Artificial intelligence real-time news recommendation system based on model

Abstract

The invention relates to the technical field of artificial intelligence, and provides a model-based artificial intelligence real-time news recommendation system, which comprises a data acquisition module, a preprocessing module, a model training module and a recommendation module. The system collects news data from multiple sources, cleans and converts the news data, constructs a recommendation model by adopting a collaborative filtering algorithm, and generates a personalized recommendation list to be pushed to a user. According to the method, the news reading correlation and the user satisfaction can be improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically, to a model-based artificial intelligence real-time news recommendation system. Background Art

[0002] Existing news recommendation systems mostly use algorithms such as content-based recommendation or collaborative filtering, which aim to recommend news content based on the user's historical behavior or the behavior of similar users. These systems can provide personalized news recommendations to a certain extent, but due to the massive amount of news data and the diversity of user behavior, existing technologies still have shortcomings in terms of real-time, accuracy, and personalized recommendations. For example, some systems may not be able to update recommended content in real time, or may not be able to effectively capture the user's latest interest changes when processing user behavior data. In addition, the diversity and complexity of news content also require recommendation systems to be able to understand the content of news data more accurately, while traditional recommendation algorithms may not be deep enough in text processing and semantic understanding.

[0003] In the process of implementing the embodiments of the present invention, it is found that there are at least the following problems or defects in the prior art: the existing system often ignores the real-time nature of news content and the multi-dimensional analysis of user behavior data when recommending news, resulting in the recommendation results being inaccurate and in time. At the same time, the in-depth understanding of news content and the capture of dynamic changes in user interests are insufficient, which limits the performance of the recommendation system and the improvement of user experience. Summary of the invention

[0004] The present invention provides a model-based artificial intelligence real-time news recommendation system, comprising: A data collection module, used to collect news data from multiple news sources in real time; Data preprocessing module, used to clean and convert the collected news data; Model training module, used to build recommendation models using collaborative filtering algorithms; The recommendation module is used to generate and push personalized news recommendation lists for each user based on the recommendation model.

[0005] Furthermore, the news data collected by the data collection module includes news title, news content, news release time, news category, news source and user news browsing behavior data, and the user news browsing behavior data includes user identification, browsing time, browsing duration, likes, comments and sharing operation information.

[0006] Furthermore, the data preprocessing module cleans the collected news data to remove noise data, duplicate data and data with incorrect format.

[0007] Furthermore, the data preprocessing module converts the news data into a structured data format, and the specific steps include: Perform word segmentation on news data and extract keywords; The extracted keywords are converted into vector form using the Word2Vec model. The specific steps are as follows: an initial random vector is assigned to each keyword, and the vector of each keyword is adjusted through training data to capture its context information. Finally, each keyword is represented as a vector of fixed dimension. The news content is represented as a vector, where each dimension of the vector corresponds to a specific keyword or feature, and its weight is based on the frequency of occurrence of the keyword in the news. and inverse document frequency Determine, among which is the total number of news, To contain keywords The number of news items; news Content Vector Keywords Weight .

[0008] Furthermore, the data preprocessing module constructs a user-news browsing matrix The rows of the matrix represent users, the columns represent news, and the matrix elements Indicates user About News Quantitative value of browsing behavior. User-news browsing matrix The specific form is:

[0009] in, is the number of users, is the number of news, The specific meaning is as follows: If the user About News If the browsing time of exceeds a certain threshold, ,otherwise ; Like, comment or share behaviors correspond to weighted values. ,Comment ,share .

[0010] Furthermore, the matrix elements The quantification method is: If the user About News If the browsing time of exceeds a certain threshold, ,otherwise ; Like, comment or share behaviors correspond to weighted values. ,Comment ,share .

[0011] Furthermore, the model training module uses a collaborative filtering algorithm to construct a recommendation model, including the following steps: To calculate the similarity between users, use the formula:

[0012] in, and Represents two different users. and Respectively represent users and users Browsed news collection, Indicates user About News Quantified value of browsing behavior, Indicates user About News The quantified value of browsing behavior. In order to avoid the impact of data sparsity on similarity calculation, a user activity penalty factor is introduced and The corrected user similarity formula is:

[0013] Predict target users For unread news Interest level rating , the calculation formula is: in, Represents the target user A user set with high similarity (similarity greater than the set threshold) , For target users Average rating of viewed news in For target users Browsed news collection, For users Average rating of viewed news .

[0014] Furthermore, the user similarity calculation formula is: At the same time, in order to avoid the impact of data sparsity on similarity calculation, a user activity penalty factor is introduced. and The corrected user similarity formula is: Furthermore, the news recommendation prediction formula is: in, Represents the target user A collection of users with high similarity. For target users Average rating of viewed news, For users Average rating of viewed news. Average rating of viewed news.

[0015] Furthermore, the recommendation module selects the recommended model with a higher predicted score (greater than the set recommendation threshold) for each user according to the recommendation model obtained by the model training module. ) as the recommended news list, and sort the recommended news list according to factors such as news release time and popularity, and push the sorted recommended news to the corresponding users in real time, where the popularity calculation formula is , For News Number of views, For the number of likes, For the number of comments, For sharing amount, is the corresponding weight coefficient, and .

[0016] The above embodiments of the present invention have at least the following beneficial effects: the system of the present invention can provide users with more accurate and personalized news recommendations by collecting multi-source news data in real time and combining it with user behavior data for in-depth analysis. The collaborative filtering algorithm adopted by the system can capture changes in user interests. At the same time, the news data is structured through the data preprocessing module, which can more effectively understand and match the user's interests and improve the relevance of the recommendation.

[0017] In addition, by constructing a user-news browsing matrix, quantifying the user's browsing behavior, and sorting the recommended news based on factors such as popularity, the system can ensure that the news received by the user not only meets his or her personal interests, but also has high timeliness and popularity. This recommendation mechanism that comprehensively considers user behavior and news popularity can improve the user's reading experience and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which: Figure 1A schematic diagram of the structure of a model-based artificial intelligence real-time news recommendation system provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0019] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0020] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present invention can be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0021] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0022] Reference below Figure 1 , Figure 1 The structure diagram of the model-based artificial intelligence real-time news recommendation system provided by one embodiment of the present invention is as follows. Figure 1 As shown, a model-based artificial intelligence real-time news recommendation system 100 includes: The data collection module 101 is used to collect news data from multiple news sources in real time; The data preprocessing module 102 is used to clean and convert the collected news data; A model training module 103 is used to construct a recommendation model using a collaborative filtering algorithm; The recommendation module 104 is used to generate and push a personalized news recommendation list for each user according to the recommendation model.

[0023] It should be noted that the system includes a data collection module, whose function is to collect news data from multiple news sources in real time. The data collection module here refers to a part of the system that is responsible for obtaining the latest news content from different news platforms on the Internet. News data includes not only news titles and content, but also release time, category, source, and user behavior data on interaction with news, such as browsing, liking, commenting, and sharing.

[0024] Specifically, the data collection module will decide which news sources to collect data from based on preset rules and parameters, such as the frequency of news updates and the importance of the data. For example, parameters can be set to prioritize the collection of news platforms with high user traffic and fast update speed. At the same time, the collection of user behavior data can be achieved by cooperating with the API interface of the news platform or using web crawler technology to ensure the real-time and integrity of the data.

[0025] Preferably, the data collection module can also include some intelligent recognition technologies, such as natural language processing (NLP) technology, to identify and filter out irrelevant noise data and improve collection efficiency. In addition, some abnormal detection mechanisms can be set, such as marking abnormally high views or likes, so that the subsequent data preprocessing module can further analyze whether the data is true and valid. This can ensure the quality of the collected data and provide an accurate basis for subsequent data processing and recommendation.

[0026] In some embodiments, the news data collected by the data collection module includes news titles, news content, news release time, news categories, news sources, and user news browsing behavior data. The user news browsing behavior data includes user identification, browsing time, browsing duration, likes, comments, and sharing operation information.

[0027] It should be noted that the news data collected by the data collection module of the system includes news titles, news content, news release time, news categories, news sources, and user browsing behavior data. The news data here refers to all relevant information that constitutes a news report, while the user browsing behavior data refers to the interactive information generated by users when reading news, such as user identification, browsing time, browsing duration, likes, comments, and sharing operation information. These data provide the system with a detailed view of user behavior, which helps to build a personalized recommendation model.

[0028] Specifically, the news title and news content in the news data are the main information carriers of the news, while the news release time, news category and news source provide the contextual information of the news. The user ID in the user behavior data refers to the user's unique identification code, which is used to distinguish different users; the browsing time and browsing duration record the specific time and duration of the user's reading of the news; likes, comments and shares are the user's feedback behavior on the news content. These parameters can be set through the background management system of the news platform, or collected through the front-end interface where users interact with news content.

[0029] Preferably, the system can further refine the collection of user behavior data, for example, by setting specific algorithms to identify users' browsing habits, or by using machine learning models to predict users' preferences. In addition, the system can also use timestamps to accurately record each user interaction in order to analyze the time series changes of user behavior. For the analysis of user behavior data, cluster analysis can be used to identify user groups with similar behavior patterns, or association rule mining can be used to discover potential connections between user behaviors. These refined operation steps and alternatives help improve the system's ability to understand and predict user behavior, thereby improving the accuracy and personalization of the recommendation system.

[0030] In some embodiments, the data preprocessing module cleans the collected news data to remove noise data, duplicate data, and data with incorrect format.

[0031] It should be noted that the data preprocessing module of the system cleans the collected news data to remove noise data, duplicate data and data with incorrect format. The data preprocessing module here refers to the component in the system responsible for processing raw data to improve data quality and availability. Noise data refers to abnormal data points that do not contain useful information or interfere with data analysis; duplicate data refers to the same data that appears multiple times in the data set; and incorrectly formatted data refers to data that does not conform to the predetermined format specifications.

[0032] Specifically, the data preprocessing module performs a series of data processing operations, such as using regular expressions to identify and remove noise data, using data deduplication algorithms to remove duplicate data entries, and using data validation rules to identify and correct malformed data. These operations can be adjusted by setting specific parameters, such as defining thresholds for noise data, determining the accuracy requirements for data deduplication, and formulating specific rules for format validation. Conceptually, data cleaning can include text normalization, such as unifying the use of decimal points and commas, and standardizing date formats.

[0033] Preferably, the data preprocessing module can further refine its operation steps, such as introducing machine learning algorithms to automatically identify and classify noisy data, or using more advanced deduplication techniques, such as semantic-based deduplication, which can not only identify literally identical data, but also identify data with the same meaning but different expressions.

[0034] Furthermore, for malformed data, intelligent formatting tools can be used, which can identify multiple data formats and convert them into a unified format. These alternatives and refinement steps help improve the efficiency and accuracy of data preprocessing, providing a higher quality data foundation for subsequent data analysis and recommendations.

[0035] In some embodiments, the data preprocessing module converts the news data into a structured data format, and the specific steps include: Perform word segmentation on news data and extract keywords; Use the Word2Vec model to convert the extracted keywords into vector form. The specific steps are: Assign an initial random vector to each keyword, adjust the vector of each keyword through training data to capture its context information, and finally each keyword is represented as a vector of fixed dimension; The news content is represented as a vector, where each dimension of the vector corresponds to a specific keyword or feature, and its weight is based on the frequency of occurrence of the keyword in the news. and inverse document frequency Determine, among which is the total number of news, To contain keywords The number of news items; news Content Vector Keywords The weight of . .

[0036] It should be noted that the data preprocessing module of the system converts news data into a structured data format. The specific steps include word segmentation of news data and extraction of keywords. The structured data format here refers to a standardized way of organizing data, which makes data easier to store, process and analyze. Word segmentation refers to the process of dividing continuous text strings into independent and meaningful words or phrases. Keywords refer to words that appear frequently in the text or are crucial to understanding the text content.

[0037] Specifically, the data preprocessing module first performs word segmentation on the news data. This process can be achieved through natural language processing technology, such as using a dictionary-based matching method or a machine learning-based word segmentation algorithm. Then, keywords are extracted from the word segmentation results. This can be done through statistical methods (such as TF-IDF) or machine learning methods (such as topic models) to identify important words in the text. Parameter settings may include the selection of word segmentation algorithms, threshold settings for keyword extraction, and parameter adjustments for statistical methods such as TF-IDF. Conceptually, keyword extraction includes not only the vocabulary itself, but may also involve part-of-speech tagging, synonym recognition, etc.

[0038] Preferably, the data preprocessing module can further refine the keyword extraction process, for example, by introducing a semantic understanding model based on deep learning to more accurately identify and extract keywords. In addition, the Word2Vec model can be used to convert the extracted keywords into vector form, which involves assigning an initial random vector to each keyword and adjusting these vectors through training data to capture the contextual information of the keyword.

[0039] Furthermore, different training parameters such as window size, learning rate, and number of iterations can be set to optimize model performance. Alternatives may include using other types of word embedding models such as GloVe or FastText, which can also effectively convert text data into numerical vectors to support subsequent similarity calculations and recommendations.

[0040] In some embodiments, the data preprocessing module constructs a user-news browsing matrix , where the rows of the matrix represent users, the columns represent news, and the matrix elements Indicates user About News Quantitative value of browsing behavior. User-news browsing matrix The specific form is: in, is the number of users, is the number of news, The specific meanings are as follows: If the user About News If the browsing time of exceeds a certain threshold, ,otherwise ; Like, comment or share behaviors correspond to weighted values, such as like ,Comment ,share .

[0041] It should be noted that the data preprocessing module of the system constructs a user-news browsing matrix, where the rows of the matrix represent users, the columns represent news, and the matrix elements represent the quantitative values ​​of users' browsing behavior on news. The user-news browsing matrix here is a mathematical model used to represent the interactive relationship between users and news, and the quantitative value refers to the conversion of user behaviors (such as browsing, liking, commenting, and sharing) into quantifiable values ​​for calculation and analysis.

[0042] Specifically, the user-news browsing matrix is ​​a two-dimensional array, whose number of rows corresponds to the number of users and the number of columns corresponds to the number of news. Each element in the matrix represents the quantitative value of a specific user's browsing behavior for a specific news. For example, if the user's browsing time for the news exceeds a certain threshold, the element can be set to 1, otherwise it is 0; likes, comments or sharing behaviors can correspond to different weighted values, such as 2 for likes, 3 for comments, and 5 for sharing. Parameter settings may include determining the threshold for browsing time, weights for likes, comments and sharing, etc. Conceptually, this matrix can contain all browsing behaviors of users and how these behaviors are quantified into numerical values. Preferably, the construction of the user-news browsing matrix can be further refined, for example, by introducing a time decay factor to consider the timeliness of user behavior, that is, earlier behaviors may have less impact on current recommendations. In addition, more complex quantification methods can be used, such as considering the user's activity in different time periods, or adjusting the weight value based on user feedback. Alternatives may include using other types of matrices, such as tensors, to represent the three-dimensional relationship between users, news and time, which can more comprehensively capture the dynamic changes of user behavior. These refined steps and alternatives help improve the accuracy and adaptability of the recommendation system. In some embodiments, the matrix elements The quantification method is: If the user About News If the browsing time of exceeds a certain threshold, ,otherwise ; Like, comment or share behaviors correspond to weighted values, such as like ,Comment ,share .

[0043] It should be noted that in the data preprocessing module of the system, the specific quantification method of the matrix elements is: if the user's browsing time for news exceeds a specific threshold, the matrix element is 1, otherwise it is 0; like, comment or share behavior corresponds to a weighted value, for example, like is 2, comment is 3, and share is 5. The quantification method here refers to converting the user's behavior into a numerical value for use in the recommendation system. The threshold refers to a preset standard used to distinguish different levels of user behavior. Specifically, the specific threshold involved in the quantification method can be set according to the statistical data of the average browsing time of users, for example, it can be set to a certain percentile of the average browsing time of users. The setting of the weighted value reflects the importance of different user behaviors to the user's interest indication, and these weights can be adjusted according to the relationship between user behavior and subsequent click-through rate in historical data. Conceptually, this quantification method not only includes the user's direct interaction with the news, but can also be extended to include the results of the user's sentiment analysis of the news, such as the quantification of positive or negative comments.

[0044] Preferably, the quantification method can be further refined, such as introducing a machine learning model to dynamically adjust thresholds and weighting values. The model can continuously optimize these parameters based on user feedback and system performance. In addition, more complex user behavior models can be adopted, such as considering user activity in different time periods or preferences for different news categories. Alternatives may include the use of model-based quantification methods, such as latent factor models, which can capture the potential relationship between users and news and generate a quantitative value for each user-news pair. These refined steps and alternatives help improve the personalization and accuracy of the recommendation system.

[0045] In some embodiments, the model training module uses a collaborative filtering algorithm to build a recommendation model, including the following steps: calculating the similarity between users, using the formula: in, and Represents two different users. and Respectively represent users and users Browsed news collection, Indicates user About News Quantified value of browsing behavior, Indicates user About News The quantified value of browsing behavior. In order to avoid the impact of data sparsity on similarity calculation, a user activity penalty factor is introduced and The corrected user similarity formula is: Predict target users For unread news Interest level rating , the calculation formula is: in, Represents the target user A user set with high similarity (similarity greater than the set threshold) , For target users Average rating of viewed news in For target users Browsed news collection, For users Average rating of viewed news .

[0046] It should be noted that the model training module of the system uses collaborative filtering algorithm to build a recommendation model, including calculating the similarity between users and predicting the score of the target user's interest in the unbrowsed news. The collaborative filtering algorithm is a recommendation system algorithm that predicts the items that users may be interested in by analyzing the similarity between users. User similarity refers to the degree of similarity between two users calculated based on the users' historical behavior data.

[0047] Specifically, the model training module first calculates the similarity between users, using a specific formula to measure. This formula involves the quantitative value of the user's browsing behavior on news, and the user activity penalty factor. The user activity penalty factor is an adjustment parameter used to reduce the impact of data sparsity on similarity calculation. Parameter settings may include the weights of each item in the similarity calculation, as well as the specific calculation method of the activity penalty factor. Conceptually, the calculation of user similarity can be based on a variety of factors, such as the user's rating of news, browsing history, etc.

[0048] Preferably, the operation steps of the model training module can be further refined, such as introducing more complex similarity calculation methods, such as cosine similarity or Jaccard similarity. In addition, machine learning methods can be used to optimize the calculation of similarity, such as using neural networks to learn complex relationships between users. Alternatives include using model-based collaborative filtering methods, such as matrix decomposition techniques, which can discover potential patterns in the user-news matrix and use them to predict users' interest in unviewed news. These refined steps and alternatives help improve the accuracy and adaptability of the recommendation system.

[0049] In some embodiments, the user similarity calculation formula is: At the same time, in order to avoid the impact of data sparsity on similarity calculation, a user activity penalty factor is introduced. and The corrected user similarity formula is: It should be noted that the system uses a specific formula to measure the similarity between users. This formula involves the quantitative value of the user's browsing behavior for news and the user activity penalty factor. The user similarity here refers to the degree of similarity between two users calculated based on the user's historical behavior data, and the user activity penalty factor is an adjustment parameter used to reduce the impact of data sparsity on similarity calculation.

[0050] Specifically, the calculation formula for user similarity involves the quantitative value of the user's browsing behavior for news, that is, the user's rating of the news or the intensity of the behavior. The calculation of the user activity penalty factor can be based on the number of user behaviors, for example, using a logarithmic function to adjust the impact of user activity. Parameter settings may include the weights of each item in the similarity calculation, as well as the specific calculation method of the activity penalty factor. Conceptually, the calculation of user similarity can be based on a variety of factors, such as the user's rating of news, browsing history, etc., and the activity penalty factor is used to adjust these similarity calculations to reflect the user's actual activity level.

[0051] Preferably, the calculation of user similarity can be further refined, for example, by introducing more complex similarity calculation methods, such as cosine similarity or Jaccard similarity. In addition, machine learning methods can be used to optimize the calculation of similarity, such as using neural networks to learn complex relationships between users. Alternatives include using model-based collaborative filtering methods, such as matrix decomposition techniques, which can discover potential patterns in the user-news matrix and use them to predict user interest in unviewed news. These refined steps and alternatives help improve the accuracy and adaptability of the recommendation system.

[0052] In some embodiments, the news recommendation prediction formula is: It should be noted that the system uses a specific formula to measure when calculating the similarity between users. This formula involves the quantitative value of the user's browsing behavior for news and the user activity penalty factor. The user similarity here refers to the degree of similarity between two users calculated based on the user's historical behavior data, and the user activity penalty factor is an adjustment parameter used to reduce the impact of data sparsity problems on similarity calculations.

[0053] Specifically, the calculation formula for user similarity involves the quantitative value of the user's browsing behavior for news, that is, the user's rating of the news or the intensity of the behavior. The calculation of the user activity penalty factor can be based on the number of user behaviors, for example, using a logarithmic function to adjust the impact of user activity. Parameter settings may include the weights of each item in the similarity calculation, as well as the specific calculation method of the activity penalty factor. Conceptually, the calculation of user similarity can be based on a variety of factors, such as the user's rating of news, browsing history, etc., and the activity penalty factor is used to adjust these similarity calculations to reflect the user's actual activity level.

[0054] Preferably, the calculation of user similarity can be further refined, for example, by introducing more complex similarity calculation methods, such as cosine similarity or Jaccard similarity. In addition, machine learning methods can be used to optimize the calculation of similarity, such as using neural networks to learn complex relationships between users. Alternatives include using model-based collaborative filtering methods, such as matrix decomposition techniques, which can discover potential patterns in the user-news matrix and use them to predict user interest in unviewed news. These refined steps and alternatives help improve the accuracy and adaptability of the recommendation system.

[0055] In some embodiments, the news recommendation prediction formula is: in, Represents the target user A collection of users with high similarity. For target users Average rating of viewed news, For users The average rating of the viewed news.

[0056] It should be noted that the system predicts the target user's interest rating for unbrowsed news through a news recommendation prediction formula. The news recommendation prediction formula here is a mathematical model used to predict which news a user may be interested in based on the user's browsing history and the preferences of similar users. The target user refers to the user for whom the system is currently generating a recommendation list.

[0057] Specifically, the news recommendation prediction formula involves a set of users with high similarity to the target user, as well as the ratings of these users on the news. This formula predicts the ratings that the target user may give to unviewed news by calculating the similarity between the target user and similar users, as well as the ratings of similar users on the news. Parameter settings may include similarity thresholds, weight distribution of ratings, etc. Conceptually, this prediction formula can include factors such as the user's historical rating data, the user's activity, and the popularity of the news.

[0058] Preferably, the operation steps of the news recommendation prediction formula can be further refined, such as introducing a time decay factor to consider the timeliness of user behavior, that is, earlier behavior may have less impact on the current recommendation. In addition, a more complex user behavior model can be adopted, such as considering the user's activity in different time periods or preferences for different news categories. Alternatives include using model-based recommendation methods, such as latent factor models or deep learning models, which can capture the potential relationship between users and news and generate a prediction score for each user-news pair. These refined steps and alternatives help improve the personalization and accuracy of the recommendation system.

[0059] In some embodiments, the recommendation module selects a user with a higher predicted score (greater than a set recommendation threshold) for each user based on the recommendation model obtained by the model training module. ) as the recommended news list, and sort the recommended news list according to factors such as news release time and popularity, and push the sorted recommended news to the corresponding users in real time, where the popularity calculation formula is , For News Number of views, For the number of likes, For the number of comments, For sharing amount, is the corresponding weight coefficient, and .

[0060] It should be noted that the recommendation module of the system selects news with higher predicted scores as the recommended news list for each user based on the recommendation model obtained by the model training module, and sorts the recommended news list according to factors such as the release time and popularity of the news. The recommendation module here refers to the part of the system that is responsible for generating a personalized news recommendation list based on the user's interests and behavior data. The predicted score refers to the numerical value of the user's interest in the news predicted by the system based on the recommendation model.

[0061] Specifically, the recommendation module will select news with a score higher than the set recommendation threshold based on the predicted score to form a recommendation list. The popularity of news refers to the popularity of news calculated based on indicators such as the number of views, likes, comments, and shares. Parameter settings may include the determination of the recommendation threshold, the weight coefficients of various indicators in the popularity calculation, etc. Conceptually, the popularity calculation can include a variety of factors, such as the timeliness of the news, the frequency and quality of user interaction, etc.

[0062] Preferably, the operation steps of the recommendation module can be further refined, for example, by introducing more user behavior indicators to adjust the heat calculation, such as the user's dwell time on the news or the number of times they return to read. In addition, machine learning methods can be used to dynamically adjust the recommendation threshold and the weight coefficient of the heat calculation to optimize the recommendation effect. Alternatives include using content-based recommendation methods, such as text analysis and natural language processing technology, to enhance the understanding of news content and provide more accurate recommendations. These refined steps and alternatives help improve the personalization and accuracy of the recommendation system while ensuring the timeliness and relevance of the recommended content.

[0063] The above-mentioned embodiments of the present invention have the following beneficial effects: the system described in the present invention can effectively collect news data from multiple news sources in real time, and perform deep preprocessing, including data cleaning and text vectorization, to construct structured news data. The recommendation model constructed by the collaborative filtering algorithm can predict the user's news preferences based on the user's historical behavior and the behavior of similar users, thereby generating a personalized news recommendation list. This method can improve the accuracy of news recommendations and user satisfaction, while enhancing the user experience.

[0064] In addition, the system sorts the recommendation list by considering factors such as the release time and popularity of the news, and pushes it to users in real time, ensuring that users can obtain the most relevant and popular news content in a timely manner. The popularity calculation formula combines the number of views, likes, comments, and shares. By adjusting the weight coefficient, it can more accurately reflect the popularity of the news and further optimize the relevance and timeliness of the recommendation results. This comprehensive consideration can improve the quality of news recommendations and user interaction and participation.

[0065] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or terminal devices such as a computer, a server, a mobile phone, and a tablet.

[0066] The above descriptions are only some preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A model-based artificial intelligence real-time news recommendation system, characterized in that: Includes the following modules: A data collection module, used to collect news data from multiple news sources in real time; The data preprocessing module is used to clean the collected news data and convert the news data into a structured data format; Model training module, used to build recommendation models using collaborative filtering algorithms; The recommendation module is used to generate and push a personalized news recommendation list for each user based on the recommendation model.

2. The system according to claim 1, characterized in that The news data collected by the data collection module includes news title, news content, news release time, news category, news source and user browsing behavior data of news. The user browsing behavior data of news includes user identification, browsing time, browsing duration, likes, comments and sharing operation information.

3. The system according to claim 1, characterized in that The data preprocessing module cleans the collected news data, including removing noise data, duplicate data and data with incorrect formats.

4. The system according to claim 1, characterized in that The data preprocessing module converts the news data into a structured data format, and the specific steps include: Perform word segmentation on news data and extract keywords; Use the Word2Vec model to convert the extracted keywords into vector form. The specific steps are: Assign an initial random vector to each keyword, adjust the vector of each keyword through training data to capture its context information, and finally each keyword is represented as a vector of fixed dimension; represent the news content in the form of vectors, each dimension of the vector corresponds to a keyword or feature, and its weight is based on the frequency of occurrence of the keyword in the news. and inverse document frequency Determine, among which is the total number of news, To contain keywords The number of news; news Content Vector Keywords Weight .

5. The system according to claim 1, characterized in that The data preprocessing module constructs a user-news browsing matrix , where the rows of the matrix represent users, the columns represent news, and the matrix elements Indicates user About News Quantitative value of browsing behavior, user-news browsing matrix The specific form is: in, is the number of users, is the number of news, The specific meaning is as follows: If the user About News If the browsing time of exceeds the threshold, ,otherwise ; Like, comment or share behaviors correspond to weighted values, like ,Comment ,share .

6. The system according to claim 5, characterized in that Matrix Elements The quantitative method is: if the user About News If the browsing time of exceeds the threshold, ,otherwise ; Like, comment or share behaviors correspond to weighted values. ,Comment ,share .

7. The system according to claim 1, characterized in that The model training module uses a collaborative filtering algorithm to build a recommendation model, including the following steps: To calculate the similarity between users, use the formula: in, and Represents two different users. and Respectively represent users and users Browsed news collection, Indicates user About News Quantified value of browsing behavior, Indicates user About News The browsing behavior quantification value is used to avoid the impact of data sparsity on similarity calculation. The user activity penalty factor is introduced and The corrected user similarity formula is: Predict target users For unread news Interest level rating , the calculation formula is: in, Represents the target user The similarity of user sets with high similarity is greater than the set threshold , For target users Average rating of viewed news in For target users Browsed news collection, For users Average rating of viewed news .

8. The system according to claim 7, characterized in that The formula for calculating user similarity is: At the same time, in order to avoid the impact of data sparsity on similarity calculation, a user activity penalty factor is introduced. and The corrected user similarity formula is: .

9. The system according to claim 7, characterized in that The news recommendation prediction formula is: in, Represents the target user A collection of users with high similarity. For target users Average rating of viewed news, For users The average rating of the viewed news.

10. The system according to claim 1, characterized in that The recommendation module selects news with high prediction scores as the recommended news list for each user based on the recommendation model obtained by the model training module, and sorts the recommended news list according to factors such as news release time and popularity, and pushes the sorted recommended news to the corresponding user in real time, where the popularity calculation formula is: , For News Number of views, For the number of likes, For the number of comments, For sharing amount, is the corresponding weight coefficient, and .

Images