A network news classification management system based on feature recognition

Abstract

This invention relates to the field of data processing technology, specifically a feature-based online news classification management system, comprising: a first classification module for acquiring a news data set, including the acquired news titles, text content, and user behavior data; classifying the news according to the news titles and text content to obtain a first classification result; a second classification module for classifying the news according to the news user behavior data to obtain a second classification result; the user behavior data includes the number of clicks, views, shares, and comments of the news; and a classification detection module for detecting the sentiment tendency of the first classification result; if the detected event in the first classification result has a positive sentiment tendency, then the system is based on the positive sentiment tendency of the event in the first classification result; thus improving the accuracy and efficiency of news classification.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically a network news classification and management system based on feature recognition. Background Technology

[0002] Currently, there are numerous text classification and sentiment analysis models available in the latest technologies, and their algorithms are relatively mature. Existing text classification and sentiment analysis models are independent algorithms. Mainstream text classification models employ Bi-LSTM, CNN, and FastText algorithms, all of which use the entire news article as training data based on characters and words. Because they use the entire text as training data, there is only one classification for a specific news article. However, when multiple subjects appear in a news article, different subjects may have different classifications. For example, if a news article describes negative information about A and positive information about B, classifying the entire article will only yield one category. This classification might be correct for A, but when A and B are in different categories (A is negative, B is positive), existing classification methods cannot adequately label and classify different subjects within the same news article. Sentiment analysis, on the other hand, often uses the Bi-LSTM algorithm. Sentiment analysis typically only outputs the overall sentiment tendency of the article, including positive and negative probabilities; it does not provide more specific distinctions between sentiment categories. Therefore, relying entirely on a single model for prediction is highly dependent on the preparation of news corpus data. Given the wide variety of news styles, the same news written by different authors may have completely different styles, thus having limitations.

[0003] For example, Chinese Patent Publication No. CN109472470A discloses a method for risk classification of corporate news data that combines deep learning and logical rules, including the following steps: extracting sentences containing relevant attributes of a specific enterprise from news materials; inputting sentences containing relevant attributes into a CNN sentence classification model to obtain the sentence classification of each sentence; weighting each sentence classification to obtain the news classification of the current news; inputting news materials into a Bi-LSTM article sentiment classification model on a per-article basis to obtain the sentiment tendency value of each news article; and combining the news classification and sentiment tendency value of each news article to obtain the news risk category of each news article.

[0004] However, when classifying news, it is not possible to identify each sentence and corresponding keyword in the news based on its specific content, determine the tendency of the news to be classified, and find the most suitable classification label for the current news classification in order to improve the management effect of news. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a network news classification management system based on feature recognition, comprising: a first classification module, used to acquire a news data set, the news data set including the acquired news titles, text content, and user behavior data; classifying the news according to the news titles and text content to obtain the first classification result of the news.

[0006] The second classification module is used to classify news based on user behavior data to obtain the second classification result of the news; user behavior data includes the number of clicks, views, shares, and comments of the news.

[0007] The classification detection module is used to detect the sentiment tendency of the first classification result. If the detected event in the first classification result has a positive sentiment tendency, the module extracts the event type corresponding to the event in the first classification result from the preset database based on the positive sentiment tendency of the event in the first classification result, and determines the theme and category expressed by the positive sentiment tendency.

[0008] If the detected event in the first classification result has a negative sentiment tendency, then based on the negative sentiment tendency of the event in the first classification result, the scene type corresponding to the event in the first classification result is extracted from the preset database, and the theme and category expressed by the negative sentiment tendency are determined.

[0009] The first classification adjustment module is used to adjust the first classification results based on the processing of corresponding themes and categories in positive and negative emotional tendencies.

[0010] The final classification module is used to compare the adjusted first classification result with the second classification result to determine the current final classification result.

[0011] The beneficial effects of this invention are as follows: First, by using multi-dimensional classification, sentiment analysis, and determination of master-slave relationships, this invention significantly improves the accuracy of news classification; it can complete the classification of a large number of news items more quickly, thus improving classification efficiency; and by adjusting the classification results according to sentiment, the classification becomes more flexible and adaptable.

[0012] Second, this invention classifies news based on its headlines and text content to obtain a first classification result, and also combines user behavior data for classification to obtain a second classification result; this multi-dimensional method makes the classification more comprehensive and accurate.

[0013] Third, by performing sentiment analysis on news events, this invention can identify positive and negative sentiment tendencies and extract corresponding event or scene types from a preset database accordingly, further adjusting the first classification result; this helps to more accurately understand the emotional tone and theme of news content.

[0014] Fourth, in the process of event extraction and keyword connection, the present invention constructs a clearer and more hierarchical classification structure by determining the master-slave relationship between keywords.

[0015] Fifth, by calculating the growth and popularity indicators of news, this invention can more objectively assess the influence and audience acceptance of news, thereby optimizing the second classification results. Attached Figure Description

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] Figure 1 This is a schematic diagram of a feature-based online news classification and management system.

[0018] Figure 2 This is a system module diagram of a feature-based online news classification and management system. Detailed Implementation

[0019] The embodiments of the present invention are described in detail below. The embodiments described below are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. Where specific techniques or conditions are not specified in the embodiments, they shall be performed in accordance with the techniques or conditions described in the literature in the art or in accordance with the product manual.

[0020] See Figure 1 , Figure 2 A feature-based online news classification management system includes: a first classification module for acquiring a news data set, which includes the acquired news titles, text content, and user behavior data; classifying the news according to the news titles and text content to obtain the first classification result of the news.

[0021] The second classification module is used to classify news based on user behavior data to obtain the second classification result of the news; user behavior data includes the number of clicks, views, shares, and comments of the news.

[0022] The classification detection module is used to detect the sentiment tendency of the first classification result. If the detected event in the first classification result has a positive sentiment tendency, the module extracts the event type corresponding to the event in the first classification result from the preset database based on the positive sentiment tendency of the event in the first classification result, and determines the theme and category expressed by the positive sentiment tendency.

[0023] If the detected event in the first classification result has a negative sentiment tendency, then based on the negative sentiment tendency of the event in the first classification result, the scene type corresponding to the event in the first classification result is extracted from the preset database, and the theme and category expressed by the negative sentiment tendency are determined.

[0024] The first classification adjustment module is used to adjust the first classification results based on the processing of corresponding themes and categories in positive and negative emotional tendencies.

[0025] The final classification module is used to compare the adjusted first classification result with the second classification result to determine the current final classification result.

[0026] The news data set is first sent to the first classification module for preliminary classification.

[0027] Meanwhile, user behavior data from the news dataset is sent to the second classification module for categorization.

[0028] The output of the first classification module is sent to the classification detection module for sentiment analysis.

[0029] The output of the classification detection module, along with the first classification result, is sent to the first classification adjustment module for adjustment.

[0030] Finally, the adjusted first classification result, along with the output of the second classification module, is fed into the final classification module to determine the final classification result.

[0031] First, news data needs to be collected from various news sources (such as news websites, social media, etc.). This data typically includes news headlines, text content, and user behavior data (such as clicks, views, shares, and comments).

[0032] For example, the news headline reads: "Major Breakthrough in New Vaccine Development"; the text content reads: "Scientists recently announced that they have successfully developed a new vaccine that is highly effective against a certain virus..."; user behavior data includes: 10,000 clicks, 8,000 reads, 500 shares, and 200 comments.

[0033] Natural Language Processing (NLP) techniques are used to classify news headlines and text content, which typically involves text analysis, keyword extraction, and sentiment analysis.

[0034] For example, the first category result indicates science / medical news with a positive sentiment (because the news mentions a major breakthrough in vaccine development).

[0035] News can be categorized based on user behavior data (such as clicks and views), which can provide additional information about the popularity, impact, or user interest of the news.

[0036] For example, the second category result: trending news (based on high click-through rates and readership).

[0037] Based on the sentiment (positive or negative) in the first classification results, the event type or scene type related to the news event is extracted from the preset database, and the theme and category expressed by the sentiment are determined.

[0038] For example, for news with a positive sentiment, the system may extract "technological innovation" as the event type from the database and determine the theme as "medical progress".

[0039] The adjusted first classification result is compared with the second classification result based on user behavior, and the two are combined to determine the final news classification.

[0040] For example, the final classification result might be "Hot Medical Technology Breakthrough News", which takes into account both the news content (medical technology breakthrough) and its popularity (hot).

[0041] The processing method for the detected first classification result includes: extracting events from the news data set and constructing an event dictionary corresponding to the events. The event dictionary is divided into multiple categories according to the title and text content of the corresponding news, and the first keyword corresponding to the title and the second keyword corresponding to the text content under each category are determined; wherein, the number of the first keyword and the second keyword is at least two.

[0042] The first keyword is compared with a preset thesaurus, and the keyword with the highest similarity among the first keywords is selected as the initial node. The initial node is used to determine the master-slave relationship among the first keywords and to determine the main and secondary categories of the current title. At the same time, the obtained initial node needs to be matched with the subsequent second keyword. All words that can represent the current news category obtained from the title and text content are matched to find the master-slave relationship of the words in the overall dataset under the relative position and frequency of occurrence, so as to understand how the corresponding news should be classified and managed under the corresponding sentence structure and paragraph description.

[0043] Connect the first keyword to the initial node, and determine the master-slave relationship between the first keyword and the initial node according to the word frequency and importance of the first keyword. Connect the connected first keyword to the second keyword, determine the master-slave relationship between the connected first keyword and the second keyword, and output the connected first keyword and the second keyword as the first category result. Now, after connecting the first keyword and the second keyword, set multiple category labels for the current title and text content based on their contextual master-slave relationship, and divide the title and text content into multiple categories based on these category labels as the first category result.

[0044] The method for determining the master-slave relationship between the first keyword and the initial node includes: connecting the initial node and the first keyword sequentially based on the performance weight of the first keyword. The performance weight represents the weight that the current first keyword can express under the corresponding category topic, and whether it can have representativeness under the current category. This performance weight is preset in the database according to the corresponding words. When connecting the initial node and the first keyword, the magnitude of the performance weight corresponding to the first keyword is first determined. Then, the first keyword, the second keyword and the initial node are connected in the way of connecting nodes in a directed graph, thereby iterating out the corresponding connection relationship.

[0045] The method for determining the master-slave relationship between the first keyword and the second keyword is as follows: determine the relevance of the corresponding category topics of the first keyword and the second keyword; if the second keyword depends on the first keyword, then the second keyword is taken as the slave node of the first keyword; if the second keyword does not depend on the first keyword, then based on the relevance of the corresponding category topics of the first keyword and the second keyword, the second keyword with the greatest relevance is taken as the slave node of the first keyword.

[0046] set up Let V be a graph, where V is the set of nodes, representing the set of the first and second keywords, and E is the set of edges, representing the edges connecting the first and second keywords; for any two nodes... The connection between them can be represented by a function: .

[0047] in: Represents a node and The edge between; and These are nodes and word frequency; and These are nodes and The importance of the nodes (which could be TF-IDF value, PageRank value, etc.); the function (f) determines whether there is an edge between the two nodes and the weight or type of the edge based on these parameters; based on the edge obtained, the connection between the corresponding keywords is adjusted until each keyword is connected.

[0048] After completing the specific connection processing of the first keyword and the second keyword, the method for obtaining the sentiment tendency of the events in the first classification result includes: distributing the event dictionary corresponding to the first classification result according to the events that occurred, determining the cosine similarity corresponding to each event, and clustering the events to obtain different event clusters; determining the occurrence frequency and inverse document probability corresponding to the event cluster, and when the product of the occurrence frequency and inverse document probability corresponding to the event cluster is greater than a preset frequency value, extracting the corresponding feature value as a feature dictionary; identifying the first keyword and the second keyword under the corresponding feature dictionary, identifying the sentiment tendency of the first keyword and the second keyword corresponding to the feature dictionary, and using the sentiment tendency as the sentiment tendency of the events in the first classification result.

[0049] The method for obtaining news event clusters is to convert different events into an event vector form, which can capture the semantic features of the events; calculate the cosine similarity between event vectors to measure the semantic similarity between different events, and group the events according to the obtained cosine similarity to form different event clusters.

[0050] The methods for obtaining the occurrence frequency and inverse document probability of an event cluster include: for each news article in the news data set, traversing the event clusters it marks; maintaining a counter for each event cluster; incrementing the counter value by 1 whenever the event cluster is encountered; after traversal, the counter value of each event cluster is the occurrence frequency of the event cluster in the document set.

[0051] Statistical inverse document probability: First, determine the number of news items in the news dataset. For each event cluster, calculate the number of news items that contain that event cluster.

[0052] The inverse document probability is then expressed as: .

[0053] in, Represents the inverse document probability. Represents a cluster of events. It is the total number of news items in the news data set. It contains event clusters The number of news items.

[0054] The word frequency obtained here is used to determine the prevalence of a word or character in news by querying the number of times it appears and the total number of characters appearing in the corresponding news dataset. Meanwhile, the inverse document probability is used to determine the frequency of a word in news. The fewer news articles the selected word contains, the higher the inverse document probability value, indicating that the selected word has a better ability to distinguish categories. The product of the word frequency and the inverse document probability obtained for this word is then used. The larger the product value, the higher the importance of the word in the corresponding news, making it easier to distinguish the current news and making the obtained news more clearly distinguishable, which is convenient for subsequent management.

[0055] At the same time, the current classification needs to be adjusted based on the emotional tendency identified in the news at this time, so that the distinction between news is more obvious. The same or multiple news items can be separated according to different emotions, which makes it easier to manage news with corresponding emotional tendencies.

[0056] The processing method for adjusting the first category results includes: when there is a positive sentiment tendency, based on the event type corresponding to the event in the first category results, the order of each category in the first category results is adjusted according to the priority of the event type.

[0057] When there is a negative sentiment tendency, based on the scene type corresponding to the event in the first classification result, scene labels are set for each category in the first classification result according to different scene types, and the corresponding categories in the first classification result are clustered according to the scene labels in order to adjust the first classification result.

[0058] The priority of the above-mentioned event types is determined by the urgency of the event type. The urgency here is obtained through preset conditions, such as traffic accidents > sports > festival celebrations > cultural activities. Then, the obtained news is set according to the corresponding priority. Based on the results of these settings, the order of each category in the first category result can be obtained.

[0059] When there is a negative emotional bias, the news we receive will be categorized according to different scenarios to determine the appropriate category for the news in the current context.

[0060] After obtaining the first category, the processing of the second category results includes determining the specific category of the second category result based on the number of clicks, views, shares, and comments of the news, and calculating the growth index and popularity index corresponding to the news under each category based on user behavior data.

[0061] The growth and popularity indicators are used as input features, and the specific category of the second classification result is used as the input label. These are input into the classification model to calculate the probability distribution of each category, and the category with the highest probability is taken as the second classification result.

[0062] The methods for determining growth indicators include: generating trend curves based on the number of clicks, reads, shares, and comments of news at different points in time; determining the growth value of the trend curve within each time period; and using the slope of the trend curve growth value as the growth indicator of the news within the corresponding time period.

[0063] When determining growth indicators, the corresponding growth indicators are represented by trend curves of clicks, views, shares, and comments over time. The slope of the growth indicator represents the corresponding growth value of these four curves. Based on the growth of these four trend curves, it is determined which attribute the growth trend of different types of news mainly depends on when classifying training samples, thus facilitating the classification of different news.

[0064] The welcome index is determined by weighting and summing the number of clicks, views, shares, and comments. The weighted sum is then used as the welcome index.

[0065] The popularity of a news story is determined by a comprehensive score based on clicks, views, shares, and comments, as users interact with the story. Furthermore, the growth potential of each news story is determined by analyzing the numerical growth of user behavior data across multiple event points. Growth potential is expressed as the percentage increase in clicks, views, and other data within a fixed time period, as well as the duration of this growth trend within each time period. This allows for the identification of whether the current news topic and category are likely to be popular and experience growth.

[0066] The calculation method for the classification model is as follows: Assume there are n training samples, and each sample has two input features: growth index and popularity metrics and an output label This indicates the specific category of the second category result of the news.

[0067] In the classification model, a linear combination of growth and popularity indicators is first determined. The probability of the corresponding output label is determined based on the obtained linear combination, and the corresponding cross-entropy loss is obtained based on the probability of the output label. Based on the obtained cross-entropy loss, the total loss of the classification model is obtained to determine whether the current classification is reasonable.

[0068] For the i-th sample, the linear combination of the growth index and the popularity index can be expressed as: .

[0069] in, It is a bias term. and is with and The corresponding weighting coefficients, It is a linear combination of growth indicators and popularity indicators. It is a popularity metric. It is a growth indicator.

[0070] Then, the sigmoid function is used to combine the linear combinations. Convert to probability output The probability of outputting the label is: .

[0071] in, It represents the probability of the output label; at this point, it predicts which category the growth and popularity indicators of the input will be classified into, thus determining the corresponding classification label when performing the current classification.

[0072] For classification problems, the commonly used loss function is the cross-entropy loss function. In the binary classification case, for the i-th sample, the cross-entropy loss can be expressed as: .

[0073] in, It is the probability of outputting the label. It is the loss value of the i-th sample.

[0074] If it is a multi-class classification problem and softmax regression is used, the loss function will be the sum of the cross-entropy losses over all classes, but the loss calculation for each sample will involve the predicted probabilities and true labels of all classes.

[0075] For the entire training set, the total loss is the average of the losses of all samples. The total loss of the classification model is expressed as: .

[0076] in, This is the total loss value.

[0077] The classification model is iterated continuously until the loss rate of the final output classification label is minimized. The classification result corresponding to the final output label is then used as the second classification result.

[0078] The processing method for the first classification result and the second classification result includes comparing the first classification result and the second classification result based on the obtained first classification result and second classification result. If the classification corresponding to the first classification result and the second classification result is consistent, the data in the first classification result and the second classification result are superimposed to obtain the final classification result.

[0079] If the classification results of the first classification result and the second classification result are inconsistent, the intersection of the first classification result and the second classification result is determined. If the number of categories in the intersection of the first classification result and the second classification result is greater than the preset number of categories, the event type corresponding to the first classification result is determined. Based on the event type of the first classification result, the category corresponding to the second classification result is added to the database. The added category and the category in the intersection of the first classification result and the second classification result are taken as the final classification result.

[0080] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention, which are still covered within the protection scope of the present invention.

Claims

1. A feature-based online news classification and management system, characterized in that, include: The first classification module is used to obtain a news data set, which includes the title, text content, and user behavior data of the obtained news. Based on the headlines and text content, the news is categorized to obtain the first category result. The second classification module is used to classify news based on user behavior data to obtain the second classification result of the news; user behavior data includes the number of clicks, views, shares, and comments of the news. The classification detection module is used to detect the sentiment tendency of the first classification result. If the detected event in the first classification result has a positive sentiment tendency, the module extracts the event type corresponding to the event in the first classification result from the preset database based on the positive sentiment tendency of the event in the first classification result, and determines the theme and category expressed by the positive sentiment tendency. If the detected event in the first classification result has a negative sentiment tendency, then based on the negative sentiment tendency of the event in the first classification result, the scene type corresponding to the event in the first classification result is extracted from the preset database, and the theme and category expressed by the negative sentiment tendency are determined. The first classification adjustment module is used to adjust the first classification results based on the processing of corresponding themes and categories in positive and negative emotional tendencies; The final classification module is used to compare the adjusted first classification result with the second classification result to determine the current final classification result; The processing method for the first classification result and the second classification result includes: based on the obtained first classification result and second classification result, comparing the first classification result and the second classification result; if the classification corresponding to the first classification result and the second classification result is consistent, then the data in the first classification result and the second classification result are superimposed to obtain the final classification result. If the classification results of the first classification result and the second classification result are inconsistent, the intersection of the first classification result and the second classification result is determined. If the number of categories in the intersection of the first classification result and the second classification result is greater than the preset number of categories, the event type corresponding to the first classification result is determined. Based on the event type of the first classification result, the category corresponding to the second classification result is added to the database. The added category and the category in the intersection of the first classification result and the second classification result are taken as the final classification result.

2. The network news classification and management system based on feature recognition according to claim 1, characterized in that, The processing methods for the detected first classification result include: For news datasets, events are extracted, and an event dictionary is constructed for each event. The event dictionary is divided into multiple categories according to the title and text content of the corresponding news articles. The first keyword corresponding to the title and the second keyword corresponding to the text content are determined for each category. The number of first and second keywords is at least two. The first keyword is compared with a preset thesaurus, and the keyword with the highest similarity to the preset thesaurus is selected as the initial node. The first keyword is connected to the initial node, and the master-slave relationship between the first keyword and the initial node is determined according to the word frequency and importance of the first keyword. The first keyword after connection is connected to the second keyword, and the master-slave relationship between the first keyword and the second keyword after connection is determined. The first keyword and the second keyword after connection are output as the first classification result.

3. The network news classification and management system based on feature recognition according to claim 2, characterized in that, The methods for determining the master-slave relationship between the primary key and the initial node include: Connect the initial node to the first keyword in sequence based on the performance weight of the first keyword. The performance weight represents the weight that the current first keyword can express under the corresponding category topic. The methods for determining the primary and secondary relationship between the primary and secondary keywords include: Determine the relevance of the category topics corresponding to the first keyword and the second keyword. If the second keyword depends on the first keyword, then the second keyword is taken as a slave node of the first keyword. If the second keyword does not depend on the first keyword, then based on the relevance of the category topics corresponding to the first keyword and the second keyword, the second keyword with the greatest relevance is taken as a slave node of the first keyword.

4. The network news classification and management system based on feature recognition according to claim 1, characterized in that, The method for obtaining the sentiment tendency of events in the first classification result includes: distributing the event dictionary corresponding to the first classification result according to the events that occurred, determining the cosine similarity corresponding to each event, and clustering the events to obtain different event clusters; determining the occurrence frequency and inverse document probability corresponding to the event cluster, and extracting the corresponding feature value when the product of the occurrence frequency and inverse document probability corresponding to the event cluster is greater than a preset frequency value, as a feature dictionary; identifying the first keyword and the second keyword under the corresponding feature dictionary, identifying the sentiment tendency of the first keyword and the second keyword corresponding to the feature dictionary, and using the sentiment tendency as the sentiment tendency of the events in the first classification result.

5. The network news classification and management system based on feature recognition according to claim 1, characterized in that, The processing method for adjusting the first category results includes: when there is a positive sentiment tendency, based on the event type corresponding to the event in the first category results, the order of each category in the first category results is adjusted according to the priority of the event type; When there is a negative sentiment tendency, based on the scene type corresponding to the event in the first classification result, scene labels are set for each category in the first classification result according to different scene types, and the corresponding categories in the first classification result are clustered according to the scene labels in order to adjust the first classification result.

6. The network news classification and management system based on feature recognition according to claim 1, characterized in that, The processing method for the second category results includes determining the specific category of the second category results based on the number of clicks, views, shares, and comments of the news, and calculating the growth index and popularity index corresponding to the news under each category based on user behavior data; The growth and popularity indicators are used as input features, and the specific category of the second classification result is used as the input label. These are input into the classification model to calculate the probability distribution of each category, and the category with the highest probability is taken as the second classification result.

7. The network news classification and management system based on feature recognition according to claim 6, characterized in that, Methods for determining growth indicators include: The number of clicks, views, shares, and comments of news articles are used to generate trend curves based on the corresponding values ​​at different time points. The growth value of the trend curve in each time period is determined, and the slope of the trend curve growth value is used as the growth indicator of news in the corresponding time period. Methods for determining popularity metrics include: The number of clicks, views, shares, and comments are weighted and summed, and the resulting weighted sum is used as the popularity index.

8. A network news classification and management system based on feature recognition according to claim 6, characterized in that, In the classification model, the linear combination of growth index and popularity index is first determined. The probability of the corresponding output label is determined based on the obtained linear combination. The corresponding cross-entropy loss is obtained based on the probability of the output label. The total loss of the classification model is obtained based on the obtained cross-entropy loss. In the classification model, for the i-th sample, the forward propagation calculation formula can be expressed as: ； in, It is a bias term. and is with and The corresponding weighting coefficients, It is a linear combination. It is a popularity metric. It is a growth indicator; The probability of outputting the label is: ； in, It represents the probability of outputting the label.

9. A network news classification and management system based on feature recognition according to claim 8, characterized in that, For the i-th sample, the cross-entropy loss can be expressed as: ； in, It is the probability of outputting the label. It is the loss value of the i-th sample; The total loss of the classification model is expressed as: ； in, This is the total loss value.

Images