News recommendation mobile phone application implementation method and system based on deep learning

Abstract

The invention provides a news recommendation mobile phone application implementation method and system based on deep learning, and relates to the technical field of intelligent information recommendation. The method comprises the following steps: acquiring multi-source data, including hot news ranking list data, user calendar application data, APP use data, news stay time data and facial expression data; and analyzing facial expressions by using an emotion evaluation model constructed by deep learning to determine user emotions, and judging that the user is not interested in current pushing when the news stays for a short time and the emotions are poor. And by combining calendar and APP use data, determining interest information through a news push model constructed by deep learning, and by applying multi-dimensional data fusion and the deep learning model, an interest change sign of a user can be captured in time, a recommendation strategy can be quickly adjusted, and accurate push is performed in combination with a hot news ranking list.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent information recommendation, and in particular to a method and system for implementing a news recommendation mobile phone application based on deep learning. Background Art

[0002] With the rapid development of mobile Internet, the way people obtain information has undergone tremendous changes, and mobile phones have become the main terminal for news reading. In the era of information explosion, massive news content continues to emerge, and users are facing the dilemma of information overload, and it is difficult to quickly and accurately find news that suits their interests. With its powerful data processing and pattern recognition capabilities, deep learning technology has shown significant advantages in the field of personalized recommendations, providing an effective way to solve the problem of news recommendation. Through in-depth analysis of a large amount of user behavior data and news content data, deep learning models can learn users' interest preferences, realize personalized news recommendations, and enhance users' news reading experience.

[0003] Traditional deep learning-based technologies mainly extract features from news content, such as themes, keywords, and text sentiment, and then match them with users' existing interest tags to recommend highly matched news to users. For example, for a user who often reads positive and travel-themed news, the system will recommend travel news with similar features.

[0004] However, traditional technologies can easily trap users in information cocoons and often fail to capture the rapid changes in user interests in real time. For example, a user may suddenly become interested in news in a certain field due to an emergency in his or her life, but the recommendation system may still make recommendations based on the user's past long-term interests and fail to adjust the recommended content in time to match the user's current interests. Summary of the invention

[0005] The present application provides a method and system for implementing a news recommendation mobile application based on deep learning, which is used to timely capture signs of user interest changes so as to quickly adjust the recommendation strategy and accurately push news in combination with hot news rankings.

[0006] First aspect, the present application provides a method for implementing a news recommendation mobile application based on deep learning, which is applied to a news recommendation system. The method includes: obtaining hot news ranking list data, calendar application data of the user within a recently set time, APP usage data of the user within a recently set time, and multiple first residence time data of the user browsing news within a recently set time; obtaining user facial expression data, and determining the current emotion value of the user through an emotion evaluation model, which is constructed in advance through deep learning from multiple expression data sets with emotion value annotations; if multiple pieces of the first residence time data are all lower than a set residence time threshold, and the current emotion value is lower than a set emotion threshold, it is determined that the user is not interested in the current news push; combining the calendar application data and the APP usage data, and determining potential information that the user is suspected to be interested in according to a preset news push model, which is constructed in advance through deep learning from calendar application data and APP usage data sets with potential information annotations; combining the hot news ranking list data, and pushing news to the user terminal according to the potential information.

[0007] By adopting the above technical solution, multi-source data is obtained, including hot news ranking list data, user calendar application data, APP usage data, news residence time data, and facial expression data. The emotion evaluation model constructed by deep learning is used to analyze facial expressions to determine the user's emotion. When the news residence time is short and the emotion is not good, it is determined that the user is not interested in the current push. Subsequently, combining the calendar and APP usage data, the news push model constructed by deep learning determines the potential information. The application of multi-dimensional data fusion and deep learning model can more accurately capture the user's potential interest, laying a foundation for subsequent pushing news that meets the user's interest, thereby improving the matching degree between news recommendation and user interest.

[0008] Combined with some embodiments of the first aspect, in some embodiments, after the step of obtaining the calendar application data of the user within a recently set time, it further includes: performing semantic analysis on the calendar application data to obtain user daily activity information; if it is detected that the user daily activity information has a first keyword related to a set news field, the corresponding field is marked as the suspected interest change direction of the user, and combined with the hot news ranking list data, relevant news is pushed to the corresponding user terminal.

[0009] By adopting the above technical solution, since the calendar application data reflects the user's daily arrangements, the appearance of specific field keywords means that the user may be interested in that field. Combining the hot news ranking list data to push relevant news can quickly adjust the news push content according to the potential interest signals in the user's calendar in a timely manner, making the news recommendation more in line with the user's possible current interest changes, and improving the timeliness and pertinence of the recommendation.

[0010] In some embodiments in combination with some embodiments of the first aspect, after the step of obtaining the user's APP usage data within the recently set time, the method further includes: analyzing the changing pattern of the APP usage frequency; if the increase amplitude of the usage frequency of the APP in the set domain within the set time exceeds the set frequency change amplitude threshold, marking the corresponding domain as the suspected interest change direction of the user, and combining the hot news ranking list data to push relevant news to the corresponding user terminal.

[0011] By adopting the above technical solution, the change in the APP usage frequency intuitively reflects the change in the user's behavior tendency. High-frequency use of an APP in a certain domain indicates that the user's interest in that domain has increased. Combining the hot news ranking list data to push relevant news can capture the interest change in real time based on the user's behavior change, timely adjust the news recommendation strategy, make the recommended content more in line with the user's ever-changing interest needs, and improve the user experience.

[0012] In some embodiments in combination with some embodiments of the first aspect, in the step of combining the hot news ranking list data and pushing news to the user according to the potential information, it specifically includes: classifying and screening the hot news ranking list data to determine the news category information and the heat weight data; determining the news category information with the highest matching degree with the potential information through a text matching algorithm; according to the news category information with the highest matching degree with the potential information, selecting news with heat weight data greater than the set heat weight threshold from the hot news ranking list data to generate a candidate news list, and pushing the news in the candidate news list to the user terminal.

[0013] By adopting the above technical solution, the classification and screening make the news data more organized. The text matching algorithm accurately locates the news category that matches the user's interest, and the news with high heat weight is more attractive. This method can quickly screen out the news that not only meets the user's interest but also attracts public attention from a large number of hot news, optimize the news push content, ensure the news quality while meeting the user's interest, and improve the user's attention and satisfaction with the news recommendation.

[0014] In some embodiments in combination with some embodiments of the first aspect, after the step of combining the hot news ranking list data and pushing news to the user according to the potential information, the method further includes: obtaining multiple second residence time data of the user browsing the news in real time, calculating the total residence time data based on the multiple second residence time data within the set time; if the total residence time data shows a continuous downward trend, starting a fast interest detection mechanism; pushing a group of short news summaries covering multiple popular news domains to the user terminal; determining the actual interest change direction of the user according to the user's click behavior, and pushing the matching news.

[0015] By adopting the above technical solution, the real-time monitoring of the residence time can promptly detect the fluctuations in the user's interests. The fast interest detection mechanism can detect the user's new interests within a short time. The determination of the interest points based on the click behavior has a high accuracy, so as to quickly adjust the recommended content, closely follow the rapid changes in the user's interests, and ensure that the news recommendations always meet the user's needs.

[0016] In combination with some embodiments of the first aspect, in some embodiments, after the step of pushing news to the user according to the potential information by combining the hot news ranking data, the method further includes: obtaining the voice feedback data of the user; judging whether the user raises a question when browsing the news according to the voice feedback data. If the user raises a question when browsing the news, relevant news is pushed in combination with the question.

[0017] By adopting the above technical solution, the user's voice feedback directly reflects the current focus of attention. Associating the question with the news recommendation can provide relevant news for the user's immediate needs. This interaction method breaks the traditional one-way recommendation mode, makes the news recommendation more interactive and targeted, meets the user's immediate information needs during the browsing process, enhances the interaction between the user and the news recommendation system, and improves the user's recognition of the news recommendation service.

[0018] In combination with some embodiments of the first aspect, in some embodiments, if it is judged whether the user raises a question when browsing the news according to the voice feedback data, and if the user raises a question when browsing the news, the step of pushing relevant news in combination with the question specifically includes: using natural language processing technology to extract the second keywords in the question; combining the second keywords, using the full-text retrieval algorithm to match the news, and pushing the news to the user terminal in the order of the set matching degree.

[0019] By adopting the above technical solution, the natural language processing technology converts the user's natural language into keywords that can be understood by the machine. The full-text retrieval algorithm comprehensively searches and matches the news in the news library. The accurate extraction of keywords ensures the accuracy of the search direction, the full-text retrieval ensures comprehensive coverage, and the pushing according to the matching degree makes the most relevant news presented first, which can efficiently meet the user's news needs based on the question, quickly provide the news content that meets the user's immediate interests for the user, and improve the accuracy and efficiency of the news recommendation.

[0020] In a second aspect, the present application provides a news recommendation system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, and the computer program code includes computer instructions. The one or more processors call the computer instructions to enable the news recommendation system to execute the method described in the first aspect and any possible implementation manner in the first aspect.

[0021] In a third aspect, the present application provides a computer-readable storage medium including instructions that, when running on a news recommendation system, cause the news recommendation system to execute the methods described in the first aspect and any possible implementation manner of the first aspect.

[0022] In a fourth aspect, the present application provides a computer program product that, when running on a news recommendation system, causes the news recommendation system to execute the methods described in the first aspect and any possible implementation manner of the first aspect.

[0023] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. By adopting the technical means of obtaining multi-source data (hot news ranking data, user calendar application data, APP usage data, news stay time data, and facial expression data), and using deep learning to construct an emotion evaluation model and a news push model to analyze and process the data, the technical problem that the existing news recommendation system is difficult to accurately grasp the user's interests, resulting in a low matching degree between the recommended content and the user's actual interests, is effectively solved. Furthermore, the technical effect of accurately capturing the user's potential interests and significantly improving the matching degree between news recommendations and user interests is achieved.

[0024] 2. By adopting the technical means of obtaining APP usage data and analyzing its usage frequency change law, marking the interest change direction when the usage frequency of the APP in the set field changes beyond the threshold, and combining the hot news ranking data for news push, the technical problem that the existing technology cannot adjust news recommendations in a timely manner according to user behavior changes, resulting in the recommendation lagging behind the user's interest change, is effectively solved. Furthermore, the technical effect of real-time capturing of user interest changes and timely adjustment of recommended content to meet the user's interest needs is achieved.

[0025] 3. By adopting the technical means of obtaining user voice feedback data, when the user browses the news and asks questions, using natural language processing technology and full-text retrieval algorithms to push relevant news in combination with the questions, the technical problem that the existing news recommendation system lacks interactivity and cannot provide news for the user's immediate needs is effectively solved. Furthermore, the technical effect of enhancing the interaction between the user and the news recommendation system, meeting the user's immediate information needs, and improving the user's recognition is achieved. Brief Description of the Drawings

[0026] Figure 1 is a flowchart showing an implementation method of a news recommendation mobile application based on deep learning in an embodiment of the present application; Figure 2 is another flowchart showing an implementation method of a news recommendation mobile application based on deep learning in an embodiment of the present application; Figure 3It is a schematic structural diagram of an entity device in an embodiment of the present application's news recommendation system. Detailed implementation manners

[0027] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification and appended claims of the present application, the singular forms "a", "an", "the", "above-mentioned", "said", and "this" are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to and includes any or all possible combinations of one or more of the listed items.

[0028] Hereinafter, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as implying or suggesting relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality" is two or more.

[0029] For ease of understanding, the method provided in this embodiment is described in terms of a process below. Please refer to Figure 1 , which is a schematic flowchart of a method for implementing a news recommendation mobile application based on deep learning in an embodiment of the present application.

[0030] S101. Obtain hot news ranking list data, calendar application data of the user within a recently set time, APP usage data of the user within a recently set time, and multiple first residence time data of the user browsing news within a recently set time; The news recommendation system first obtains multiple hot news ranking list data. Specifically, when obtaining multiple hot news ranking list data, the news recommendation system will comprehensively use a variety of technical means and strategies. The system will first establish data docking with major authoritative news platforms. Taking the cooperation with mainstream news websites to obtain data as an example, the API (Application Programming Interface) docking method is usually adopted. The system sends a request to the API server of the news website, and the request contains authentication information such as a key or a token to ensure the legality and security of the request. At the same time, in the request parameters, set the type of data to be obtained (such as news ranking list data sorted by popularity, click volume, etc.), the time range (for example, hot news data in the past 24 hours, within this week), and the data format.

[0031] To obtain data comprehensively and in a timely manner, the system does not rely solely on a single news platform. Instead, it interfaces with multiple news sources of different types, including comprehensive news websites and vertical professional news platforms (such as technology, finance, entertainment, etc.). This is because different platforms have different focuses and audiences, and their hot news rankings also have their own characteristics. The integration of data from multiple platforms can more comprehensively reflect the overall picture of news hotspots.

[0032] The news recommendation system also obtains the user's calendar application data within a recently set time period. Specifically, When the news recommendation system obtains the user's calendar application data, it needs to perform secure and compliant data interaction with the calendar application on the user's mobile phone. First, according to different operating systems, the corresponding authorization mechanism is adopted to obtain the data access permission. The system guides the user to perform an authorization operation on the mobile phone. After the user agrees, the system obtains an access token. With this token, the system can send a data request to the API of the Android calendar application. In the request, the time range for obtaining data (such as calendar event data in the past week or the next two weeks) and the required data fields (such as event title, start time, end time, location, etc.) are clearly specified.

[0033] When the system receives the data returned by the calendar application, it will perform local encrypted storage on the data. For example, the AES (Advanced Encryption Standard) algorithm is used to encrypt the data to ensure the security of the data during the storage stage in the system. To extract valuable information from the calendar data, the system uses natural language processing (NLP) technology to perform semantic analysis on the calendar event titles. First, lexical analysis is carried out to decompose the title into individual words or phrases and mark their parts of speech. Then, syntactic analysis is performed to understand the grammar structure of the sentence and identify the key subject, predicate, object, etc. components. Then, through named entity recognition technology, the entities in the title, such as person names, organization names, location names, etc., are identified. For example, for an event title like "Attend the new product launch of XX Technology Company", the system can identify "XX Technology Company" as the organization name and "new product launch" as the event type, and initially judge that the event is related to the technology field.

[0034] In some embodiments, in-depth analysis can also be combined with geographical location information. After the system obtains the location information in the calendar event, it uses Geographic Information System (GIS) technology to analyze the areas where the user often activities. If the user frequently participates in activities in a certain science and technology park, it can be inferred that the user has a relatively high interest in news in the technology field. At the same time, combined with the time when the activity occurs, such as often participating in business activities on weekdays and leisure activities on weekends, the user's interest scenario can be further refined. In addition, a time series analysis algorithm is introduced to model the time pattern of the user's calendar events. By analyzing the changes in the activity frequency and type of the user in different time periods, the possible future activity arrangements of the user can be predicted. For example, if the user regularly participates in fitness activities on certain days of each month, the system can recommend fitness-related news in advance for the user, such as fitness tips, sports equipment recommendations, etc., to improve the accuracy and timeliness of the recommendation.

[0035] The news recommendation system will also obtain the user's APP usage data within the recently set time. Specifically, when the news recommendation system obtains the user's APP usage data, it interacts with the application usage record interface of the mobile operating system. Taking the iOS system as an example, the system uses the AppAnalytics API provided by Apple to obtain data. Before obtaining the data, the user needs to authorize the system to access the application usage record in the mobile phone settings. After the system obtains the authorization, it sends a request to the AppAnalytics API, and the request parameters include the time range for obtaining the data (such as the APP usage data in the past 30 days), the data granularity (such as statistics of usage data by day or by hour), etc.

[0036] In order to understand the user's behavior patterns and interest preferences more comprehensively, the system not only obtains basic data such as the number of times the APP is launched and the usage duration, but also obtains specific behavior data within the APP, such as the product search keywords and browsed product categories in the e-commerce APP, and the article classification browsing records in the reading APP. By analyzing these detailed data, the user's interest points can be grasped more accurately. For example, if the user frequently searches for electronic products and browses product pages such as mobile phones and computers in the e-commerce APP, it indicates that the user may be interested in news related to electronic products.

[0037] In terms of data processing, the system uses data mining techniques to analyze APP usage data. By applying clustering algorithms such as the K-Means clustering algorithm, the APPs used by users are clustered according to functions, user behavior patterns, etc. For example, reading APPs with similar user behaviors (such as involving a large amount of text reading and information acquisition) are grouped into one category. By analyzing the usage time distribution and behavior characteristics of users on various APPs, a user's APP usage profile is constructed, thereby inferring the user's interest areas. At the same time, using the association rule mining algorithm, the association relationships between different APP usage behaviors are discovered. For example, if a user often uses a health diet APP after using a fitness APP, then the system can consider that the user is interested in news related to a healthy lifestyle.

[0038] The news recommendation system also obtains multiple first stay time data of users browsing news. Specifically, on the mobile application side, data collection is achieved through code embedding technology. When a user opens a news detail page, the system records the start timestamp; when the user leaves the news page, whether it is closing the page, switching to other news, or performing other operations, the system records the end timestamp, and the difference between the two can obtain the browsing stay time of the news. To ensure data accuracy and integrity, the system also processes abnormal situations. For example, when the user suddenly closes the application or encounters a network interruption, the unsuccessfully uploaded stay time data will be uploaded to the server when reconnecting or restarting the application next time. After the data is transmitted to the news recommendation system server, a data cleaning strategy is adopted to remove obviously abnormal data, such as extreme values with negative stay times or far exceeding the normal reading time, which may be caused by system failures or user misoperations.

[0039] S102. Obtain the user's facial expression data and determine the user's current emotion value through an emotion assessment model, which is constructed in advance through deep learning using multiple expression data sets with emotion value annotations; The system collects the user's facial expression data by calling the camera permission of the mobile device. To ensure the stability and accuracy of data collection, advanced computer vision technologies are adopted, such as facial detection algorithms based on convolutional neural networks (CNNs). Quickly locate the position of the human face in the video stream, and then combine more complex deep learning models to accurately detect the key feature points of the face, including the contour and position information of parts such as eyes, eyebrows, and mouth.

[0040] The collected facial image data will be preliminarily processed on the local device to reduce the amount of data transmission and protect user privacy. Using image preprocessing techniques, such as grayscale and normalization, the image is converted into a format suitable for model processing. After that, the preprocessed data is transmitted to the server of the news recommendation system through a secure network channel, such as the HTTPS protocol. On the server side, the system uses an emotion evaluation model that has been constructed through deep learning using multiple expression datasets with emotion value annotations to analyze facial expression data. The dataset contains a large number of facial images with different expressions and is labeled with corresponding emotion categories (such as anger, disgust, fear, happiness, sadness, surprise, and neutrality). Based on this dataset, a deep convolutional neural network (DCNN) is used for model training. During the training process, the model learns the complex mapping relationship between facial features and emotions, and improves the accuracy of emotion recognition by continuously adjusting network parameters.

[0041] After receiving the user's facial expression data, the emotion assessment model performs forward propagation calculations on the data and outputs the probability that the user's current emotion belongs to each predefined emotion category. The system selects the emotion category with the highest probability as the user's current emotion value.

[0042] S103: if the plurality of first stay time data are all lower than the set stay time threshold, and the current emotion value is lower than the set emotion threshold, it is determined that the user is not interested in the current news push; In the news recommendation system, to determine whether the user is interested in the current news push, it is necessary to comprehensively consider the first dwell time and emotional value of browsing the news. The system first sets a reasonable dwell time threshold and emotional threshold. The dwell time threshold is determined based on statistical analysis of the average reading time of a large number of users for various types of news. For example, by collecting and analyzing the dwell time data of a large number of users reading different types of news (such as current affairs, entertainment, technology, etc.), the average dwell time of various types of news is calculated, and the threshold is set in combination with a certain standard deviation or percentile to adapt to the differences in reading habits of different users.

[0043] The emotion threshold is determined based on the output of the emotion assessment model through experiments and data analysis. For example, the emotion confidence output by the emotion assessment model is divided into multiple intervals. After multiple tests and user feedback, a lower confidence interval is set as the emotion threshold for judging that the user is not interested. For example, when the output confidence of the emotion assessment model for a certain negative emotion (such as boredom, annoyance) is lower than 0.3, it is considered that the current emotion value is lower than the set emotion threshold.

[0044] After the system obtains the user's browsing stay time data and mood value, it uses an efficient conditional judgment algorithm for judgment. On the server side, by writing logic code, each of the multiple stay time data is compared with the set stay time threshold one by one, and at the same time the current mood value is compared with the set mood threshold. If all the stay time data are lower than the set stay time threshold, and the current mood value is also lower than the set mood threshold, the system determines that the user is not interested in the current news push.

[0045] S104. Combine the calendar application data and the APP usage data, and determine the potential information that the user is suspected to be interested in according to the preset news push model. The news push model is constructed in advance through deep learning based on the calendar application data and APP usage data sets with potential information annotations. For the calendar application data, natural language processing (NLP) technology is used for comprehensive analysis. With the help of a lexical analysis tool, the calendar event title is split into individual words or phrases, and the part of speech is marked to identify key nouns, verbs, etc. For example, for the event title "Attend the XX Technology Forum Lecture", key information such as "Technology Forum Lecture" can be accurately identified. Then, through syntactic analysis, the grammatical structure of the sentence is understood, the core semantics is determined, and the nature and relationship of the event are further clarified. Using named entity recognition (NER) technology, entity information such as the organization name "XX Technology" is extracted from the title, and it can be initially judged that the event is related to the technology field. In addition to text analysis, the system also combines the time and location information of the calendar event. Using geographic information system (GIS) technology, the areas where the user often activities are analyzed. If the user frequently participates in activities in the technology park and the activity time is concentrated on weekdays, this further strengthens the possibility that the user is interested in business news in the technology field.

[0046] For the APP usage data, the system uses data mining technology for in-depth analysis. The clustering algorithm, such as the K-Means algorithm, is used to classify the APPs used by the user according to the functions of the APP, user behavior patterns, etc. For example, reading APPs with similar user behaviors, such as those involving a large amount of text reading and information acquisition, are grouped into one category. By analyzing the usage time distribution, usage frequency of the user on various APPs, and specific behavior data within the APP, such as the product search keywords and browsing product categories in the e-commerce APP, and the article classification browsing records in the reading APP, etc., a user's APP usage portrait is constructed. For example, if the user frequently searches for electronic products and browses product pages such as mobile phones and computers in the e-commerce APP, combined with the clustering analysis results, it can be judged that the user may be interested in news related to electronic products.

[0047] After completing data processing and feature extraction, the system inputs the processed calendar application data and APP usage data into a preset news push model. This news push model is constructed in advance based on multiple calendar application data and APP usage data sets with potential information annotations, using deep learning techniques. Taking models based on recurrent neural networks (RNNs), such as long short-term memory networks (LSTMs) or gated recurrent units (GRUs), as an example, these models can effectively process sequential data and learn the time series features and long-term dependencies in the data. During the training process, a large amount of calendar application data and APP usage data are used as inputs, and the corresponding potential information is used as output labels. The model parameters are continuously adjusted through the backpropagation algorithm, enabling the model to accurately predict user interest information based on the input data.

[0048] When the system obtains real-time calendar application data and APP usage data, it inputs them into the trained news push model. The model outputs multiple potential information suspected by the user. These potential information can be specific news topics, such as "The latest developments in artificial intelligence" and "Technological breakthroughs in new energy vehicles"; they can also be news categories, such as "Science and technology news" and "Financial news"; or interest tags, such as "Digital enthusiasts" and "Travel experts", etc.

[0049] In some embodiments, the following technical means can be used to achieve real-time monitoring and precise recommendation of user interest changes: The system first maintains a preset APP domain classification table, classifying each APP according to its main functions and service attributes, such as classifying into categories like "News", "Video", "Game", etc. The server divides the user's APP usage data according to this classification table to obtain different categories of APPs used by the user, that is, multiple classified APP usage domains. By collecting the user's long-term various behavior data, a fingerprint portrait of each user is constructed, that is, an interest feature set. The user's fingerprint portrait is matched with the preset domain features to determine which features in the fingerprint portrait are strongly correlated with a certain domain, thereby determining the user's stable preferences in which domains, that is, determining multiple hobby domains that the user is interested in. The user's hobby domains are compared and analyzed with the classified APP usage domains: (1) Statistically analyze the usage duration or usage frequency of each classified APP usage domain; (2) Calculate the weights of each classified APP usage domain according to the preset weight algorithm; (3) Select the classified APP usage domains with higher weight rankings but not belonging to the hobby domains; (4) Determine these selected classified APP usage domains as the user's suspected interest change domains.

[0050] For each suspected conversion domain, the server counts the usage frequency of this type of APP in the recent period and compares it with the average usage frequency of this type of APP in the historical period: (1) collect the number of times this type of APP is used in the past week and calculate the average daily usage frequency A; (2) Query historical statistical data to obtain the average daily usage frequency B of this type of APP in the past month; (3) Calculate the frequency increase = (AB) / B; (4) If the magnitude exceeds a preset threshold, it is confirmed that the user’s interest has changed and relevant news is pushed.

[0051] Through the above-mentioned technical means, the direction of user interest change can be effectively determined, and hot news can be used for recommendation, so that the recommended content meets the user's current actual interest needs.

[0052] S105: Combine the hot news ranking data and push news to the user terminal according to the potential information.

[0053] The system first classifies and screens the hot news ranking data. Using a classification algorithm, hot news is divided into different categories, such as political news, entertainment news, technology news, sports news, etc., according to the subject, field, type and other attributes of the news. At the same time, each piece of news is given a heat weight data. The calculation of the heat weight takes into account multiple factors, including the number of clicks, comments, and shares of the news. For example, if a piece of news receives a large number of clicks and comments in a short period of time, its heat weight will increase accordingly.

[0054] Next, the system uses a text matching algorithm to determine the news category information that best matches the potential information. The word vector model in natural language processing is used to convert the potential information and news category information into vector representations. By calculating the similarity between vectors, such as cosine similarity, the news category that best matches the potential information is found. For example, if the potential information is "innovative breakthroughs in the field of science and technology", the system may determine that the "science and technology news" category is the best matching news category through a text matching algorithm.

[0055] From the matching news category information, the system selects multiple news with heat weight data greater than the set heat weight threshold as the candidate news list. The heat weight threshold is set to ensure that the recommended news has a certain degree of heat and attention, and avoid recommending news that is too unpopular. For example, if the heat weight threshold is set to 500, only news with a heat weight exceeding 500 will be selected into the candidate news list.

[0056] After determining the candidate news list, the system pushes these news to the user side. During the pushing process, considering the network status and performance of the user device, an adaptive pushing strategy is adopted. For users with good network conditions and strong device performance, news content in high-definition and rich media forms, such as news containing pictures and videos, is pushed; for users with poor network conditions or limited device performance, a simplified version of the news content, such as pure text news, is pushed to ensure that users can quickly load and view the news.

[0057] In the embodiments of the present application, by performing semantic analysis on calendar application data and monitoring APP usage, the system can keenly detect signs of changes in user interests, and timely adjust the recommended content in combination with the hot news ranking list, so that the recommendation closely follows the changes in user interests, solves the problem that the recommendation lags behind the change in user interests, and improves the user experience.

[0058] In some embodiments, in order to better meet the immediate needs of users during the news browsing process, the system can continuously obtain user facial expression data, and then input the facial expression data into the aforementioned emotion assessment model to obtain a real-time emotion value. When it is detected that the real-time emotion value rises compared with the previous moment and the change amplitude exceeds a preset emotion change threshold (such as the emotion probability increase ≥ 20%), it is determined that the potential information currently recommended may conform to the user's interests and is marked as "effective".

[0059] It is also possible to record the time data of user emotion changes; obtain multiple historical time data of user emotion changes; calculate the potential information game strength in combination with the time data and historical time data. The potential information game strength refers to the confidence in the conversion of user interests into potential information; if the potential information game strength is lower than the set standard game strength, a fast interest detection mechanism is started after a set time. Specifically, when a significant change in the emotion value is detected (whether it rises or falls), the timestamp (accurate to seconds), the current emotion value, the potential information label corresponding to the recommendation, etc. are automatically recorded and stored in the user behavior log database. The historical time data is managed through a time series database, which supports quick query of emotion fluctuation records within the past N days (such as the time points of all emotion increase events in the past 30 days).

[0060] Initialize according to the frequency of the same type of potential information triggering an emotion increase in historical data (for example, if a certain technology-related potential information triggers an emotion increase 8 times in the past 10 recommendations, the basic confidence is 80%). The recent emotion change data has a greater impact on the confidence. Through the formula decay coefficient = (t is the current time, is the historical event time, is the decay constant) to weight the historical data, and the older the data, the lower the weight.

[0061] Then determine , if the calculation result is lower than the set standard game strength (such as 60%), it indicates that the system has insufficient confidence in the user's interest turning to this potential information, and it is necessary to activate the fast interest detection mechanism. This interest detection mechanism includes pushing a group of short news summaries covering multiple popular news fields to the user side; determining the actual interest change direction of the user according to the user's click behavior, and pushing the matching news. Through the above technical means, the system can dynamically verify the effectiveness of potential information, evaluate the recommendation confidence in combination with the time law of user emotion change, and quickly calibrate the interest model through the lightweight detection mechanism when necessary, so as to achieve the dynamic matching of accurate recommendation and user needs.

[0062] After combining the above content, the following is a further and more specific process description of the method provided in this embodiment. Please refer to Figure 2 , which is another process schematic diagram of the method for implementing a news recommendation mobile application based on deep learning in the embodiments of the present application.

[0063] S201. Obtain multiple second stay time data of the user browsing news in real time, and calculate the total stay time data according to the multiple second stay time data within the set time; After the step of pushing news to the user according to the potential information in combination with the hot news ranking list data, the system can use code embedding technology to record the user's browsing behavior. Specifically, when the user opens a news, the application will automatically trigger an event to record the timestamp of opening the news; when the user closes the news page, swipes to other news or has no operation after a certain period of time, the event is triggered again to record the closing timestamp, and the difference between the two is the second stay time of the user browsing this news. After the user browses multiple news, the system will set a specific time period. Within this set time, the system will collect multiple second stay time data generated by the user browsing each news. Then, these data are accumulated and calculated, and the final result is the total stay time data. This code embedding method can achieve accurate timing and can adapt to different application scenarios and user operation habits.

[0064] S202. If the total stay time data shows a continuous downward trend, activate the fast interest detection mechanism; The news recommendation system will regularly monitor and analyze the total browsing time of users within a set time period, which can be adjusted according to the system's requirements and users' behavioral habits. To determine whether the total stay time data shows a continuous downward trend, the system uses a time series analysis algorithm (such as the ARIMA model). The ARIMA model can model and predict time series data, and by analyzing features such as the trend, seasonality, and periodicity of the data, it can judge whether there is a downward trend in the data. When the system detects that the total browsing time data of users within the set time period shows a continuous downward trend, it indicates that the users' interest in the currently recommended news is gradually decreasing. At this time, it is necessary to activate the fast interest detection mechanism.

[0065] S203. Push a set of short news summaries covering multiple popular fields to the user side; After the news recommendation system determines to activate the fast interest detection mechanism, the system first collects news data from multiple data sources, including mainstream news websites, social media platforms, industry vertical websites, etc., and automatically crawls the news content on these websites according to preset rules. During the crawling process, key information such as the title, summary, release time, and field of the news is extracted.

[0066] To generate high-quality short news summaries, the system uses natural language processing (NLP) technology. Among them, the TextRank algorithm is widely used. It is based on a graph model, regards the sentences in the news text as nodes in the graph, evaluates the importance of sentences by calculating the edge weights between nodes, and then extracts the key sentences as news summaries.

[0067] S204. Determine the new interest points of users according to their click behaviors and push matching news.

[0068] After pushing the short news summaries to the user side, the system records the click behaviors of users on the news summary display page through code embedding technology. When a user clicks on a news summary, the system will immediately capture this event and send relevant information, such as the click time, the field and title of the clicked news summary, etc., to the server side for analysis.

[0069] After receiving the click data, the server side judges whether the user has developed a new interest in the field according to the field of the clicked news and the user's previous interest preferences. For example, if the user has rarely paid attention to the travel field before but clicks on a travel-related news summary, the system will focus on analyzing the travel field as a potential new interest point. At the same time, the system will analyze the user's click frequency and duration. If the user frequently clicks on news summaries in the same field, or reads a certain news for a long time, it indicates that the user has a strong interest in this field.

[0070] In the embodiments of the present application, by obtaining the dwell time data of the user browsing news in real time, starting the fast interest detection mechanism, pushing short news summaries covering multiple fields, and determining new interest points based on click behavior to push matching news, the dynamic tracking and accurate grasp of the user's interests are realized. It not only effectively solves the problems that the traditional news recommendation system is difficult to capture the changes of the user's interests in real time and the recommended content is lagging behind, but also significantly improves the matching degree between news recommendation and the user's interests, enhancing the user experience; it breaks the limitation of the information cocoon, broadens the fields for the user to obtain news, enables the user to access richer and more diverse news content, and meets the diverse needs of the user in information acquisition; moreover, it improves the interactivity and intelligence level of the news recommendation system, enables the system to adjust the recommendation strategy in a timely manner according to the user's real-time feedback, provides a more personalized and considerate news recommendation service for the user, and enhances the competitiveness of the news application in the fierce market competition.

[0071] The news recommendation system in the embodiments of the present invention application will be described from the perspective of hardware processing. Please refer to Figure 3 , which is a schematic structural diagram of an entity device of the news recommendation system in the embodiments of the present application.

[0072] It should be noted that Figure 3 the structure of the news recommendation system shown is only an example, and should not bring any limitations to the functions and usage scopes of the embodiments of the present invention.

[0073] As Figure 3 shown, the news recommendation system includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 302 or the program loaded from the storage section 308 into the random access memory (RAM) 303, such as executing the method described in the above embodiments. In the RAM 303, various programs and data required for system operation are also stored. The CPU 301, ROM 302, and RAM 303 are connected to each other through a bus 304. The input / output (I / O) interface 305 is also connected to the bus 304.

[0074] The following components are connected to the I / O interface 305: an input section 306 including an audio input device, a button switch, etc.; an output section 307 including a liquid crystal display (LCD), an audio output device, an indicator light, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. The drive 310 is also connected to the I / O interface 305 as needed. A removable medium 311 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is installed on the drive 310 as needed so that a computer program read from it can be installed into the storage section 308 as needed.

[0075] Specifically, according to an embodiment of the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present invention includes a computer program product that includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for performing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication section 309, and / or installed from the removable medium 311. When the computer program is executed by the central processing unit (CPU) 301, various functions defined in the present invention are executed.

[0076] It should be noted that specific examples of the computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, the computer-readable storage medium can be any tangible medium that contains or stores a program, and the program can be used by or combined with an instruction execution system, apparatus, or device.

[0077] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present invention. Among them, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and the above-mentioned module, program segment, or part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order from that marked in the accompanying drawings.

[0078] Specifically, the news recommendation system of this embodiment includes a processor and a memory. A computer program is stored on the memory. When the computer program is executed by the processor, it implements the method for implementing a news recommendation mobile application based on deep learning provided in the above-mentioned embodiment.

[0079] On the other hand, the present invention also provides a computer-readable storage medium. This storage medium may be included in the news recommendation system described in the above-mentioned embodiment; or it may exist alone without being assembled into the news recommendation system. The above storage medium carries one or more computer programs. When the above one or more computer programs are executed by a processor of the news recommendation system, the news recommendation system is enabled to implement the method for implementing a news recommendation mobile application based on deep learning provided in the above-mentioned embodiment.

[0080] As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the various embodiments of the present application.

[0081] As used in the above embodiments, depending on the context, the term "when..." may be interpreted to mean "if...", or "after...", or "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if detecting (the stated condition or event)" may be interpreted to mean "if determining...", or "in response to determining...", or "when detecting (the stated condition or event)", or "in response to detecting (the stated condition or event)".

[0082] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above-described embodiments can be implemented by a computer program instructing relevant hardware. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-described method embodiments. The foregoing storage medium includes various media that can store program codes, such as ROM, random access memory (RAM), magnetic disks, or optical discs.

Claims

1. A method for implementing a news recommendation mobile application based on deep learning, applied to a news recommendation system, characterized in that The method includes: Obtaining hot news ranking data, calendar application data of the user within a recently set time, APP usage data of the user within a recently set time, and multiple first stay time data of the user browsing news within a recently set time; Obtaining user facial expression data, and determining the current emotion value of the user through an emotion evaluation model, where the emotion evaluation model is pre-constructed through deep learning using multiple expression data sets with emotion value annotations; If multiple of the first stay time data are all lower than a set stay time threshold, and the current emotion value is lower than a set emotion threshold, it is determined that the user is not interested in the current news push; Combining the calendar application data and the APP usage data, and determining potential information that the user is suspected to be interested in according to a preset news push model, where the news push model is pre-constructed through deep learning using calendar application data and APP usage data sets with potential information annotations; Combining the hot news ranking data, and pushing news to the user terminal according to the potential information.

2. The method according to claim 1, wherein After the step of obtaining the calendar application data of the user within a recently set time, it further includes: Performing semantic analysis on the calendar application data to obtain user daily activity information; If a first keyword related to a set news field is detected in the user daily activity information, the corresponding field is marked as the suspected interest change direction of the user, and relevant news is pushed to the corresponding user terminal in combination with the hot news ranking data.

3. The method according to claim 1, wherein After the step of obtaining the APP usage data of the user within a recently set time, it further includes: Classifying the APP usage data according to a preset domain classification table to obtain multiple classified APP usage domains; Obtaining a user fingerprint portrait, where the user fingerprint portrait is a pre-extracted set of user interest characteristics; Determining multiple hobby fields that the user is interested in in combination with the user fingerprint portrait; Combining the hobby fields and the classified APP usage domains to determine multiple suspected transition fields that the user is involved in except for the hobby fields within a recently set time; If the increase amplitude of the APP usage frequency corresponding to the suspected transition field within a set time exceeds a set frequency change amplitude threshold, the corresponding suspected transition field is marked as the user's interest change direction, and relevant news is pushed to the corresponding user terminal in combination with the hot news ranking data.

4. The method according to claim 1, wherein After the step of combining the hot news ranking data and pushing news to the user according to the potential information, it further includes: Real-time obtaining multiple second stay time data of the user browsing news, and calculating total stay time data based on the multiple second stay time data within a set time; If the total stay time data shows a continuous downward trend, start a fast interest detection mechanism; Pushing a set of short news summaries covering multiple popular news fields to the user terminal; Determining the actual interest change direction of the user according to the user's click behavior, and pushing matching news.

5. The method according to claim 1, characterized in that, After the step of combining the hot news ranking data and pushing news to the user according to the potential information, it further includes: Obtaining the voice feedback data of the user; Determine whether the user asks questions while browsing news according to the voice feedback data. If the user asks questions while browsing news, relevant news is pushed in combination with the questions.

6. The method according to claim 1, wherein After the step of pushing news to the user terminal according to the potential information in combination with the hot news ranking list data, the method further includes: Continuously obtain the user's facial expression data, and determine the user's real-time emotion value according to the user's facial expression data; If it is detected that the real-time emotion value rises and the change range exceeds the set emotion transformation threshold, it is determined that the potential information is valid.

7. The method according to claim 6, characterized in that, After the step of determining that the potential information is valid if it is detected that the real-time emotion value rises and the change range exceeds the set emotion transformation threshold, the method further includes: Record the time data of the user's emotion change; Obtain multiple historical time data of the user's emotion change; Calculate the potential information game strength in combination with the time data and the historical time data. The potential information game strength refers to the confidence level of the user's interest turning into potential information; If the potential information game strength is lower than the set standard game strength, start the fast interest detection mechanism after a set time.

8. A news recommendation system, characterized in that, The news recommendation system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the news recommendation system to execute the method according to any one of claims 1-7.

9. A computer-readable storage medium, comprising instructions, characterized in that, When the instruction runs on the news recommendation system, the news recommendation system is enabled to execute the method according to any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product runs on the news recommendation system, the news recommendation system is enabled to execute the method according to any one of claims 1-7.

Images