Spam SMS classification method and device based on pseudo-label fusion clustering
Through the method based on pseudo-label fusion clustering, combined with TF-IDF and clustering model, the accuracy and coverage of spam message classification are solved, efficient identification of spam message and rapid acquisition of effective information are achieved, and noise is reduced Influence.
Patent Information
- Application Number
- CN202410565826.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-09
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-05-09
AI Technical Summary
The existing technology is difficult to effectively identify and classify spam messages, which makes it difficult for users to quickly obtain effective information in massive text messages, and there is a risk of being cheated.
Using a method based on pseudo-label fusion clustering, a spam message classification model is constructed, and a fixed number knowledge base and a pseudo-label training model is used, combining TF-IDF algorithm, clustering model and data enhancement algorithm to reduce noise and improve classification accuracy.
It realizes accurate and efficient identification of spam messages, improves the coverage rate of spam messages fixed numbers in all numbers, improves the generalization ability and accuracy of the classification model, and reduces pseudo-label noise.
Smart Images

Figure CN118503419B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of communication data processing, and more particularly to a junk text message classification method based on pseudo-label fusion clustering. Background Art
[0002] At present, with the continuous development of communication technology, whether it is express delivery, popular activities, withdrawal and transfer, or business promotion, most of the information is transmitted by sending text messages. More and more companies tend to establish service communication with users through text messages.
[0003] However, the text messages received by users include a large number of spam text messages that are sent without permission and involve commercial promotion, telecommunications fraud, pornographic violence, reactionary propaganda, etc.; not only does this result in the omission of valid information and hinder effective communication between users and companies; it also causes distress to users and increases insecurity, such as the risk of being deceived.
[0004] Therefore, how to classify junk text messages and help users quickly identify valid information is a problem that technical personnel in this field need to solve urgently. Summary of the invention
[0005] In view of this, the present invention provides a junk SMS classification method and device based on pseudo-label fusion clustering, which can realize SMS text classification when labeled data is scarce.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] A method for classifying spam text messages based on pseudo-label fusion clustering includes the following steps:
[0008] Build a spam SMS classification model;
[0009] Determine the fixed number knowledge base, and train the spam SMS classification model based on pseudo-label fusion clustering according to the data set corresponding to the fixed number;
[0010] Use the trained spam message classification model to classify spam messages.
[0011] Preferably, the process of determining the fixed number knowledge base includes:
[0012] Obtain massive amounts of SMS data and unify them into the following format: number, corresponding organization name, and number of occurrences within a set period;
[0013] The management bit in the number is extracted, a fixed number segment and a corresponding organization name are determined based on the management bit, and a fixed number knowledge base is determined according to the fixed number segment and the corresponding organization name.
[0014] Preferably, the method for determining the fixed number segment and the corresponding organization name based on the management bit is:
[0015] For service-type SMS data, the data in a unified format are grouped according to the management position. If there is only one organization in the group, the only organization is the organization corresponding to the current management position; if there are multiple organizations in the group, the organization corresponding to the management position is the organization corresponding to the current management position;
[0016] Preferably, for service platform SMS data, the data in a unified format are grouped according to the management bit. If there is only one organization in the group, it is determined whether the number of times the unique organization appears within a set period is greater than n. If so, the unique organization is the organization corresponding to the current management bit.
[0017] If there are multiple organizations in the group, the data is deduplicated, and it is determined whether the number of times the remaining organizations appear within the set period is greater than n. If so, the remaining organizations are used as the organizations corresponding to the current management position.
[0018] Preferably, before training, the data set corresponding to the fixed number is dimensionality reduced and vectorized based on the TF-IDF algorithm to obtain the more important features in the text messages, and the sparseness problem of high-dimensional vectors is solved by dimensionality reduction.
[0019] Preferably, the process of training the spam message classification model based on pseudo-label fusion clustering includes:
[0020] Dividing the data set corresponding to the fixed number into labeled data and unlabeled data, and determining the true label of the labeled data;
[0021] Based on the labeled data and the corresponding true labels, the clustering model is used to obtain the pseudo labels of the unlabeled data;
[0022] Using the real label and the pseudo label data, the spam message classification model is trained according to the following loss function:
[0023]
[0024] Where, X P represents a pseudo-label dataset, X L represents the real label data set; θ1, θ2 are model parameters, y * represents the predicted label, Represents the output distribution of the clustering task corresponding to the dataset generated by back-translation of the unlabeled dataset.
[0025] As a preference, the labeled data is enhanced by a data enhancement algorithm and evaluated by a relative entropy maximization mechanism and a conditional entropy minimization mechanism. The evaluation function is:
[0026]
[0027] In the formula, represents the jth enhanced sample generated by the i-th real sample, represents the i-th true sample, y i represents the true label, n is the number of real sample sets, h is the number of enhanced samples generated for each real sample, φ:R d →R D represents feature mapping, w is a hyperparameter, H is Shannon entropy, which is used to represent the information gain of enhanced samples, D is relative entropy, I is mutual information, and p is probability distribution. Used to calculate and p(y i ) is the difference between .
[0028] Preferably, back translation is used to perform sample enhancement on unlabeled data, and the metric between the two sample distributions is minimized by a consistency loss function, and the consistency loss function is as follows:
[0029]
[0030] Where θ is a model parameter, is the data enhancement function, X u represents an unlabeled dataset, is the pseudo-label dataset distribution, represents the distribution of data augmentation samples output by the clustering algorithm, D KL represents the distance between two distributions measured by KL divergence, and E represents expectation.
[0031] Preferably, the spam message classification model is trained using the real label and the pseudo label according to the following loss function:
[0032] L′=L+L u +L E .
[0033] A junk SMS fixed number classification device based on pseudo-label fusion clustering, comprising:
[0034] A junk text message classification model building unit, used to build a junk text message classification model;
[0035] A spam message classification model training unit is used to determine a fixed number knowledge base, and train a spam message classification model based on pseudo-label fusion clustering according to a data set corresponding to the fixed number;
[0036] The spam message classification unit is used to classify spam messages using a trained spam message classification model.
[0037] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a spam text message classification method and device based on pseudo-label fusion clustering, which can accurately and efficiently identify spam text messages and improve the coverage rate of spam text message fixed numbers among all numbers.
[0038] Specifically, this application is mainly based on the pseudo-label fusion clustering method, which enables the spam SMS classification model to make full use of unlabeled data and mine its potential features, thereby providing additional training samples for the classification model, enabling the classification model to better learn the data distribution, thereby improving its generalization ability and classification accuracy.
[0039] On the other hand, since pseudo-labels introduce more noise, the present invention proposes to reduce the noise of pseudo-label datasets by training with the consistency of enhanced samples, thereby improving the shortcomings of pseudo-label methods. At the same time, based on data enhancement, classifier feedback is also used to achieve relative entropy maximization (REM) and conditional entropy minimization (CEM), balancing the diversity and quality of enhanced labeled samples.
[0040] At the same time, the fixed number identification method of the present invention improves the coverage rate of fixed numbers in the total number of SMS senders, and provides more raw data for the subsequent spam SMS classification model. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0042] Figure 1 This is a flow chart of fixed number identification for short message services in the present invention;
[0043] Figure 2 This is a flow chart of fixed number identification for a short message service platform in the present invention;
[0044] Figure 3 This is a flow chart of the spam SMS classification model training in the present invention. DETAILED DESCRIPTION
[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the 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.
[0046] The embodiment of the present invention discloses a junk text message classification method based on pseudo-label fusion clustering, which aims to accurately and efficiently identify junk text messages from massive data, thereby avoiding the omission of effective information.
[0047] In this embodiment, the method for classifying fixed numbers of junk text messages is implemented by the following steps:
[0048] S1. Build a spam SMS classification model;
[0049] S2. Determine a fixed number knowledge base, and train a spam SMS classification model based on pseudo-label fusion clustering according to the data set corresponding to the fixed number;
[0050] S3. Use the trained spam message classification model to classify spam messages.
[0051] The specific execution process of each step is as follows:
[0052] The first step is to construct a spam SMS classification model. This embodiment does not specifically limit the type of spam SMS classification model. As long as the model can achieve SMS classification, the classification method disclosed in this application can be applied.
[0053] The second step mainly includes the following two parts:
[0054] S21, determining a fixed number knowledge base,
[0055] This embodiment obtains part of the data within a set life cycle from the massive SMS data as initial data for determining the fixed number knowledge base;
[0056] The process of determining the fixed number knowledge base includes:
[0057] S211, obtaining a large amount of SMS data, unified into the following format: number, corresponding organization name, number of occurrences within a set period;
[0058] This application mainly completes data preprocessing through a merging algorithm. That is, first delete the first two digits of the number starting with 86 to get the real SMS sender number, then obtain information that can represent the SMS sender organization from the SMS content as the initial SMS content sender organization name, and at the same time merge different expressions or abbreviations of the same organization in the organization name to get the final sender organization name. After that, calculate the number of days that the SMS sender number corresponds to a unique organization during the life cycle.
[0059] S212: extract the management bit in the number, determine the fixed number segment and the corresponding organization name based on the management bit, and determine the fixed number knowledge base according to the fixed number segment and the corresponding organization name.
[0060] In an exemplary embodiment, the number is divided into a management bit and a text message sequence number, the management bit is the head, and the extension bit is the tail. The specific division method can be based on the management bit length of each number segment specified by the Ministry of Industry and Information Technology, as shown in Table 1.
[0061] Table 1
[0062] Number segment Management position length 95 5 / 6 digit 96 5 / 6 digit 1062 8-bit 1063(0-4) 8-bit 1063(5-8) 12-bit 1065 5-digit 1066 8-bit 1068 / 1069(0-4) 8-bit 10696 12-bit 10698 12-bit 10639 / 10699 8-bit
[0063] According to the specific plan for the 106 number segment in the "Telecommunications Network Numbering Plan (2017 Edition)" issued by the Ministry of Industry and Information Technology, as shown in Table 2, these number segments are divided into two categories: short message services and short message service access platforms. Since the short message services operated by basic operators also belong to third-party platforms, they are classified into the short message service access platform category. In addition, the 95 and 96 number segments can also be applied for as short message service access codes, so the 95 and 96 number segments are classified as a separate category.
[0064] Table 2
[0065]
[0066]
[0067] In this embodiment, the fixed number identification method for short message services is as follows: Figure 1 As shown, the data with the same head are grouped together. If there is only one organization in the group, the numbers with this head as the management bit belong to this type of organization, and the head is a fixed number segment. If there are multiple organization names in the group, the dictionary is used to query the corresponding application organization of the head, and the head is the fixed number segment of the corresponding organization.
[0068] The fixed number identification method for short message service platforms is as follows: Figure 2 As shown, the data with the same head are grouped together, and all the organizations in the group are traversed. If the organization in the group contains only one organization, it is further determined whether the number of days in the corresponding life cycle is greater than or equal to 2. If the condition is met, it is a fixed number. If the group contains multiple organizations, further processing is required. Continue to deduplicate this type of data, delete the data with the same SMS sender number, and traverse the remaining data. If the number of days in the corresponding life cycle of the data is greater than or equal to 2, it is determined that the number corresponding to the data is a fixed number.
[0069] For the data in the 95 and 96 number segments, the same numbers are grouped together. If the group of data contains a unique organization, the number corresponding to the data is a fixed number. If the organization corresponding to the data contains multiple organization names, the corresponding organization name is obtained by querying the dictionary for the organization corresponding to the number, and the number is also a fixed number.
[0070] Finally, a fixed number knowledge base is determined according to the fixed number segment and the corresponding organization name.
[0071] S22, according to the data set corresponding to the fixed number, the spam SMS classification model is trained based on pseudo-label fusion clustering; the training steps are as follows Figure 3 As shown, including:
[0072] S221, dividing the data set corresponding to the fixed number into labeled data and unlabeled data, and determining the true label of the labeled data;
[0073] Specifically, the data set is partially manually labeled, that is, based on the content of the text messages, the text messages containing marketing, promotion, fraud, violence, reactionary and other content are set to 1, that is, spam text messages, and other text message labels are set to 0.
[0074] In a preferred embodiment, the TF-IDF algorithm is used to perform dimension reduction processing on the SMS text before labeling to extract the key information of the SMS, so as to reduce the influence of noise words and meaningless words in the SMS on the subsequent classification model.
[0075] TF refers to the frequency of a word in a text message, and IDF refers to the inverse of the frequency of the word in all text messages, which is calculated by the following formula:
[0076]
[0077] Among them, TF(d,w) represents the frequency of word w in SMS text d, N represents the total number of texts, and N(w) represents the number of times word w appears in the sentence. According to the calculation results, the words appearing in the SMS are sorted, and the top-ranked words are selected to replace the SMS text and vectorized.
[0078] S222, based on the labeled data, using the clustering model to obtain pseudo labels for the unlabeled data;
[0079] In this embodiment, the implementation process is as follows:
[0080] 1) Divide the data into labeled data X according to the obtained data set l With unlabeled data X u , respectively expressed as follows:
[0081]
[0082]
[0083] Where m is the number of unlabeled data, is the unlabeled text, n is the number of labeled data, is the marked text, is the true label, is a one-hot vector, C is the number of categories;
[0084] 2) Using labeled data X l Train the clustering model and then pass the unlabeled data X u Get a pseudo-labeled dataset as input;
[0085] In one embodiment, the clustering model adopts the k-means clustering algorithm. The k-means clustering algorithm realizes data classification by calculating the distance between each data. First, multiple center points are randomly initialized according to the spam message classification requirements, and the samples are classified according to the distance between each sample point and the center point. The center point is dynamically adjusted according to the positional relationship between the center point and the sample point, so that the distance between each class is as large as possible to realize sample clustering. The specific implementation process is as follows:
[0086] Step 1: Take the target SMS data and the required number of categories k as model input;
[0087] Step 2: Select k points as the initial centroids of the clusters. Randomly select a data point as the first cluster center, then for each data point, calculate its distance from the selected cluster center, and select the data point with the farthest distance as the next cluster center, until k cluster centers are selected;
[0088] Step 3: Calculate the distance between each sample and the centroid according to the following formula and assign it to the cluster with the closest distance;
[0089] d c =|cc i |2
[0090] Step 4: After calculating the mean of all sample points in each cluster, find the new centroid to replace the previous centroid using the following formula;
[0091]
[0092] Step5: Continue to iterate Step3, Step4, and Step5 until the model converges.
[0093] Among them, c represents the value of the data set sample, C i represents the i-th cluster, Represents the center of the i-th cluster. The label corresponding to the cluster center is the pseudo-label. The unlabeled data is passed as output to the k-means clustering model. The model assigns the unlabeled data to the nearest cluster center to obtain a pseudo-label dataset. Since the pseudo-label dataset obtained at this time has a certain amount of noise, it is necessary to improve the data quality of the pseudo-label dataset by using consistency training on its enhanced samples.
[0094] In a preferred embodiment, data enhancement is performed on the labeled data and the unlabeled data respectively.
[0095] For labeled data, the present invention uses the EPiDA framework to perform data enhancement on labeled data from two aspects: data diversity and data semantic consistency. EPiDA consists of three parts: data enhancement algorithm, classifier or classification model, and sample evaluation and selection module. The sample evaluation and selection module evaluates the enhanced samples generated by the data enhancement algorithm through the relative entropy maximization mechanism and the conditional entropy minimization mechanism based on the feedback from the classifier, and finally selects suitable enhanced samples to improve the classification ability of the subsequent model.
[0096] The data enhancement algorithm selects the back translation method and the context enhancement algorithm. The data enhancement algorithm is used to obtain data enhancement samples. In order to improve the diversity of enhanced sample data, it is necessary to select data that is significantly different from the original sample. According to the relative entropy maximization mechanism, the corresponding objective function is as follows:
[0097]
[0098] in, represents the enhanced sample, y i represents the true label, φ:R d →R D represents feature mapping, w is a hyperparameter, H is Shannon entropy, which is used to represent the information gain of enhanced samples, and D is the relative entropy. Used to calculate and p(y i ). Since p(y i ) is a one-hot vector, so H(p(y i ))=0, in order to maximize the difference between the original sample and the enhanced sample, the relative entropy needs to be maximized Therefore, the diversity score of the enhanced sample is:
[0099]
[0100] In order to keep the semantic consistency of the data, it is necessary to select samples with smaller semantic deviation from the original samples. This can be achieved by minimizing the conditional entropy mechanism. The objective function is shown in the following formula:
[0101]
[0102] As can be seen from the formula, minimizing the conditional entropy is actually minimizing the selected sample The entropy of the selected sample is maximized With the original sample The mutual information between them, the quality score of the enhanced sample is:
[0103]
[0104] Finally, the final score of the enhanced sample is:
[0105]
[0106] The loss function of the EPiDA data enhancement framework is:
[0107] L E =L C +L R
[0108] According to the score, select appropriate enhanced samples and add them to the label data set to expand the data. The expanded data set is recorded as X L .
[0109] For unlabeled data, the present invention obtains unlabeled data enhancement samples through the back translation method. The specific steps are as follows:
[0110] Step 1: Translate the data set language text into the target language text;
[0111] Step 2: Translate the target language text back to the source language. The model input is the target language text and the output is the translated source language. The translation models in Step 1 and Step 2 are two different trained translation models.
[0112] Step 3: Compare the original text with the back-translated text. If there is a large difference between the two, use the difference as a feedback signal to improve the translation model.
[0113] Step 4: Get the final reverse translation result as an enhanced sample for subsequent processing.
[0114] After obtaining the enhanced samples according to the above steps, the unlabeled data samples and the data augmented samples are trained for consistency, and the consistency loss function is used to minimize the metric between the two sample distributions. The pseudo-label consistency loss function is shown in the following formula:
[0115]
[0116] Among them, θ is a model parameter, is the data enhancement function, X u represents an unlabeled dataset, is the pseudo-label dataset distribution, It represents the distribution of data augmentation samples output by the clustering algorithm, and measures the distance between the two distributions by using KL divergence. Consistency training makes full use of a large amount of unlabeled data to force the alignment of unlabeled samples with the augmented sample clustering results, thus improving the generalization ability of the model.
[0117] S223, using the real label and the pseudo label, the spam message classification model is trained according to the following loss function:
[0118] In this embodiment, Bert is used as the basic model in the pre-training stage, and English is used as the intermediate language for back-translation for data enhancement involved in the method. The k value of the clustering algorithm is set to 2 according to the classification requirements.
[0119] In the training phase, the pseudo-label dataset X after noise reduction is p As input passed to the classification model for training, the weakly supervised cross entropy loss function is obtained by the cross entropy of the classification label and the pseudo label:
[0120]
[0121] Keep the parameters, then merge the original labeled data with the enhanced samples after EPiDA screening, and merge the merged data X L Passed as input to the classification model for training, the supervised cross entropy loss function is passed through the dataset X L The cross entropy between the predicted label and the true label output in the classification model is obtained.
[0122]
[0123] The final loss function is:
[0124]
[0125] Where, X P represents a pseudo-label dataset, X L represents the real label data set, θ1, θ2 are model parameters, y * represents the predicted label, Represents the output distribution of the clustering task corresponding to the dataset generated by back-translation of the unlabeled dataset.
[0126] In another embodiment, the final loss function is determined as:
[0127] L′=L+L u +L E
[0128] Among them, L Erepresents the loss function for data enhancement using EPiDA data, L represents the sum of the weakly supervised cross entropy loss function and the supervised cross entropy loss function, L u represents the pseudo-label consistency loss function.
[0129] The third step is to use the trained spam message classification model to classify spam messages.
[0130] The present invention realizes the classification of fixed number text messages through the above-mentioned classification model based on pseudo-label fusion clustering. Overall, the present application realizes a text classification algorithm based on pseudo-label fusion clustering. The algorithm is improved for labeled data and unlabeled data respectively, which solves the problem of low classification accuracy caused by the small amount of labeled data in the data set and the limited learning ability of the model during the pre-training process.
[0131] In another embodiment, a junk SMS fixed number classification device based on pseudo-label fusion clustering includes:
[0132] A junk text message classification model building unit, used to build a junk text message classification model;
[0133] A spam message classification model training unit is used to determine a fixed number knowledge base, and train a spam message classification model based on pseudo-label fusion clustering according to a data set corresponding to the fixed number;
[0134] The spam message classification unit is used to classify spam messages using a trained spam message classification model.
[0135] The details of the execution steps of each unit are consistent with the contents disclosed in the method for classifying spam messages based on pseudo-label fusion clustering, so they will not be repeated here.
[0136] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0137] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A spam SMS classification method based on pseudo-label fusion clustering, It is characterized in that The steps include: Build a spam SMS classification model; Determine the fixed number knowledge base, and train the spam SMS classification model based on pseudo-label fusion clustering according to the data set corresponding to the fixed number; process include; Dividing the data set corresponding to the fixed number into labeled data and unlabeled data, and determining the true label of the labeled data; Based on the labeled data and the corresponding true labels, the clustering model is used to obtain the pseudo labels of the unlabeled data; Using the real label and the pseudo label, the spam message classification model is trained according to the following loss function: ; In the formula, represents a pseudo-label dataset, represents the real label dataset, , are model parameters, represents the predicted label, Represents the output distribution of the clustering task corresponding to the dataset generated by back-translation of the unlabeled dataset Use the trained spam message classification model to classify spam messages.
2. According to claim 1, a junk text message classification method based on pseudo-label fusion clustering, It is characterized in that The process of determining the fixed number knowledge base includes: Obtain massive amounts of SMS data and unify them into the following format: number, corresponding organization name, and number of occurrences within a set period; The management bit in the number is extracted, a fixed number segment and a corresponding organization name are determined based on the management bit, and a fixed number knowledge base is determined according to the fixed number segment and the corresponding organization name.
3. According to claim 2, a junk text message classification method based on pseudo-label fusion clustering, It is characterized in that The method for determining the fixed number segment and the corresponding organization name based on the management bit is: For service-related SMS data, data in a unified format are grouped according to the management bit. If there is only one organization in the group, the only organization is the organization corresponding to the current management bit; if there are multiple organizations in the group, the organization corresponding to the management bit is the organization corresponding to the current management bit.
4. According to claim 2, a junk text message classification method based on pseudo-label fusion clustering, It is characterized in that For service platform SMS data, the data in a unified format are grouped according to the management bit. If there is only one organization in the group, it is determined whether the number of times the unique organization appears within the set period is greater than n. If so, the unique organization is the organization corresponding to the current management bit. If there are multiple organizations in the group, the data is deduplicated, and it is determined whether the number of times the remaining organizations appear within the set period is greater than n. If so, the remaining organizations are used as the organizations corresponding to the current management position.
5. According to claim 1, a junk text message classification method based on pseudo-label fusion clustering, It is characterized in that Before training, the data set corresponding to the fixed number is prioritized for dimensionality reduction and vectorization.
6. The junk text message classification method based on pseudo-label fusion clustering according to claim 1, It is characterized in that The labeled data is enhanced through the data enhancement algorithm and evaluated through the relative entropy maximization mechanism and the conditional entropy minimization mechanism. The evaluation function is: ; In the formula, represents the jth enhanced sample generated by the i-th real sample, represents the i-th true sample, represents the true label, n is the number of true sample sets, h is the number of enhanced samples generated for each true sample, represents feature mapping, w is a hyperparameter, H is Shannon entropy, which is used to represent the information gain of enhanced samples, D is relative entropy, I is mutual information, and p is probability distribution. Used to calculate and The difference.
7. The junk text message classification method based on pseudo-label fusion clustering according to claim 6, It is characterized in that Back-translation is used to perform sample augmentation on unlabeled data, and the metric between the two sample distributions is minimized by a consistency loss function, which is as follows: ; In the formula, is a model parameter, is the data enhancement function, represents an unlabeled dataset, is the pseudo-label dataset distribution, represents the distribution of data augmentation samples output by the clustering algorithm, D KL represents the distance between two distributions measured by KL divergence, and E represents expectation.
8. According to claim 7, a junk text message classification method based on pseudo-label fusion clustering, It is characterized in that Using the real label and the pseudo label, the spam message classification model is trained according to the following loss function: 。 9. A fixed number spam SMS classification device based on pseudo-label fusion clustering, It is characterized in that A method for classifying fixed numbers of spam short messages based on pseudo-label fusion clustering as described in any one of claims 1 to 8 is applied, comprising: A junk text message classification model building unit, used to build a junk text message classification model; A spam message classification model training unit is used to determine a fixed number knowledge base, and train a spam message classification model based on pseudo-label fusion clustering according to a data set corresponding to the fixed number; The spam message classification unit is used to classify spam messages using a trained spam message classification model.
