value is 1, then the classifier is selected as a classifier with classification ability and placed into the classifier pool P. test By iterating in this way, we can obtain the final classifier pool P for this film review. test And by adjusting the final classifier pool P test The prediction results of the classifier are used to obtain the final prediction result of the film review by majority voting.

[0157] Based on the film review sentiment analysis method described in the embodiments of the present invention, the embodiments of the present invention also provide a film review sentiment analysis system based on classifiers and feature integration, including:

[0158] Module M1 (i.e., the conversion module): Converts the movie reviews into digital vectors that computers can recognize. The specific conversion process can be found in step S1 above, and will not be repeated here.

[0159] Module M2 (i.e., the recognition module): This module uses the Feature Selection Ensemble (FSE) framework to automatically identify important film review features. The specific recognition process can be found in step S2 above and will not be elaborated upon here.

[0160] Module M3 (i.e., the prediction module): Based on the META-DES framework, an improved Dynamic Ensemble Selection (DES) framework is developed and applied to the training and selection phases of the classifier, and then used for the final prediction. The specific prediction process can be found in step S3 above, and will not be elaborated here.

[0161] Specifically, for ease of understanding and implementation of this invention, the sentiment scale dataset of the movie review dataset (http: / / www.cs.cornell.edu / people / pabo / movie-review-data / ) is used as an example. This dataset contains four corpora of movie reviews. In each corpus, each review is assigned a sentiment class based on 3 categories and a sentiment class based on 4 categories. Therefore, these four corpora can be divided into eight distinct data subsets, denoted as Data subset 1, Data subset 2, ..., Data subset 8. See Table 2 for details. In Table 2, Cla* represents the number of sentiment classes assigned to the text. Text* represents the total number of texts in each data subset. N Cl1 N Cl2 N Cl3 N Cl4 These represent the number of texts for each emotion category.

[0162] Table 1. Details of the experimental data.

[0163]

[0164] For each dataset, a training set (75%) and a validation set (25%) need to be split. Each dataset is then randomly and independently split 20 times, and the average of the splits is taken as the final prediction result for that dataset. The classification model is set to K=7, as used in improvedMETA-DES for calculating θ. The time frame uses K=5, the number of base classifiers in the base classifier pool P is set to 100, the base classifier uses SVM, and the threshold h is... c =0.5, when obtaining meta-training training data Then, it was divided into 75% training and 25% validation, and a Multi-Layer Perceptron neural network (containing 10 neurons in the hidden layer) was used as the meta-classifier.

[0165] For evaluating the experimental results of this invention, we selected accuracy and kappa value as indicators, and we used Wilcoxon signed-rank as a nonparametric statistical test.

[0166] To demonstrate the effectiveness of feature selection in film review texts, we compared FSE with six individual methods within FSE (IG, GR, CHI, MI, Relief-F, and ANOVA F-value). In contrast, all seven methods used META-DES as their classifiers, and we selected a threshold of 10% for each of the six individual feature selection algorithms for a fair comparison. The classification accuracy of different feature selection methods is shown in Table 2, and the Kappa values ​​are shown in Table 4. The best results for each dataset are highlighted in bold.

[0167] To arrive at meaningful conclusions, we employed appropriate statistical tests. We performed Wilcoxon signed-rank tests on pairwise comparisons between the proposed FSE-META-DES and other META-DES-based methods. The results of the Wilcoxon signed-rank tests are shown in Tables 3 (accuracy) and 4 (Kappa values), respectively.

[0168] Table 2. Average classification accuracy of FSE-META-DES and other META-DES-based single feature selection algorithms

[0169]

[0170]

[0171] Table 3 shows the Wilcoxon signed-rank test results (classification accuracy) comparing FSE-META-DES with a single feature selection algorithm based on META-DES.

[0172]

[0173] Table 4. Average Kappa values ​​of FSE-META-DES and other META-DES-based single feature selection algorithms.

[0174]

[0175]

[0176] Table 5 shows the Wilcoxon signed-rank test results (Kappa values) comparing FSE-META-DES with a single feature selection algorithm based on META-DES.

[0177]

[0178]

[0179] In comparing accuracy and Kappa value, using FSE as the feature selection method is significantly superior to other single feature selection algorithms. In the Wilcoxon signed-rank test results, R... + R represents the sum of FSE-meta-des. - This represents the sum of the ranks of the other methods. Observing these tables, the Wilcoxon test has very low accuracy, with the null hypothesis in 4 out of 6 cases (except for CHI-META-DES). In the case of CHI-META-DES, although the test is not rejected, we can still emphasize the behavior of FSE-META-DES because the ranking favors FSE-META-DES, and the corresponding p-value is low. Similar to the kappa performance measurement, it is again only the case of CHI-META-DES that is not rejected, and FSE-META-DES has a low p-value and a high ranking in the comparison results with CHI-META-DES. Therefore, the results demonstrate the superiority of FSE over other single filters.

[0180] To analyze whether the META-DES ensemble system outperforms other methods within the same series, we compared META-DES with seven different classification methods (KNORA-E, KNORA-U, LCA, OLA, MCB, RANK, and DESP), using FSE for fair comparison on text features in all cases. The comparisons of classification accuracy and Kappa scores are shown in Tables 6 and 8, respectively, with the best results for each dataset highlighted in bold. The corresponding Wilcoxon signed-rank tests are shown in Tables 7 and 9.

[0181] Table 6. Average classification accuracy of FSE-META-DES and other FSE-based DES algorithms

[0182]

[0183]

[0184] Table 7 shows the Wilcoxon signed-rank test results (classification accuracy) comparing FSE-META-DES with other DES-based algorithms.

[0185]

[0186] Table 8. Average Kappa values ​​of FSE-META-DES and other FSE-based DES algorithms.

[0187]

[0188]

[0189] Table 9 shows the Wilcoxon signed-rank test results (Kappa values) comparing FSE-META-DES with other DES-based algorithms.

[0190]

[0191] Observing these tables, the FSE-based META-DES is significantly superior to other methods in the same series as META-DES in terms of mean accuracy, mean Kappa value, and the number of best-performing datasets. Furthermore, to draw meaningful conclusions, we still performed Wilcoxon signed-rank tests on pairwise comparisons between META-DES and other methods. The results of the statistical tests are shown in Table 7 (classification accuracy) and Table 9 (Kappa value). Based on these tables, we can easily see that all equivalence hypotheses for the two performance metrics were rejected, with low p-values. In other words, the FSE-based META-DES is significantly better than other methods.

[0192] In the meta-DES framework, meta-features are key to determining the performance of the meta-classifier λ. In this method, we expand the number of meta-feature types from 5 to 15, and apply the FSE framework to the expanded meta-features. To analyze whether the improved META-DES ensemble system outperforms the initial META-DES, we still conducted comparative experiments. Classification accuracy and Kappa values ​​are shown in Tables 10 and 12, and the corresponding Wilcoxon signed-rank tests are shown in Tables 11 and 13.

[0193] Table 10 Comparison of average classification accuracy of improved META-DES, META-DES and other advanced methods.

[0194]

[0195]

[0196] Table 11 shows the Wilcoxon signed-rank test results for improved META-DES, META-DES, and other advanced methods.

[0197] Classification accuracy

[0198]

[0199] Table 12 Comparison of mean kappa values ​​of improved META-DES, META-DES, and other advanced methods.

[0200]

[0201]

[0202] Table 13 shows the Wilcoxon signed-rank test results (kappa values) comparing improved META-DES, META-DES, and other state-of-the-art methods.

[0203]

[0204]

[0205] As can be seen from the tables, the improved META-DES and the original META-DES framework perform almost identically in terms of accuracy and kappa value. To obtain clearer and more convincing comparison results, we still use Wilcoxon signed rank to perform pairwise comparisons between META-DES and the improved META-DES. From both tables, no comparisons for accuracy and kappa were rejected; the p-values ​​for accuracy were all below 0.5, while the p-values ​​for kappa were all above 0.5. Furthermore, the rankings between the two tables did not show any trend. Therefore, the difference between the two methods is not significant. Based on the results, our proposed method performs at least as well as the original META-DES. Meanwhile, we used FSE in the meta-features extraction process of the improved META-DES, so our proposed method still has some advantages in terms of operational efficiency.

[0206] To analyze the differences between our proposed improved META-DES and state-of-the-art sentiment classification methods, we compared our improved method with five well-known benchmark methods, including OVA-SVM, MCST-SVM with Culling, MCST-SVM without Culling, Regression, and Improved OVO. For a fair comparison, we followed the recommendations of the original paper in setting the parameters of the selected methods. The full results of these comparisons are shown in Tables 10 (Classification Accuracy) and 12 (Kappa Values), where the best results for each dataset are highlighted in bold. Overall, the proposed method outperforms the other methods in terms of average classification accuracy and Kappa value, and is at least as good as some of the best-performing methods. Statistical test results are shown in Tables 11 (Classification Accuracy) and 13 (Kappa Values). In both tables, the Wilcoxon test rejected the null hypothesis of accuracy equivalence in 4 out of 6 cases (except for MCST and IG+OVO+SVM), but the corresponding p-values ​​were very low, and the improved META-DES had a higher rank. Therefore, the results demonstrate the superiority of the improved META-DES over the other five advanced methods.

[0207] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention. For those skilled in the art, there may be changes in specific implementation methods based on the ideas provided by the present invention, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A film review sentiment analysis method based on classifiers and feature ensemble, characterized in that, Includes the following steps: S1. Convert film reviews into recognizable digital vectors; S2. Integrate the identifiable digital vector input features into the FSE model to identify key film review features; S3. Input key film review features into a dynamic ensemble selection DES model based on the META-DES framework for sentiment analysis and prediction. Step S3 includes the following steps: S31. Establish the base classifier pool: Train the base classifier pool using the feature-selected training dataset T. λ With validation dataset D SEL The data is divided into two parts; an initial base classifier pool P is established, and the base classifiers uniformly use the SVM algorithm, with the training dataset T as the input. λ The base classifiers are trained by sampling different datasets and assigning them to each base classifier. The heterogeneity between the base classifiers is reflected by the differences in the datasets, and thus the initial base classifier pool P is obtained. S32. Training Data Filtering: Filtering the training dataset T λ Chinese data x j，train The process iterates, using all base classifiers in the initial base classifier pool P to predict the data; if the percentage of correctly classified classifiers in the base classifier pool exceeds a set threshold h... c If the film reviews are selected, the training data is chosen; otherwise, they are discarded. This process is repeated until all film reviews in the training data have been filtered out. The filtered training data is then T′. λ ; S33: Computational domain θ and output file domain Using the k-nn algorithm, each classifier C in the initial base classifier pool P is used to train the dataset T′. λ Chinese data x j，train Calculate its domain θ and the domain of the output file First, iterate through the classifiers in the classifier pool P, and for a given classifier C... i Then for T′ λ The classifier iterates through all the data and calculates x. j，train The domain θ; θ is derived from the validation dataset D SEL The data is composed of the K closest samples, θ = {x1, x2, ..., x...} k Next, x j，train Convert to output file Similarly, all data in the validation set are transformed and set to... Then, the k-nn algorithm is used for calculation. Depend on Zhongyu Nearest K p Composed of 1 sample, S34. Establish Meta-features: Classifier C i For training data x j，train This yields the domain θ = {x1, x2, ..., x}. k } and output file fields Then, 15 meta-features are calculated and stored in a vector v. i,j ={f1,f2,f3,…,f 15 In the process, if the classifier predicts correctly, the class attribute α = 1; otherwise, it is 0. Then, the base classifier P in the classifier pool is continuously compared with the training data T′. λ Iterate through the data to generate vectors v consisting of all 15 meta-features. i,j The data is aggregated into meta-training data with dimensions Q×N, representing the number of base classifiers. As training data for the meta classifier needed for subsequent base classification, the label of each training data is the corresponding α; S35, Feature Selection: In calculation Then, the FSE model from step S2 is used to select important features; S36. Training Phase: Using the features selected in step 35 Recorded as First of all The dataset is divided into 75% training data and 25% validation data. The 75% training data is used as input to train the meta classifier. During training, a prediction is made on the validation set after each training round. If the performance on the validation set decreases or fails to improve compared to the previous prediction result at a certain point, the training stops. S37. Final Prediction: Predict the film review using the trained meta-classifier. First, use each classifier C from the classifier pool. i Calculate the film review x test Obtain its meta-features vector v test Input meta classifier, and if the output is α test If the value is 1, then the classifier is selected as a classifier with classification ability and placed into the classifier pool P. test This process is repeated until the final classifier pool P for the film review is obtained. test And by adjusting the final classifier pool P test The prediction results of the middle classifier are used to obtain the final prediction result of the film review by majority voting.

2. The film review sentiment analysis method according to claim 1, characterized in that, Step S1 includes the following steps: S11. The BOW model is used to extract textual features from the film reviews as the initial feature set, and each film review is converted into a vector of 0 or 1 based on whether the feature exists in the film review. S12. Using the TF-IDF model, calculate the TF-IDF value of each feature in each film review, and convert each film review into a feature vector composed of TF-IDF values.

3. The film review sentiment analysis method according to claim 2, characterized in that, The method for calculating the tf-idf value is as follows: Among them, f l Representing the l-th film review characteristic, tf l Representative characteristics of film reviews f l The frequency of occurrence in film review texts, where H represents the total number of film review texts, h l Represents all features f l The number of training texts.

4. The film review sentiment analysis method according to claim 3, characterized in that, Step S2 includes the following steps: S21. The Feature Selection Ensemble (FSE) model is applied to reduce the dimensionality of the high-dimensional feature vectors. The training film reviews are randomly divided into A parts, and B feature selection algorithms are combined with the Feature Selection Ensemble (FSE) model. These algorithms are Information Gain (IG), Gain Ratio (GR), Chi-Square Distribution (CHI), Mutual Information (MI), Relief-F algorithm, and F-test. The calculation methods are as follows: IG: GR: Where m represents the total number of film review categories, cl i Let f represent the i-th film review. l Representing the l-th film review characteristic, P(cl) i ) indicates that a film review belongs to cl i The probability of the class, P(f) l ) represents containing f l The probability of a film review appearing in all collections of film reviews. This means that f is not present. l The probability of a film review appearing in all film review collections; WHO: WHO(f l )=max i (WHO(f l ,Cl i )) Where n represents the total number of film review features, and a represents the number of features included in the review. l And belongs to category cl i The number of film reviews, b indicates that it includes f. l And does not belong to category cl i The number of film reviews, where c indicates that f is not included. l And belongs to category cl i The number of film reviews, where d indicates that f is not included. l And does not belong to category cl i The number of film reviews; Relief-F: The Relief-F algorithm requires three stages of computation: First, the initial weights of all features are set to 0; second, the number of sampling rounds is set to m, and in each round, a sample S is randomly selected from the training set, and then the k nearest neighbor sample H is found from the samples of the same class as S. j (j = 1, 2, ..., k), find the k nearest neighbor sample M from samples of different classes from S. j (S); The third step is to update the feature weights according to the formula in each round, where w(f l () represents the weight of each feature, initially set to 0. Let represent other classes that do not belong to class S, P(cl) represent the probability that a sample belongs to class cl, class(S) represent the class of a randomly selected sample S, and diff(f) represent the probability that a sample belongs to class cl. l R1, R2) represent feature f l The difference between sample R1 and sample R2, R1[f] l ] represents the feature f of sample R1 l Corresponding value, R2[f l Similarly, max(f) l ) represents all film review characteristics f l The maximum value among the corresponding values, min(f) l ) is the minimum value; S22. After B feature selection algorithms iteratively calculate A datasets respectively, each feature selection algorithm will produce A different weight vectors for the A datasets. Using Vec... a,b This represents the weight vector obtained by the b-th feature selection algorithm after calculating the weight vector for the a-th data point. Next, the data in all calculated weight vectors are sorted and selected, using the following calculation method: Where, τ a,b Indicates targeting Vec a,b The threshold obtained using the CDF-g algorithm is then applied to the weight vector Vec. a,b In, all values ​​less than the threshold τ a,b The features corresponding to the weights are removed, FS a,b This represents the film review features selected by the b-th feature selection algorithm after processing the a-th dataset. Then, the intersection of these A threshold-filtered feature vector sets yields the final feature selection result FS of the b-th feature selection algorithm across all training data. b Finally, the final key film review features are obtained by performing a union operation on the results of all B feature selection algorithms.

5. The film review sentiment analysis method according to claim 4, characterized in that, The 15 meta-features are as follows: neighbors'hard classification,f Hard Using a base classifier C for θ = {x1, x2, ..., x} K Classify the data; if the classification is correct, return 1; otherwise, return 0, generating a K-dimensional binary vector. Posterior probability, f Prob : Posterior probability, P(class|x) k Generate a K-dimensional binary vector; Overall local accuracy, f Overall : Throughout the entire θ = {x1, x2, ..., x k In the context of classification, the accuracy of the base classifier C is considered. Output profiles classification:f Op Using base classifier C Classify the data; if the classification is correct, return 1; otherwise, return 0, generating a K-dimensional binary vector. Classifier's confidence, f Conf : The vertical distance between the sample and the decision boundary of C, which is the largest value among the predictions made by the classifier for that sample; Conditional local accuracy,f Cond : Calculate local precision, that is, the number of correctly classified items in each category compared to the total number of items in the class θ = {x1, x2, ..., x}. k The ratio of the number of correctly classified items in}; Ambiguity, f Amb The difference between the highest probability and the second highest probability after classification; Logarithmic, f Log : Where S lk (X k ) is the classifier C for θ={x1,x2,…,x k The support value of the correctly classified samples in the classification results of}; Entropy, f Ent : For θ = {x1, x2, ..., x k The samples in} are calculated separately, and finally a K-dimensional vector is formed; where S l (X k This refers to the support value for misclassification; Minimal difference,f MD : Kullback-Leibler divergence,f KL : Exponential,f Exp : Randomized reference classifier,f RRC For θ = {x1, x2, ..., x} k For the samples in}, the conditional probability of correct classification is calculated using RRC, resulting in a K-dimensional vector; Simplified classifier rank,f Rank The validation set data are sorted from nearest to farthest from the samples, and classifier C is used to classify the validation set data until the first classification error occurs. The number of correctly classified data is then f. Rank ; classifier rank OP, f RankOP : Implemented in the decision space, that is, the data of the sample and validation sets are transformed to form the output file domain, and then compared with f. Rank The same processing is used in the middle.

6. A film review sentiment analysis system based on classifiers and feature ensemble, characterized in that, include: The conversion module is used to convert film reviews into recognizable digital vectors. The recognition module is used to select and integrate identifiable digital vector input features into the FSE model to identify key film review features; The prediction module is used to input key film review features into a dynamic ensemble selection DES model based on the META-DES framework for sentiment analysis prediction. The prediction process of the prediction module includes: (1) Establish a base classifier pool: Train the training dataset T after feature selection. λ With validation dataset D SEL The data is divided into two parts; an initial base classifier pool P is established, and the base classifiers uniformly use the SVM algorithm, with the training dataset T as the input. λ The base classifiers are trained by sampling different datasets and assigning them to each base classifier. The heterogeneity between the base classifiers is reflected by the differences in the datasets, and thus the initial base classifier pool P is obtained. (2) Training data selection: For the training dataset T λ Chinese data x j，train The process iterates, using all base classifiers in the initial base classifier pool P to predict the data; if the percentage of correctly classified classifiers in the base classifier pool exceeds a set threshold h... c If the film reviews are selected, the training data is chosen; otherwise, they are discarded. This process is repeated until all film reviews in the training data have been filtered out. The filtered training data is then T′. λ ; (3) Calculation domain θ and output file domain Using the k-nn algorithm, each classifier C in the initial base classifier pool P is used to train the dataset T′. λ Chinese data x j，train Calculate its domain θ and the domain of the output file First, iterate through the classifiers in the classifier pool P, and for a given classifier C... i Then for T′ λ The classifier iterates through all the data and calculates x. j，train The domain θ; θ is derived from the validation dataset D SEL The data is composed of the K closest samples, θ = {x1, x2, ..., x...} k Next, x j，train Convert to output file Similarly, all data in the validation set are transformed and set to... Then, the k-nn algorithm is used for calculation. Depend on Zhongyu Nearest K p Composed of 1 sample, (4) Establish Meta-features: Classifier C i For training data x j，train This yields the domain θ = {x1, x2, ..., x}. k } and output file fields Then, 15 meta-features are calculated and stored in a vector v. i,j ={f1,f2,f3,…,f 15 In the process, if the classifier predicts correctly, the class attribute α = 1; otherwise, it is 0. Then, the base classifier P in the classifier pool is continuously compared with the training data T′. λ Iterate through the data to generate vectors v consisting of all 15 meta-features. i,j The data is aggregated into meta-training data with dimensions Q×N, representing the number of base classifiers. As training data for the meta classifier needed for subsequent base classification, the label of each training data is the corresponding α; (5) Feature selection: In calculation Then, the FSE model from step S2 is used to select important features; (6) Training phase: Using the features selected in step 35 Recorded as First of all The dataset is divided into 75% training data and 25% validation data. The 75% training data is used as input to train the meta classifier. During training, a prediction is made on the validation set after each training round. If the performance on the validation set decreases or fails to improve compared to the previous prediction result at a certain point, the training stops. (7) Final prediction: The film review is predicted using the trained meta classifier. First, each classifier C in the classifier pool is used. i Calculate the film review x test Obtain its meta-features vector v test Input the meta classifier, and if the output is α test If the value is 1, then the classifier is selected as a classifier with classification ability and placed into the classifier pool P. test This process is repeated until the final classifier pool P for the film review is obtained. test And by adjusting the final classifier pool P test The prediction results of the middle classifier are used to obtain the final prediction result of the film review by majority voting.

7. The film review sentiment analysis system according to claim 6, characterized in that, The process of converting film reviews into recognizable digital vectors includes the following steps: (1) The text features in the film reviews are extracted using the BOW model as the initial feature set, and each film review is converted into a vector of 0 or 1 based on whether the feature exists in the film review; (2) Using the TF-IDF model, calculate the TF-IDF value of each feature in each film review, and convert each film review into a feature vector composed of TF-IDF values.

8. The film review sentiment analysis system according to claim 7, characterized in that, The method for calculating the tf-idf value is as follows: Among them, f l Representing the l-th film review characteristic, tf l Representative characteristics of film reviews f l The frequency of occurrence in film review texts, where H represents the total number of film review texts, h l Represents all features f l The number of training texts.

9. The film review sentiment analysis system according to claim 8, characterized in that, The identification process of the identification module includes: (1) The feature selection ensemble FSE model is applied to reduce the dimensionality of the huge feature vector. The training film reviews are randomly divided into part A, and the feature selection algorithms B are combined with the feature selection ensemble FSE model. The algorithms are information gain IG, gain ratio GR, chi-square distribution CHI, mutual information MI, Relief-F algorithm, and F test. The calculation methods are as follows: IG: GR: Where m represents the total number of film review categories, cl i Let f represent the i-th film review. l Representing the l-th film review characteristic, P(cl) i ) indicates that a film review belongs to cl i The probability of the class, P(f) l ) represents containing f l The probability of a film review appearing in all collections of film reviews. This means that f is not present. l The probability of a film review appearing in all film review collections; CHI: CH(f l ) = max i (CHI(f l ,Cl i Where n represents the total number of film review features, and a represents the number of features included in the review. l And belongs to category cl i The number of film reviews, b indicates that it includes f. l And does not belong to category cl i The number of film reviews, where c indicates that f is not included. l And belongs to category cl i The number of film reviews, where d indicates that f is not included. l And does not belong to category cl i The number of film reviews; Relief-F: The Relief-F algorithm requires three stages of computation: First, the initial weights of all features are set to 0; second, the number of sampling rounds is set to m, and in each round, a sample S is randomly selected from the training set, and then the k nearest neighbor sample H is found from the samples of the same class as S. j (j = 1, 2, ..., k), find the k nearest neighbor sample M from samples of different classes from S. j (S); The third step is to update the feature weights according to the formula in each round, where W(f l () represents the weight of each feature, initially set to 0. Let represent other classes that do not belong to class S, P(cl) represent the probability that a sample belongs to class cl, class(S) represent the class of a randomly selected sample S, and diff(f) represent the probability that a sample belongs to class cl. l R1, R2) represent feature f l The difference between sample R1 and sample R2, R1[f] l ] represents the feature f of sample R1 l Corresponding value, R2[f l Similarly, max(f) l ) represents all film review characteristics f l The maximum value among the corresponding values, min(f) l ) is the minimum value; (2) After B feature selection algorithms iteratively calculate A datasets respectively, each feature selection algorithm will produce A different weight vectors for the A datasets. (Using Vec...) a,b This represents the weight vector obtained by the b-th feature selection algorithm after calculating the weight vector for the a-th data point. Next, the data in all calculated weight vectors are sorted and selected, using the following calculation method: Where, τ a,b Indicates targeting Vec a,b The threshold obtained using the CDF-g algorithm is then applied to the weight vector Vec. a,b In, all values ​​less than the threshold τ a,b The features corresponding to the weights are removed, FS a,b This represents the film review features selected by the b-th feature selection algorithm after processing the a-th dataset. Then, the intersection of these A threshold-filtered feature vector sets yields the final feature selection result FS of the b-th feature selection algorithm across all training data. b Finally, the final key film review features are obtained by performing a union operation on the results of all B feature selection algorithms.

10. The film review sentiment analysis system according to claim 9, characterized in that, The 15 meta-features are as follows: neighbors'hard classification,f Hard Using a base classifier C for θ = {x1, x2, ..., x} K Classify the data; if the classification is correct, return 1; otherwise, return 0, generating a K-dimensional binary vector. Posterior probability, f Prob : Posterior probability, P(class|x) k Generate a K-dimensional binary vector; Overall local accuracy, f Overall : Throughout the entire θ = {x1, x2, ..., x k In the context of classification, the accuracy of the base classifier C is considered. Output profiles classification:f Op Using base classifier C Classify the data; if the classification is correct, return 1; otherwise, return 0, generating a K-dimensional binary vector. Classifier's confidence, f Conf : The vertical distance between the sample and the decision boundary of C, which is the largest value among the predictions made by the classifier for that sample; Conditional local accuracy,f Cond : Calculate local precision, that is, the number of correctly classified items in each category compared to the total number of items in the class θ = {x1, x2, ..., x}. k The ratio of the number of correctly classified items in}; Ambiguity, f Amb The difference between the highest probability and the second highest probability after classification; Logarithmic, f Log : Where S lk (X k ) is the classifier C for θ={x1,x2,…,x k The support value of the correctly classified samples in the classification results of}; Entropy, f Ent : For θ = {x1, x2, ..., x k The samples in} are calculated separately, and finally a K-dimensional vector is formed; where S l (X k This refers to the support value for misclassification; Minimal difference,f MD : Kullback-Leibler divergence,f KL : Exponential,f Exp : Randomized reference classifier,f RRC For θ = {x1, x2, ..., x} k For the samples in}, the conditional probability of correct classification is calculated using RRC, resulting in a K-dimensional vector; Simplified classifier rank,f Rank The validation set data are sorted from nearest to farthest from the samples, and classifier C is used to classify the validation set data until the first classification error occurs. The number of correctly classified data is then f. Rank ; classifier rank OP, f RankOP : Implemented in the decision space, that is, the data of the sample and validation sets are transformed to form the output file domain, and then compared with f. Rank The same processing is used in the middle.

Images