User data mining method and system for diversified customized meteorological services

Abstract

The present application relates to the field of artificial intelligence and meteorological service technology, and in particular provides a user data mining method and system for diversified customized meteorological services, which introduces a discrete structured behavior semantic embedding network and a discrete structured push preference parsing network, and achieves accurate identification of user needs and optimization of personalized push by deeply mining the semantic information and push preferences of user session behaviors. At the same time, the embodiment of the present application also uses past data of user session behaviors to debug behavioral variations, continuously improve network performance, and ensure the accuracy and timeliness of push services. In addition, a multi-decision algorithm is adopted to make push decisions by comprehensively considering multiple factors, ensuring that the push content not only meets the actual needs of users, but also improves user experience and service quality. It can be seen that by introducing advanced network technology and decision-making algorithms, accurate identification of user needs and optimization of personalized push are achieved, which significantly improves user experience and service quality.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence and meteorological service technology, and in particular to a user data mining method and system for diversified customized meteorological services. Background Art

[0002] With the rapid development of Internet technology and the advent of the big data era, digital weather services have become an important way for people to obtain weather information. However, traditional digital weather services often adopt a "one-size-fits-all" approach when pushing content, lacking in-depth understanding and satisfaction of users' personalized needs. This leads to a mismatch between pushed content and users' actual needs, reducing user experience and service quality.

[0003] In order to solve this problem, the embodiment of the present application proposes a personalized meteorological service push method based on user conversation behavior. The method first obtains the diversified customized push text of digital meteorological services and meteorological user conversation behavior data, and then identifies user needs based on this. However, existing technical methods often have problems of low accuracy and poor understanding in terms of user needs identification, and cannot effectively capture the user's true intentions and preferences. Summary of the invention

[0004] In order to improve the above problems, the present application provides a user data mining method and system for diversified customized meteorological services.

[0005] The present application embodiment provides a user data mining method for diversified customized meteorological services, which is applied to a user data mining system. The method includes:

[0006] Obtain the diversified customized push texts and meteorological user conversation behavior data of digital meteorological services;

[0007] The diverse customized push text is used as a user demand identification indication, and the meteorological user conversation behavior data is subjected to behavioral semantic embedding mining through a discrete structured behavioral semantic embedding network, and personalized demand feedback elements are configured for the output of the behavioral semantic embedding mining to obtain a number of push preference decision knowledge features of the meteorological user conversation behavior data;

[0008] The multi-modal customized push text is used as a user demand identification indication, and the push preference decision knowledge features are analyzed through a discrete structured push preference analysis network to obtain a number of meteorological content push preference viewpoints; the discrete structured behavior semantic embedding network and the discrete structured behavior semantic embedding network are obtained by debugging behavior variation based on past user session behavior data;

[0009] According to the meteorological content push preference viewpoint, a meteorological service push decision is made on the meteorological user conversation behavior data through a multivariate decision algorithm to obtain a meteorological service push label corresponding to each meteorological content push preference viewpoint.

[0010] Preferably, the diversified customized push text is used as a user demand identification indication, the meteorological user conversation behavior data is subjected to behavioral semantic embedding mining through a discrete structured behavioral semantic embedding network, and personalized demand feedback elements are configured for the output of the behavioral semantic embedding mining, so as to obtain several push preference decision knowledge features of the meteorological user conversation behavior data, including:

[0011] The diversified customized push text is used as a user demand identification indication, and a discrete structured behavior semantic embedding network is used to extract multi-level behavior semantics from the meteorological user session behavior data to obtain multi-level user session behavior semantics;

[0012] Retrieving the associated session behavior semantics of the multi-level user session behavior semantics in the push mapping feature coordinate system through a preset matching rule;

[0013] Configuring personalized demand feedback elements for the associated session behavior semantics to obtain a plurality of session behavior push associated semantics;

[0014] Generate push preference decision knowledge features corresponding to each session behavior push association semantics.

[0015] Preferably, configuring personalized demand feedback elements for the associated session behavior semantics to obtain a plurality of session behavior push associated semantics includes:

[0016] Downsampling the associated session behavior semantics through a dynamic semantic downsampling algorithm to obtain page operation downsampling behavior semantics;

[0017] Performing deep residual processing on the semantics of the downsampled behavior of the page operation, and obtaining feature influence coefficients of different attention dimensions in the semantics of the downsampled behavior of the page operation based on the output of the deep residual processing;

[0018] Based on the discrete thermal analysis strategy, the feature influence coefficients of the different attention dimensions are configured with demand feedback definition features to obtain semantic definition coefficients of the different attention dimensions;

[0019] According to the semantic definition coefficients of the different attention dimensions and the page operation downsampling behavior semantics, a plurality of session behavior push associated semantics are obtained.

[0020] Preferably, the diversified customized push text is used as a user demand identification indication, and the push preference decision knowledge features are analyzed through a discrete structured push preference analysis network to obtain several meteorological content push preference viewpoints including:

[0021] Performing feature projection on a plurality of the push preference decision knowledge features to obtain a plurality of push preference viewpoint scene projection features;

[0022] The diverse customized push text is used as an indication of user demand identification, and a preference viewpoint identification operation is performed on each push preference viewpoint scene projection feature through a discrete structured push preference analysis network to determine the meteorological content push preference viewpoint corresponding to each push preference viewpoint scene projection feature.

[0023] Preferably, the method further comprises:

[0024] Obtain several data category labels of the user's past session behavior data;

[0025] According to the data category labels, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are debugged for behavior variation, so as to obtain the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network.

[0026] Preferably, the step of performing behavior variation debugging on the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged based on the data category label to obtain the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network includes:

[0027] Sequentially configuring different levels of subscription interest point features for each of the data category tags to obtain subscription interest description text for each data category tag under the influence of the subscription interest point features at different levels;

[0028] According to the subscription interest description text under the influence of the subscription interest point features at different levels of each data category label, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are sequentially subjected to behavior description derivation to obtain the subscription interest error variables, behavior description derived error variables and confidence coefficient thermal error variables corresponding to the subscription interest point features at different levels;

[0029] Based on the subscription interest error variable, the behavior description derived error variable and the confidence coefficient thermal error variable, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are trained and debugged to obtain the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network.

[0030] Preferably, according to the subscription interest description text under the influence of the subscription interest point features at different levels of each data category label, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are sequentially subjected to behavior description derivation, and the subscription interest error variables corresponding to the subscription interest point features at different levels are obtained, including:

[0031] Combining the core subscription interest description text of each data category tag with the past data of the user session behavior to obtain the session subscription description combination text of each data category tag, wherein the core subscription interest description text is the subscription interest description text configured with the highest-level subscription interest point feature;

[0032] The session subscription description combination text is subjected to behavioral semantic embedding mining through the discrete structured behavioral semantic embedding network to be debugged, and personalized demand feedback elements are configured for the output of the behavioral semantic embedding mining to obtain the push preference decision knowledge features corresponding to each session subscription description combination text;

[0033] Performing a preference viewpoint recognition operation on the push preference decision knowledge features through the discrete structured push preference parsing network to be debugged, thereby obtaining the identified key subscription interest point features;

[0034] Based on the core subscription interest description text and the key subscription interest point features of each data category label, generating a subscription interest description prediction text set for each data category label;

[0035] Based on the subscription interest description prediction text set of each data category label and the subscription interest description priori text set of each data category label, a subscription interest error variable of key subscription interest point features is determined.

[0036] Preferably, the step of performing a preference viewpoint identification operation on the push preference decision knowledge features through the discrete structured push preference parsing network to be debugged to determine the identified key subscription interest point features includes:

[0037] Through the discrete structured push preference parsing network to be debugged, the past data of the user's session behavior is repeated in a session behavior log to obtain a user session behavior repetition log;

[0038] The push preference decision knowledge feature is mapped to the session behavior retelling feature coordinate system of the user session behavior retelling log to obtain the identified key subscription interest point features.

[0039] Preferably, the method further comprises:

[0040] Comparing the past data of the user session behavior with the retelling log of the user session behavior to obtain a behavior description derived error variable corresponding to the key subscription interest point features;

[0041] The feature influence coefficient and semantic definition coefficient of the push preference decision knowledge feature in different attention dimensions are obtained, and the confidence coefficient thermal error variable corresponding to the key subscription interest point feature is obtained.

[0042] Preferably, based on the subscription interest error variable, the behavior description derived error variable and the confidence coefficient thermal error variable, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are trained and debugged to obtain the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network, which includes:

[0043] Based on the subscription interest error variable, the behavior description derived error variable and the confidence coefficient thermal error variable, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are trained and debugged to obtain the discrete structured behavior semantic embedding network to be verified and the discrete structured push preference parsing network to be verified;

[0044] Obtaining user session behavior verification data and a priori push tags for the user session behavior verification data;

[0045] Inputting the user session behavior verification data into the discrete structured behavior semantic embedding network to be verified and the discrete structured push preference parsing network to be verified to obtain a weather service push verification tag;

[0046] Perform consistency judgment on the meteorological service push verification tag and the prior push tag to obtain a consistency judgment weight, and a first evaluation index and a second evaluation index of a probability model of the meteorological service push verification tag and the prior push tag;

[0047] Determining a discrete evaluation index according to the first evaluation index and the second evaluation index;

[0048] A global network quality evaluation value is obtained according to the consistency judgment weight and the discrete evaluation index. When the global network quality evaluation value is higher than the set evaluation value, the discrete structured behavior semantics to be verified is embedded in the network as a discrete structured behavior semantics embedding network, and the discrete structured push preference parsing network to be verified is used as a discrete structured push preference parsing network.

[0049] Preferably, the diverse customized push text includes a page push message; the method further includes: performing page push message identification processing based on the meteorological service push tag to generate a page push message identification result; and issuing the page push message identification result.

[0050] An embodiment of the present application provides a user data mining system, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above method.

[0051] An embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. The computer program implements the above method when running.

[0052] The embodiment of the present application introduces a discrete structured behavior semantic embedding network and a discrete structured push preference parsing network, and realizes accurate identification of user needs and optimization of personalized push by deeply mining the semantic information and push preferences of user session behaviors. At the same time, the embodiment of the present application also uses the past data of user session behaviors to debug behavior variations, continuously improve network performance, and ensure the accuracy and timeliness of push services.

[0053] In traditional digital weather services, push decisions are usually based on simple rules or models, which make it difficult to fully consider the personalized needs and preferences of users. To address this problem, the embodiment of the present application adopts a multi-factor decision-making algorithm to comprehensively consider multiple factors to make push decisions, ensuring that the push content not only meets the actual needs of users, but also improves user experience and service quality.

[0054] In summary, the embodiments of the present application aim to address the shortcomings of traditional digital meteorological services in terms of personalized push content. By introducing advanced network technologies and decision-making algorithms, accurate identification of user needs and optimization of personalized push are achieved, significantly improving user experience and service quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flowchart of a user data mining method for diversified customized meteorological services provided in an embodiment of the present application.

[0056] Figure 2 A schematic diagram of the structure of a user data mining system 200 provided in an embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to better understand the above technical scheme, the technical scheme of the present application is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical scheme of the present application, rather than limitations on the technical scheme of the present application. In the absence of conflict, the embodiments of the present application and the technical features in the embodiments can be combined with each other.

[0058] Figure 1A user data mining method for diversified customized meteorological services is shown, which is applied to a user data mining system. The method includes the following steps 101 to 104.

[0059] Step 101: The user data mining system obtains the diversified customized push texts of digital meteorological services and meteorological user conversation behavior data.

[0060] Step 102: The user data mining system uses the diverse customized push text as an indication of user demand identification, performs behavioral semantic embedding mining on the meteorological user conversation behavior data through a discrete structured behavioral semantic embedding network, and configures personalized demand feedback elements on the output of the behavioral semantic embedding mining to obtain a number of push preference decision knowledge features of the meteorological user conversation behavior data.

[0061] Step 103, the user data mining system uses the diverse customized push text as an indication of user demand identification, and performs push preference analysis on the push preference decision knowledge features through a discrete structured push preference analysis network to obtain a number of meteorological content push preference viewpoints; the discrete structured behavioral semantic embedding network and the discrete structured behavioral semantic embedding network are obtained by debugging behavioral variations based on past data of user session behavior.

[0062] Step 104: The user data mining system makes a meteorological service push decision on the meteorological user conversation behavior data according to the meteorological content push preference viewpoint through a multivariate decision algorithm to obtain a meteorological service push label corresponding to each meteorological content push preference viewpoint.

[0063] In one application scenario, the user data mining system is used to optimize the push strategy of digital weather services.

[0064] First, in step 101, the user data mining system obtains the diversified customized push texts and the meteorological user conversation behavior data from the digital meteorological service platform. The meteorological user conversation behavior data includes the interaction records between the user and the meteorological service platform, such as query history, browsing history, and feedback after receiving push. At the same time, the user data mining system also collects various diversified customized push texts received by the user. The diversified customized push texts are personalized meteorological service information generated based on the user's geographical location, historical query and other information.

[0065] Next, in step 102, the user data mining system uses the obtained diverse customized push texts as an indication for identifying user needs. The user data mining system uses a discrete structured behavioral semantic embedding network to perform in-depth mining of meteorological user conversation behavior data. The discrete structured behavioral semantic embedding network can capture the subtle differences in user behaviors and convert these behaviors into quantifiable data features. The user data mining system further configures personalized demand feedback elements for the outputs of these behavioral semantic embedding mining, thereby obtaining several push preference decision knowledge features of meteorological user conversation behavior data. The push preference decision knowledge features reflect the user's preferences and needs for meteorological service content.

[0066] In step 103, the user data mining system continues to use the discrete structured push preference parsing network to parse the push preference decision knowledge features obtained in step 102. This process aims to extract the user's explicit preference for meteorological content from complex behavioral data. Through the analysis, the user data mining system obtains several meteorological content push preference opinions. The meteorological content push preference opinions are based on the user's actual behavior and feedback and have a high reference value. It is worth mentioning that the discrete structured behavioral semantic embedding network and the discrete structured push preference parsing network are both obtained by debugging behavioral variations based on past data of user session behavior, which means that they can continuously learn and optimize to more accurately capture and parse user behavior.

[0067] Finally, in step 104, the user data mining system makes a weather service push decision based on these weather content push preference views through a multi-decision algorithm for the weather user session behavior data. The multi-decision algorithm comprehensively considers multiple factors, including the user's historical behavior, current needs, and the platform's service capabilities, to ensure that the pushed content meets the user's personalized needs and can achieve efficient service push. Finally, the user data mining system generates a corresponding weather service push tag for each weather content push preference view, and the weather service push tag will be used to guide subsequent weather service push work.

[0068] Through the above steps, the user data mining system realizes the whole process from mining push preferences from user behavior data to generating personalized push decisions, which not only improves the pertinence and effectiveness of meteorological services, but also improves user satisfaction and service quality.

[0069] Compared with traditional technologies, the discrete structured behavior semantic embedding network can understand the semantics of user behavior more deeply. For example, when a user queries "Will it rain tomorrow?", the network can not only recognize that this is a query about the weather, but also capture the user's possible concern about travel plans or event arrangements. In addition, the discrete structured behavior semantic embedding network converts user behavior into structured data output, which makes subsequent data processing and analysis more convenient. Each user behavior is accurately embedded in a high-dimensional space, retaining the rich semantic information of the behavior. Furthermore, the discrete structured behavior semantic embedding network has strong dynamic adaptability and can automatically adjust according to the behavior patterns of different users, which means that as user behavior continues to change, the network can continuously learn and optimize its embedded representation.

[0070] Compared with traditional push systems, the discrete structured push preference parsing network can provide more refined user preference parsing. It not only considers the user's historical behavior, but also combines the user's real-time feedback and contextual information to provide users with more personalized push content. Furthermore, the discrete structured push preference parsing network can generate multi-dimensional user preference tags, such as weather type, temperature range, wind level, etc. These tags provide the push system with a richer user portrait, thereby improving the accuracy and satisfaction of the push. In addition, the discrete structured push preference parsing network has a real-time feedback mechanism, which can make timely adjustments based on user feedback on the push content. This dynamic optimization capability makes the push system more intelligent and efficient.

[0071] Compared with a single decision-making algorithm, a multi-decision-making algorithm can make decisions by comprehensively considering multiple factors. In the push of meteorological services, the algorithm not only considers the user's personal preferences, but also combines factors such as current weather conditions and geographical location to provide users with the most suitable push content. On the other hand, the various factors in the multi-decision algorithm are not simply added together, but the weights are dynamically adjusted according to the actual situation. For example, in severe weather conditions, the algorithm can focus more on pushing information related to weather warnings. In addition, the multi-decision algorithm also has the ability to continuously optimize. By continuously collecting user feedback and data, the algorithm can automatically adjust the decision-making strategy to provide better services. This self-learning and optimization ability makes the algorithm more adaptable to changing user needs and environmental conditions.

[0072] It is worth mentioning that, combined with the above content, specific algorithm examples of discrete structured behavior semantic embedding network, discrete structured push preference parsing network and multi-decision algorithm are given below.

[0073] (1) Specific algorithm example of discrete structured behavior semantic embedding network:

[0074] Innovation: Combining graph neural network (GNN) with Transformer model to deeply embed user session behavior; Data preprocessing: Convert user session behavior data into a graph structure, where nodes represent user behaviors and edges represent the associations between behaviors; Graph neural network embedding: Use graph neural network (such as Graph Convolutional Network, GCN) to embed user behavior graphs to capture the structured information of behaviors; Transformer model: Use the output of graph neural network as the input of Transformer model to further capture the timing and contextual information of behaviors; Output: Obtain a deep embedding vector for each user behavior, which combines the structured information and timing context of the behavior.

[0075] (2) Specific algorithm example of discrete structured push preference parsing network:

[0076] Innovation: The attention mechanism is combined with the memory network to perform a detailed analysis of push preferences; Input processing: The output of the discrete structured behavior semantic embedding network is used as input; Attention mechanism: The multi-head self-attention mechanism (Multi-HeadAttention) is used to weight the embedding vector of user behavior to highlight important behavioral features; Memory network: The memory network is introduced to store and recall users' long-term preferences. Each memory slot in the memory network can represent a user's preference or interest; Output: Combine the output of the attention mechanism and the memory network to obtain the user's push preference view.

[0077] (3) Specific algorithm examples of multivariate decision-making algorithms:

[0078] Innovation: Combining reinforcement learning with deep neural networks to optimize push decisions; State definition: Define the state space as a combination of information such as the user's current session behavior embedding, historical push feedback, etc.; Action definition: Define the action space as possible push content or service tags; Reward function: Define the reward function based on user feedback (such as click-through rate, dwell time, etc.); Deep neural network: Use deep neural networks (such as DQN, Deep Q-Network) to approximate the action value function (Q-function); Reinforcement learning training: Through interaction with the environment, use the ε-greedy strategy for exploration and utilization to continuously optimize push decisions; Output: Get the optimal push decision in each state.

[0079] Based on the above, the digital meteorological service in step 101 refers to the use of modern information technology, such as big data, cloud computing, the Internet of Things, etc., to provide refined, personalized and intelligent meteorological information services. This service can collect, process and analyze a large amount of meteorological data in real time, and customize and push relevant meteorological information according to user needs and preferences. Digital meteorological services not only provide basic weather forecasts, climate monitoring and other functions, but also provide users with personalized meteorological advice and warnings to help users better plan activities and travel.

[0080] Diverse customized push texts refer to push information texts in various formats and styles customized for users based on their personalized needs and preferences, as well as current weather conditions. These texts may include simple weather forecasts, weather warnings, clothing suggestions, travel tips, etc. They are pushed to users in various forms such as text messages, emails, and APP notifications. This type of push text is characterized by high personalization and customization, and is designed to provide weather information that is closely related to users' lives, helping users make more informed decisions.

[0081] Meteorological user session behavior data refers to all interactive data generated by users in the process of using digital meteorological services, including but not limited to users' query history, browsing history, click behavior, feedback, etc. By analyzing this data, we can gain a deep understanding of users' needs and preferences, thereby providing users with more accurate meteorological services. Meteorological user session behavior data is an important basis for the optimization and personalization of digital meteorological services.

[0082] In detail, in step 101, the main task of the user data mining system is to obtain the diversified customized push texts and meteorological user conversation behavior data of the digital meteorological service. First, the user data mining system is connected to the database or data warehouse of the digital meteorological service to extract the push texts that have been customized for users. These texts may include weather forecasts, meteorological disaster warnings, life index suggestions, etc. for different user groups, and the formats and styles are different to meet the needs and preferences of different users. At the same time, the user data mining system will also collect the conversation behavior data generated by users in the process of using digital meteorological services. These data record every interaction between users and the meteorological service platform, such as query history, browsing history, clicked push texts, feedback provided, etc. These data are crucial for understanding user needs and optimizing push strategies. After obtaining these data, the user data mining system performs preliminary data cleaning and preprocessing to ensure the accuracy and consistency of the data. For example, the user data mining system removes duplicate or invalid data records, processes missing values ​​or outliers, and converts the data into a format suitable for subsequent analysis. Through this step, the user data mining system provides a rich and accurate data foundation for subsequent user behavior semantic embedding mining, push preference analysis, and meteorological service push decision-making. These data will help the user data mining system to understand user needs more deeply, thereby providing more personalized and intelligent meteorological services.

[0083] Furthermore, the user demand identification indication in step 102 and step 103 refers to a signal or indicator used by the system to identify and understand the specific needs of the user. In the context of digital weather services, the user demand identification indication may come from the user's input query, browsing history, click behavior, etc., which together provide the system with the user's current and potential weather information service needs.

[0084] Behavioral semantic embedding mining is a data analysis technology that aims to extract deep semantic information from user behavior data. In the field of meteorological services, it can help the system understand the true intention of user query, browsing and feedback behaviors, and then optimize service push.

[0085] The personalized demand feedback elements reflect the personalized needs and feedback of users, which may include the type of weather information preferred by users, push time, information presentation method, etc. These elements are the key references for the system to customize push content.

[0086] Push preference decision knowledge features are extracted from user behavior data and are key information used to guide push decisions. They may include user interests, usage habits, and attention to specific weather information.

[0087] Push preference analysis aims to analyze users’ specific preferences in weather information push based on their behavioral data and feedback. These preferences will directly affect the customization of push content.

[0088] Weather content push preference opinions refer to the opinions formed by the system about the type and method of weather information that users want to receive based on their historical behaviors and current needs. These opinions are an important basis for the push system to customize content.

[0089] The past data of user session behavior records the user's past interactions with the system, including queries, browsing, clicks, feedback, etc. They are valuable resources for the system to optimize push strategies and learn user preferences.

[0090] Behavior variation debugging improves the system's push strategy and service quality by analyzing and debugging user behavior data. It can help the system better adapt to the changes and diversity of user behavior.

[0091] Based on step 102, the user data mining system first obtains the multi-state customized push texts, which are regarded as user demand identification indications. The user data mining system then uses the discrete structured behavior semantic embedding network to conduct an in-depth analysis of the user's conversation behavior data with the meteorological service. This network can capture the deep semantics of user behavior, so as to more accurately understand the user's needs and intentions. Through the behavioral semantic embedding mining of the meteorological user conversation behavior data, the user data mining system can extract the user's personalized demand feedback elements. These elements may include the user's attention to different types of meteorological information, usage habits, preferred information presentation methods, etc. Finally, the user data mining system generates a number of push preference decision knowledge features based on these personalized demand feedback elements. These features provide a key basis for subsequent push preference analysis and meteorological content customization. In step 103, the user data mining system continues to use the multi-state customized push text as a user demand identification indication. Then, the user data mining system conducts an in-depth analysis of the push preference decision knowledge features obtained in the previous step through the discrete structured push preference analysis network. The discrete structured push preference analysis network can analyze the user's specific preferences in meteorological information push based on the user's personalized needs and historical behavior data. Through this process, the user data mining system can form several weather content push preference opinions. It is worth noting that the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network are both obtained by debugging behavior variation based on past data of user session behavior. This means that the two networks can continuously learn and optimize to better adapt to the changes and diversity of user behavior, thereby providing more accurate and personalized weather information services.

[0092] In other possible application scenarios, the personalized demand feedback elements can be concretized as a series of numerical feature vectors, which describe the personalized needs of users for meteorological information. For example, the personalized demand feedback elements can be: [0.8, 0.3, 0.6, 0.9, 0.2]. Among them, the first value (0.8) represents the user's attention to weather forecasts. The closer the value is to 1, the more the user pays attention to weather forecast information; the second value (0.3) represents the user's demand for meteorological disaster warnings. The higher the value, the more the user needs timely meteorological disaster warning information; the third value (0.6) reflects the user's interest in air quality information. The larger the value, the more interested the user is in air quality information; the fourth value (0.9): represents the user's acceptance of clothing suggestions. The value close to 1 means that the user is very willing to accept clothing suggestions provided based on weather conditions; the fifth value (0.2) represents the user's interest in tourism meteorological information. The lower the value, the lower the user's interest in this type of information.

[0093] The push preference decision knowledge feature is extracted based on the user's behavior data and feedback, and is used to guide push decisions. For example, the push preference decision knowledge feature can be: [0.7, 0.5, 0.4, 0.8, 0.3]. Among them, the first value (0.7) represents the degree to which the user prefers to receive weather information in the morning. The higher the value, the more the user likes to receive weather push in the morning; the second value (0.5) reflects the user's preference for push content that combines pictures and texts. A moderate value indicates that the user has a certain preference for the push method that combines pictures and texts, but it is not the strongest choice; the third value (0.4): indicates the user's acceptance of weather push in the form of video. A relatively low value indicates that the user is generally interested in this form of push; the fourth value (0.8) represents the user's expectation for personalized push. A high value means that the user is very much looking forward to receiving personalized weather push; the fifth value (0.3) reflects the user's tolerance for advertising content. A low value means that the user does not want to see advertising content in weather push.

[0094] The weather content push preference view is a push strategy formed by the system based on the user's historical behavior and current needs. The following is a specific weather content push preference view: [0.9, 0.6, 0.2, 0.7, 0.5]. Among them, the first value (0.9) indicates that the system believes that the user is in great need of timely weather forecast push; the second value (0.6): indicates that the system judges that the user has a certain demand for the push of meteorological disaster warnings; the third value (0.2) reflects that the system believes that the user has a low demand for the push of meteorological science knowledge; the fourth value (0.7) indicates that the system believes that the user is more in need of the push of travel suggestions based on weather conditions; the fifth value (0.5) indicates that the system judges that the user has a certain interest in the push of healthy life suggestions provided according to weather conditions.

[0095] Based on the above content, the weather service push decision in step 104 refers to the process of finally determining what kind of weather service content to push to the user through a series of algorithms and analyses based on the user's behavior data, preferences and needs. This decision-making process may involve comprehensive consideration of multiple data, including but not limited to the user's historical behavior data, current weather conditions, and the user's personalized needs. The purpose of the weather service push decision is to ensure that the information pushed to the user is accurate, timely, and meets the user's needs. The weather service push tag is a tag used to identify and classify weather service push content. These tags are usually generated based on the user's behavior data, preferences, and needs, and are used to guide the system to push the most appropriate weather service content to the user. For example, if a user often queries the weather forecast of a certain city, the system can label the user as "pay attention to the weather in a certain city" so that information can be pushed to the user in a timely manner when the relevant weather changes.

[0096] In detail, the user data mining system begins to conduct in-depth analysis and decision-making on the meteorological user session behavior data through a multivariate decision-making algorithm based on the meteorological content push preference views obtained in the previous steps. First, the user data mining system reviews and organizes all the user's session behavior data during the use of meteorological services. These data include the user's query records, browsing history, click behavior, feedback, etc., which together constitute the overall picture of the user's behavior. Next, the user data mining system uses a multivariate decision-making algorithm, which is an algorithm that can comprehensively consider multiple factors and weigh the pros and cons to make the best decision. In this scenario, the algorithm will combine the user's meteorological content push preference views to analyze which meteorological service content best meets the user's current needs and long-term preferences. In the decision-making process, the algorithm will refer to multiple external factors such as weather conditions, seasonal changes, and the user's geographical location, as well as internal factors such as the user's personalized needs and historical behavior patterns. By combining these factors, the algorithm can customize the most appropriate meteorological service push content for the user. Finally, after processing by the multivariate decision-making algorithm, the user data mining system generates corresponding meteorological service push tags for each meteorological content push preference view. These tags not only reflect the user's personalized needs, but also provide clear guidance for subsequent meteorological service push. For example, if a user's preference views show a strong concern for air quality information, the user data mining system can generate a push tag of "concerned about air quality" to promptly notify the user when the air quality changes significantly.

[0097] The embodiment of the present application achieves accurate identification of user needs and effective push of personalized services by acquiring the diversified customized push texts and meteorological user conversation behavior data of digital meteorological services. The deep semantics of user conversation behavior are deeply mined by using the discrete structured behavioral semantic embedding network, which not only improves the recognition accuracy of user needs, but also provides a solid data foundation for personalized push. Furthermore, the push preference decision knowledge features are analyzed through the discrete structured push preference parsing network, and several meteorological content push preference viewpoints are obtained, which makes the push content more in line with the actual preferences and needs of users.

[0098] In addition, the embodiment of the present application also adopts a method of debugging behavior variation based on past data of user session behavior, continuously optimizing network performance, and improving push accuracy and user satisfaction. Finally, a push decision is made for each meteorological content push preference viewpoint through a multivariate decision-making algorithm, and a corresponding meteorological service push tag is generated. This not only simplifies the push process, but also greatly improves the personalization and intelligence level of the push service. Overall, the embodiment of the present application has significant advantages in improving the quality of digital meteorological service push, meeting user personalized needs, and improving user experience.

[0099] In some optional embodiments, the method of using the diverse customized push text as a user demand identification indication, performing behavioral semantic embedding mining on the meteorological user session behavior data through a discrete structured behavioral semantic embedding network, and configuring personalized demand feedback elements for the output of the behavioral semantic embedding mining to obtain a number of push preference decision knowledge features of the meteorological user session behavior data includes: using the diverse customized push text as a user demand identification indication, performing multi-level behavioral semantic extraction on the meteorological user session behavior data through a discrete structured behavioral semantic embedding network to obtain multi-level user session behavior semantics; retrieving associated session behavior semantics of the multi-level user session behavior semantics in a push mapping feature coordinate system through preset matching rules; configuring personalized demand feedback elements for the associated session behavior semantics to obtain a number of session behavior push associated semantics; and generating push preference decision knowledge features corresponding to each session behavior push associated semantics.

[0100] Based on this embodiment, the user data mining system performs a series of refined operations to achieve in-depth mining of meteorological user session behavior data and personalized demand feedback:

[0101] First, the user data mining system uses the diversified customized push texts as the key indicator for identifying user needs. These push texts may be customized based on the user's historical browsing history, search behavior or other relevant data, aiming to accurately capture the user's interests and needs.

[0102] Next, the user data mining system uses the discrete structured behavior semantics embedding network to perform in-depth behavior semantic extraction on the meteorological user session behavior data. This network structure can effectively identify and extract the multi-level behavior semantics displayed by users in the session, such as the keywords used by users when querying the weather, the types of meteorological information browsed, etc., thereby obtaining multi-level user session behavior semantics. For example, when a user queries "Beijing tomorrow's weather", the user data mining system extracts key semantics such as "Beijing", "tomorrow" and "weather".

[0103] After extracting the multi-level user conversation behavior semantics, the user data mining system uses the preset matching rules to retrieve the conversation behavior semantics associated with these behavior semantics in a push mapping feature coordinate system. This push mapping feature coordinate system can be regarded as a multi-dimensional space, in which each point or area represents one or more specific user behaviors or interests. Continuing with the above example, the user data mining system can find other semantics associated with "Beijing tomorrow's weather" in the coordinate system, such as "air quality", "clothing suggestions", etc. Subsequently, the user data mining system configures personalized demand feedback elements for these associated conversation behavior semantics. These elements may include user preferences, historical behaviors, geographic locations, etc., to ensure the personalization and accuracy of the pushed content. For example, for users who often check the weather in Beijing, the user data mining system can push personalized information such as "According to historical data, the temperature in Beijing will drop tomorrow, and it is recommended to wear a warm coat."

[0104] Finally, the user data mining system generates corresponding push preference decision knowledge features for each session behavior push association semantics. These features are derived based on the user's personalized needs and historical behavior data, and they will serve as the basis for subsequent push decisions.

[0105] With this design, by deeply mining the multi-level behavior semantics in the user session behavior data and combining the personalized demand feedback elements, the user data mining system can understand user needs more accurately and provide more personalized services. This not only improves user satisfaction, but also enhances the practicality and intelligence level of the user data mining system. At the same time, by generating the knowledge features of push preference decisions, the user data mining system can more effectively push subsequent content, thereby improving user experience and service quality.

[0106] In other optional embodiments, configuring personalized demand feedback elements for the associated session behavior semantics to obtain a number of session behavior push associated semantics includes: downsampling the associated session behavior semantics through a dynamic semantic downsampling algorithm to obtain page operation downsampled behavior semantics; performing deep residual processing on the page operation downsampled behavior semantics, and obtaining feature influence coefficients of different attention dimensions in the page operation downsampled behavior semantics based on the output of the deep residual processing; configuring demand feedback definition features for the feature influence coefficients of the different attention dimensions based on a discrete thermal analysis strategy to obtain semantic definition coefficients of different attention dimensions; and obtaining a number of session behavior push associated semantics based on the semantic definition coefficients of the different attention dimensions and the page operation downsampled behavior semantics.

[0107] In the application scenario of this embodiment, the user data mining system performs a series of fine processing on the associated session behavior semantics to configure personalized demand feedback elements, and finally obtains a number of session behavior push associated semantics.

[0108] First, the user data mining system downsamples the semantics of the associated session behavior through a dynamic semantic downsampling algorithm. Downsampling is a process of reducing the amount of data while retaining as much useful information as possible in the original data. Here, the dynamic semantic downsampling algorithm can dynamically adjust the sampling frequency and method according to the characteristics and importance of the session behavior, thereby obtaining the semantics of the downsampled behavior of the page operation. For example, for a weather forecast page that users frequently click on, the algorithm will increase the sampling frequency of the page operation behavior to more accurately capture the user's behavior patterns and preferences.

[0109] Next, the user data mining system performs deep residual processing on the semantics of the downsampled behavior of page operations. Deep residual processing is an effective deep learning technology that can solve the problems of gradient vanishing and model degradation in the training process of deep neural networks. Here, through deep residual processing, the user data mining system is able to extract the feature influence coefficients of different attention dimensions in the semantics of the downsampled behavior of page operations. These feature influence coefficients reflect the degree of attention users pay to different page elements. For example, for a weather forecast page, users may pay more attention to temperature information and pay less attention to wind information. This difference is reflected in the feature influence coefficient.

[0110] Then, the user data mining system configures the demand feedback definition feature based on the feature influence coefficients of different attention dimensions based on the discrete thermal analysis strategy. Discrete thermal analysis is a method of analyzing data features based on data distribution and thermal conditions. Here, through the discrete thermal analysis strategy, the user data mining system can determine the user's real needs and preferences for each page element, and obtain the semantic definition coefficients of different attention dimensions based on this.

[0111] Finally, the user data mining system obtains several session behavior push-related semantics based on the semantic definition coefficients of different attention dimensions and the page operation downsampling behavior semantics. These related semantics not only take into account the user's behavior patterns, but also combine the user's personalized needs, so they can more accurately guide subsequent push decisions.

[0112] In this way, through technical means such as dynamic semantic downsampling algorithm and deep residual processing, the user data mining system can have a deeper understanding of the user's page operation behavior and attention allocation; at the same time, combined with the demand feedback definition characteristics obtained by the discrete thermal analysis strategy, the push content is more in line with the user's real needs and preferences; the final generated session behavior push association semantics provides strong support for subsequent precise push, improving user experience and service quality.

[0113] It is worth mentioning that the discrete thermal analysis strategy and the semantic definition coefficient play a core role in the above embodiments. The following is a detailed introduction to these two parts, and how to configure the demand feedback definition features for the feature influence coefficients of different attention dimensions through the discrete thermal analysis strategy, so as to obtain the semantic definition coefficients of different attention dimensions.

[0114] In detail, the discrete thermal analysis strategy is a data analysis method that visualizes data points according to their importance or frequency based on the principle of heat maps. In this strategy, the "heat" of a data point is usually proportional to its frequency or importance. Through discrete thermal analysis, you can intuitively see the distribution of data and find out the hot and cold spots in the data. In the user data mining system, the discrete thermal analysis strategy is used to analyze the semantics of users' page operation behavior, especially the users' attention to different page elements. By analyzing the data such as users' clicks, browsing, and dwell time on page elements, the system can generate a heat map to intuitively show the users' attention to each element.

[0115] The semantic definition coefficient is a quantitative indicator used to measure the user's attention, interest or understanding of a page element or a topic. In the user data mining system, the semantic definition coefficient is calculated based on the discrete thermal analysis strategy, which reflects the user's real needs and preferences for different page elements.

[0116] Based on this, the following introduces how to configure demand feedback definition features through a discrete thermal analysis strategy through an independently implementable technical solution.

[0117] (1) Data collection and processing: First, the user data mining system collects users’ page operation data, including clicks, browsing, dwell time, etc. These data are preprocessed and converted into a format suitable for thermal analysis.

[0118] (2) Generate a heat map: Using the discrete thermal analysis strategy, the user data mining system generates a heat map based on the collected data. In this heat map, the color depth of different areas represents the user's attention to the elements in that area.

[0119] (3) Determine the feature influence coefficient: By analyzing the heat map, the user data mining system can calculate the feature influence coefficients of different attention dimensions. These coefficients reflect the user's attention to each page element.

[0120] (4) Configure demand feedback defining features: Based on the feature influence coefficient, the user data mining system configures demand feedback defining features for each page element. These features define the user’s real needs and preferences for the element.

[0121] (5) Calculate semantic definition coefficients: Finally, the user data mining system calculates semantic definition coefficients for different attention dimensions based on the definition features of demand feedback and the user's operation data. These coefficients will be used for subsequent push decisions.

[0122] Through the above steps, the discrete thermal analysis strategy can help the user data mining system to more accurately understand the needs and preferences of users, thereby achieving more accurate push services. This strategy not only improves the user experience, but also provides strong support for the company's precision marketing.

[0123] In some other independent embodiments, the semantic definition coefficients of the different attention dimensions and the page operation downsampling behavior semantics are used to obtain a plurality of session behavior push associated semantics, including:

[0124] Constructing a multi-dimensional weight matrix; wherein each dimension in the multi-dimensional weight matrix corresponds to a target element in the semantics of the page operation downsampling behavior, and the value in the multi-dimensional weight matrix is ​​determined by the semantic definition coefficient of the corresponding target element;

[0125] Normalizing the multi-dimensional weight matrix so that the sum of the weights of each dimension is 1, so as to achieve the distribution of the influence of each dimension;

[0126] The normalized multi-dimensional weight matrix is ​​used to perform weighted processing on the semantics of the downsampled behavior of page operations: the semantic definition coefficient of each page element is used as a weight and multiplied by the corresponding downsampled behavior semantics of the page operation to obtain the weighted semantics of the page operation behavior;

[0127] Based on the weighted semantics of page operation behaviors, the weighted semantics of page operation behaviors are decomposed into several associated groups through clustering algorithm, each of which represents a user interest or demand;

[0128] Extract representative push association semantics from each association group. The representative push association semantics not only represents the user's page operation behavior, but also reflects the user's attention and preference for different page elements.

[0129] The representative push-related semantics are used as session behavior push-related semantics for personalized and accurate content push.

[0130] In detail, the user data mining system first constructs a multi-dimensional weight matrix. Each dimension of this matrix corresponds to a target element in the semantics of the page operation downsampled behavior, such as a button, a block or a function point on the page. The value in the matrix, that is, the weight of each dimension, is determined by the semantic definition coefficient of the corresponding target element. This coefficient reflects the user's attention and importance to the page element. Next, the user data mining system normalizes this multi-dimensional weight matrix. The purpose of normalization is to make the sum of the weights of each dimension equal to 1, so that in subsequent calculations, the influence of each dimension can be reasonably distributed to avoid the influence of a certain dimension being too large or too small. Then, the user data mining system uses the normalized multi-dimensional weight matrix to weight the semantics of the page operation downsampled behavior. Exemplarily, the semantic definition coefficient of each page element is used as a weight and multiplied by the corresponding semantics of the page operation downsampled behavior. In this way, the weighted semantics of the page operation behavior obtained not only takes into account the user's operation behavior, but also reflects the user's attention to different page elements. Next, based on the weighted semantics of page operation behaviors, the user data mining system uses a clustering algorithm to decompose these weighted semantics of page operation behaviors into several association groups. Each association group represents a user interest or need. For example, users who often click on weather forecasts may be classified into a group that pays attention to weather, while users who often browse travel information may be classified into another group that pays attention to travel. After that, the user data mining system extracts representative push association semantics from each association group. These representative push association semantics not only characterize the user's page operation behavior, but also reflect the user's attention and preference for different page elements. For example, in the group that pays attention to weather, the representative push association semantics may be "today's weather", "weather forecast for the next week", etc. Finally, the user data mining system uses these representative push association semantics as session behavior push association semantics for subsequent personalized and accurate content push. In this way, when the user visits the page again, the user data mining system can push the most relevant content to the user based on the user's interests and needs.

[0131] In this way, by constructing a multi-dimensional weight matrix and normalizing it, the user data mining system can more accurately measure the user's attention and importance to different page elements; by using weighted processing and clustering algorithms, the user data mining system can more effectively decompose the user's page operation behavior semantics into related groups, and extract representative push-related semantics; ultimately, these representative push-related semantics provide strong support for subsequent personalized and accurate content push, improving user experience and service quality.

[0132] Under some preferred design ideas, the diverse customized push text is used as a user demand identification indication, and the push preference decision knowledge features are pushed preference parsed through a discrete structured push preference parsing network to obtain a number of meteorological content push preference viewpoints, including: feature projection of a number of the push preference decision knowledge features to obtain a number of push preference viewpoint scene projection features; the diverse customized push text is used as a user demand identification indication, and a preference viewpoint identification operation is performed on each of the push preference viewpoint scene projection features through a discrete structured push preference parsing network to determine the meteorological content push preference viewpoint corresponding to each push preference viewpoint scene projection feature.

[0133] In this design idea, the user data mining system uses the diverse customized push texts as an indication of user demand identification, and performs push preference analysis on the push preference decision knowledge features through a discrete structured push preference analysis network to obtain several meteorological content push preference viewpoints.

[0134] First, the user data mining system will perform feature projection on several push preference decision knowledge features. Feature projection is a data dimensionality reduction technology that can map high-dimensional data to low-dimensional space while retaining the main features of the data. Here, the purpose of feature projection is to obtain several push preference viewpoint scenario projection features. These projection features can more intuitively reflect the user's push preferences and facilitate subsequent preference viewpoint identification operations.

[0135] Next, the user data mining system will use the diverse customized push texts as user demand identification indicators. These push texts may be the user's historical browsing records, search keywords, likes or comments, etc., which can reflect the user's needs and interests in meteorological content. The user data mining system inputs these push texts into the discrete structured push preference parsing network as an important basis for the network to perform preference parsing.

[0136] Then, the user data mining system uses the discrete structured push preference parsing network to perform a preference view identification operation on each push preference view scene projection feature. This network may be built by a deep learning algorithm with strong feature extraction and classification capabilities. Through the processing of the network, the user data mining system can determine the meteorological content push preference view corresponding to each push preference view scene projection feature. For example, some projection features may indicate that users are more concerned about the accuracy of weather forecasts, while other projection features may reflect users' needs for meteorological disaster warnings.

[0137] Finally, the user data mining system organizes and outputs these weather content push preference opinions for reference in the subsequent content push strategy formulation. These preference opinions not only help the user data mining system to push weather content that users are interested in more accurately, but also help the user data mining system to continuously optimize push strategies and improve user experience.

[0138] Through the above steps, the user data mining system can effectively transform the user's push preferences into specific weather content push preferences, thereby achieving personalized and accurate content push. The beneficial effects of this technical solution are: improving the accuracy and pertinence of content push, meeting the personalized needs of users; performing preference analysis through a discrete structured push preference analysis network, enhancing the intelligence and automation of the user data mining system; and at the same time, providing strong data support for the subsequent optimization and upgrade of the user data mining system.

[0139] It is worth mentioning that, in some independent embodiments, the feature projection of the plurality of push preference decision knowledge features to obtain a plurality of push preference viewpoint scene projection features includes:

[0140] Constructing a feature projection matrix, wherein the dimension of the feature projection matrix matches the number of push preference decision knowledge features and the type of push preference decision knowledge features;

[0141] Preprocessing each push preference decision knowledge feature; the preprocessing includes standardization, normalization or denoising operations to achieve scale consistency and range consistency of each push preference decision knowledge feature;

[0142] Using the feature projection matrix, the preprocessed push preference decision knowledge features are projected into the new feature space to obtain the push preference viewpoint scene projection features; wherein, in the projection process, the projection algorithm used includes at least one of principal component analysis (PCA), linear discriminant analysis (LDA), and local preservation projection (LPP), and the projection algorithm used is used to retain important information in the original features and remove redundant information;

[0143] Post-processing the obtained push preference viewpoint scene projection features; the post-processing includes feature selection, dimensionality reduction or clustering operations;

[0144] According to the result of the post-processing, a plurality of push preference viewpoint scene projection features are output.

[0145] In detail, the user data mining system first constructs a feature projection matrix. The dimension of this matrix matches the number and type of push preference decision knowledge features to ensure that all feature information can be fully mapped. Next, the user data mining system preprocesses each push preference decision knowledge feature. The preprocessing process may include standardization, normalization, or denoising. The purpose of standardization is to adjust the scale of the feature to a uniform range, while normalization is to ensure that the value range of each feature is consistent, and denoising is to eliminate outliers or noise in the data, thereby improving the accuracy of subsequent projections. Then, using the constructed feature projection matrix, the user data mining system projects the preprocessed push preference decision knowledge features into a new feature space. In this process, the user data mining system uses some projection algorithms, such as principal component analysis (PCA), linear discriminant analysis (LDA), local preservation projection (LPP), etc. The purpose of these algorithms is to retain the important information in the original features while removing redundant information, thereby obtaining a more compact and effective feature representation. After the projection is completed, the user data mining system will post-process the obtained push preference viewpoint scene projection features. Post-processing may include operations such as feature selection, dimensionality reduction, or clustering. The purpose of feature selection is to select the most representative features, dimensionality reduction is to further compress the dimensions of the features for easier processing and analysis, and clustering is to group similar features into one category for subsequent identification and analysis. Finally, based on the results of post-processing, the user data mining system will output several push preference, opinion, and scenario projection features. These features not only have lower dimensions and more compact information, but also better reflect the user's push preferences and opinion scenarios, providing strong data support for subsequent personalized push.

[0146] In this way, by constructing a feature projection matrix and adopting a suitable projection algorithm, the user data mining system can more effectively extract and represent the user's push preferences; the preprocessing and post-processing operations further improve the accuracy and availability of the features; the final output push preference viewpoint scene projection features provide an accurate data basis for subsequent personalized push, thereby improving user experience and service quality.

[0147] In some other possible technical solutions, the method also includes: obtaining several data category labels of past user session behavior data; based on the data category labels, performing behavior variation debugging on the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged to obtain the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network.

[0148] Based on this technical solution, first, the user data mining system will obtain several data category labels of the user's past session behavior data. These data category labels may be extracted based on multi-dimensional information such as the user's browsing history, search history, purchase behavior, etc., and are used to identify and classify different user behavior patterns. For example, one data category label may represent the user's preference for travel-related weather content, and another may represent the user's attention to science and technology-related weather information.

[0149] Next, the user data mining system will perform behavioral variation debugging on the discrete structured behavioral semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged based on these data category labels. "Behavior variation debugging" here is a technical term that refers to adjusting the parameters and structure of the network so that it can better adapt to and parse user data with different behavioral patterns.

[0150] During this debugging process, the user data mining system will input the past user session behavior data with data category labels into the network to be debugged, and observe the difference between the network output and the real label. Then, the user data mining system will adjust the network's weights, biases and other parameters based on these differences, and may even change the network's structure to improve the network's ability to analyze user behavior.

[0151] After multiple iterations and optimizations, the user data mining system will obtain two debugged networks: the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network. These two networks have been specially trained and optimized for user behavior patterns, so they can more accurately understand and parse user behavior semantics and push preferences.

[0152] In this way, by obtaining the data category labels of the past data of user session behavior and debugging the network for behavioral variation based on these labels, the user data mining system can more accurately grasp the user's behavior patterns and push preferences. This not only improves the accuracy and personalization of the content pushed by the user data mining system, but also enhances the user experience and satisfaction. At the same time, the debugged network has stronger generalization ability and can better adapt to the needs and behavioral changes of different user groups.

[0153] In the following technical scheme, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are debugged for behavior variation based on the data category labels to obtain the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network, including: configuring subscription interest point features of different levels for each of the data category labels in turn to obtain subscription interest description texts of each data category label under the influence of subscription interest point features of different levels; based on the subscription interest description texts of each data category label under the influence of subscription interest point features of different levels, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are derive behavior descriptions in turn to obtain subscription interest error variables, behavior description derived error variables and confidence coefficient thermal error variables corresponding to subscription interest point features of different levels; based on the subscription interest error variables, the behavior description derived error variables and the confidence coefficient thermal error variables, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are trained and debugged to obtain discrete structured behavior semantic embedding network and discrete structured push preference parsing network.

[0154] Combined with this technical solution, the user data mining system will perform refined behavioral variation debugging on the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged based on the acquired data category labels. This process is designed to ensure that the network can better understand and respond to the meteorological service needs of different users.

[0155] First, the user data mining system will configure different levels of subscription interest point features for each data category label in turn. These subscription interest point features may include the user's attention to different types of weather information, the user's preferred weather data display method, etc. By configuring these features, the user data mining system can generate subscription interest description texts for each data category label under the influence of different levels of subscription interest point features. These texts describe in detail the user's needs and preferences for weather information in different situations.

[0156] Next, the user data mining system will derive behavior descriptions for the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged based on these subscription interest description texts. The purpose of this step is to evaluate the accuracy of the network in understanding user behavior semantics and push preferences. Through behavior description derivation, the user data mining system can obtain subscription interest error variables, behavior description derivative error variables, and confidence coefficient thermal error variables corresponding to the features of subscription interest points at different levels.

[0157] Among them, the subscription interest error variable reflects the deviation of the network's understanding of the user's subscription interest; the behavior description derivative error variable reflects the error of the network when generating description text based on user behavior; and the confidence coefficient thermal error variable represents the network's confidence error in the features of subscription interest points at different levels. These error variables provide a clear direction for debugging the network for the user data mining system.

[0158] Finally, based on these error variables, the user data mining system will train and debug the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged. By continuously adjusting the parameters and structure of the network, the user data mining system aims to minimize these error variables, thereby improving the accuracy of the network's understanding of user behavior and push preferences.

[0159] After this series of steps, the user data mining system will eventually obtain a finely tuned discrete structured behavior semantic embedding network and a discrete structured push preference parsing network. These two networks will be able to more accurately parse users' weather service needs and provide users with more personalized and accurate weather content push.

[0160] In this way, through refined behavior variation debugging, the user data mining system can more deeply understand the user's weather service needs and push preferences. This will help improve the personalization and accuracy of weather services, thereby enhancing user experience and meeting the diverse needs of users. At the same time, the debugged network will have stronger adaptability and generalization capabilities, and can better serve the vast user base.

[0161] In some other preferred embodiments, according to the subscription interest description text of each data category label under the influence of the subscription interest point features at different levels, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are sequentially subjected to behavior description derivation to obtain the subscription interest error variables corresponding to the subscription interest point features at different levels, including: combining the core subscription interest description text of each data category label with the past data of the user session behavior to obtain the session subscription description combination text of each data category label, the core subscription interest description text being the subscription interest description text configured with the highest level subscription interest point features; and combining the session subscription description combination text with the discrete structured behavior semantic embedding network to be debugged. This paper conducts behavioral semantic embedding mining, and configures personalized demand feedback elements for the output of behavioral semantic embedding mining, so as to obtain the push preference decision knowledge features corresponding to each session subscription description combination text; performs preference viewpoint recognition operation on the push preference decision knowledge features through the discrete structured push preference parsing network to be debugged, so as to obtain the identified key subscription interest point features; based on the core subscription interest description text and the key subscription interest point features of each data category label, generates a subscription interest description prediction text set for each data category label; based on the subscription interest description prediction text set for each data category label and the subscription interest description prior text set for each data category label, determines the subscription interest error variable of the key subscription interest point features.

[0162] Based on this embodiment, first, the user data mining system generates core subscription interest description texts for each data category tag. These texts are generated based on the subscription interests of the user configured with the highest-level subscription interest point features, and they represent the user's most core and most concerned weather information needs. Then, the user data mining system combines these core subscription interest description texts with the user's past session behavior data to form session subscription description combination texts for each data category tag.

[0163] Subsequently, the user data mining system will use the discrete structured behavioral semantic embedding network to be debugged to conduct in-depth mining of these session subscription description combination texts. This process aims to extract the behavioral semantic information in the text and configure personalized demand feedback elements for it, so as to obtain the push preference decision knowledge features corresponding to each session subscription description combination text. These features not only reflect the user's meteorological information needs, but also reflect the user's personalized preferences.

[0164] Then, the user data mining system will further perform preference viewpoint recognition operations on these push preference decision knowledge features through the discrete structured push preference parsing network to be debugged. The purpose of this step is to identify key subscription interest point features, that is, those features that have a significant impact on user push preferences.

[0165] After identifying the key subscription interest features, the user data mining system combines the core subscription interest description text of each data category label to generate a set of subscription interest description prediction texts. These prediction text sets are based on the user's current and past behaviors and preferences to infer possible future subscription interests.

[0166] Finally, in order to evaluate the accuracy of the prediction, the user data mining system will compare the subscription interest description prior text set of each data category label (i.e., the subscription interest that the user has clearly expressed before) with the subscription interest description prediction text set. By comparing the difference between the two, the user data mining system can determine the subscription interest error variable of the key subscription interest point features. This error variable actually reflects the deviation of the user data mining system in understanding and predicting user subscription interests, providing an important basis for subsequent network debugging and optimization.

[0167] This design enables the user data mining system to more accurately capture and analyze users' weather information needs and push preferences, thereby providing more accurate and personalized weather services. This not only improves the user experience, but also enhances the user data mining system's responsiveness and adaptability to user behavior. At the same time, through continuous iteration and optimization, the performance of the user data mining system will be continuously improved to better meet users' diverse weather service needs.

[0168] It is worth mentioning that the exemplary calculation formula of the subscription interest error variable can be:

[0169] Subscription interest error variable = (q1*text similarity error) + (q2*push preference error) + (q3*user feedback error)

[0170] in:

[0171] The text similarity error is obtained by calculating the text similarity between the predicted text set and the prior text set. Specifically, the similarity of each text in the predicted text set can be compared with the corresponding text in the prior text set, and then the average value is taken. Text similarity can be calculated using methods such as cosine similarity and Jaccard similarity. Finally, the text similarity error is obtained by subtracting the average similarity from 1;

[0172] The push preference error is calculated by comparing the difference between the system's predicted push preference and the user's actual push preference. The absolute difference between the predicted value and the actual value of each push preference can be calculated and averaged to get the push preference error;

[0173] User feedback error is measured by user satisfaction with the pushed content. You can collect user satisfaction scores for each pushed content, and then calculate the average satisfaction. Finally, subtract the average satisfaction from 1 to get the user feedback error.

[0174] q1, q2, and q3 are weight coefficients used to balance the importance of different error terms, and their sum is equal to 1. These weight coefficients can be adjusted according to specific application scenarios and requirements.

[0175] Under the following design idea, the push preference decision knowledge features are subjected to a preference viewpoint recognition operation through the discrete structured push preference parsing network to be debugged to determine the identified key subscription interest point features, including: retelling the user session behavior past data through the discrete structured push preference parsing network to be debugged to obtain the user session behavior retelling log; mapping the push preference decision knowledge features to the session behavior retelling feature coordinate system of the user session behavior retelling log to obtain the identified key subscription interest point features.

[0176] In the meteorological service application scenario, the user data mining system can accurately capture users' meteorological information needs and push preferences through deep learning and data analysis technology, thereby providing users with more personalized and accurate meteorological services.

[0177] First, the user data mining system will use the discrete structured push preference parsing network to be debugged to process the past data of user session behaviors. This step is called "session behavior log retelling", and its purpose is to retell and record the user's session behaviors in the form of logs. These session behaviors include but are not limited to the types of weather information browsed by users, keywords searched, push content clicked, etc. By retelling these behaviors, the system can have a more comprehensive understanding of the user's information needs and preferences.

[0178] After completing the session behavior log retelling, the user data mining system will obtain a user session behavior retelling log. This log contains all the user's weather information related behaviors in the past period of time, which is an important basis for subsequent analysis.

[0179] Next, the system will map the push preference decision knowledge features to the conversation behavior retelling feature coordinate system of the user's conversation behavior retelling log. In other words, this step is to compare and map the user's push preference features previously mined by the system with the user's actual behavior data. Through this process, the system can more accurately identify the user's key subscription interest point features. These key subscription interest point features are the user's core needs and preferences for meteorological information. They may include weather in a specific area that the user is concerned about, warning information about a certain meteorological phenomenon (such as typhoons, heavy rains, etc.), or temperature changes in a specific time period. By accurately identifying these features, the user data mining system can provide users with more demand-oriented meteorological services.

[0180] For example, user A frequently browsed information about "Beijing weather" and "smog warning" in the past week. Through the processing of the user data mining system, the system will identify "Beijing weather" and "smog warning" as key subscription interest features of user A. When the system detects new Beijing weather forecast or smog warning information, it will be pushed to user A first, thus realizing personalized and accurate meteorological services.

[0181] In this way, the user data mining system is an important technical means in the meteorological service application scenario to identify the preference viewpoints through the discrete structured push preference parsing network to be debugged and determine the key subscription interest point features. This can not only improve the user experience and meet the diverse meteorological service needs of users, but also continuously improve the performance and responsiveness of the system through continuous iteration and optimization.

[0182] In other embodiments, the method further includes: comparing the past data of the user session behavior and the retelling log of the user session behavior to obtain a behavior description derived error variable corresponding to key subscription interest point features; obtaining the feature influence coefficient and semantic definition coefficient of the push preference decision knowledge feature in different attention dimensions to obtain a confidence coefficient thermal error variable corresponding to the key subscription interest point features.

[0183] In the meteorological service application scenario, the user data mining system further deepens its data processing and analysis capabilities to ensure that the user's meteorological information needs are met more accurately. In this process, the system not only pays attention to the user's current behavior, but also deeply mines the differences between the user's historical data and the session behavior retelling log, as well as the performance of the push preference decision knowledge characteristics in different attention dimensions.

[0184] First, the user data mining system compares the past data of user session behavior with the log of user session behavior retelling. The purpose of this step is to find the difference between the two, and then obtain the error variable derived from the behavior description corresponding to the key subscription interest point features. In other words, this error variable reflects the accuracy of the system's retelling of the user behavior log. If the error is small, it means that the system has a relatively accurate grasp of the user's behavior; otherwise, it means that the system may need further optimization to improve its ability to interpret user behavior.

[0185] For example, user B frequently checked the "Shanghai Weather Forecast" in the past month, but this information was not fully reflected in the system's user session behavior retelling log. Then, when comparing past data and retelling logs, the system will find this difference and use it as part of the behavior description derived error variable.

[0186] Next, the user data mining system will obtain the feature influence coefficients and semantic definition coefficients of the push preference decision knowledge features in different attention dimensions. These coefficients reflect the importance and clarity of different features in the user's push preference decision. By comprehensively analyzing these coefficients, the system can obtain the confidence coefficient thermal error variable corresponding to the key subscription interest point features. This error variable actually measures the system's confidence in the user's key subscription interest point features. If the confidence coefficient thermal error variable is small, it means that the system has high confidence in the user's key subscription interest points; otherwise, it means that the system may need more data to enhance its understanding of the user's interest points.

[0187] Taking user C as an example, the system finds that user C has a high degree of attention to "Guangzhou Heavy Rain Warning". When obtaining the different attention dimension coefficients of the push preference decision knowledge features, the system finds that the influence coefficient and semantic definition coefficient of the feature "Guangzhou Heavy Rain Warning" are both very high, indicating that this is a very critical and clear subscription interest point. However, if user C does not respond positively to this part of the information in the actual push, the system will re-evaluate the importance of this interest point by calculating the confidence coefficient thermal error variable.

[0188] In summary, by comparing the past data of user session behavior and the retelling log of user session behavior, and obtaining the coefficients of the push preference decision knowledge features in different attention dimensions, the user data mining system can more comprehensively understand the user's meteorological information needs and push preferences. This not only helps the system optimize its push strategy and improve user experience, but also ensures that meteorological services are more accurate and personalized. It can be seen that the user data mining system can more accurately capture and analyze the user's meteorological information needs and preferences, thereby providing more accurate and personalized meteorological services. This not only improves the user experience and enhances the system's responsiveness and adaptability to user behavior, but also through continuous iteration and optimization, the system's performance is continuously improved, better meeting the user's diverse meteorological service needs.

[0189] It is worth mentioning that in the meteorological service application scenario, in order to quantify the accuracy of the user data mining system in processing user session behavior data and pushing preference decision knowledge features, the behavior description derived error variable (behavior description derived training loss) and the confidence coefficient thermal error variable (confidence coefficient thermal training loss) are introduced. The following are exemplary calculation formulas for these two error variables.

[0190] 1) Behavior description derived error variable (behavior description derived training loss)

[0191] The behavior description derived error variable is used to measure the difference between the user session behavior log replayed by the system and the user's actual session behavior past data. The example formula of the behavior description derived error variable is:

[0192] [text{Behavior description derived error variable}=

[0193] frac{1}{N}sum_{i=1}^{N}|y_i-hat{y}_i|];

[0194] in:

[0195] (N) is the total number of user session behavior data;

[0196] (y_i) is the descriptive value of the (i)th behavior in the past data of the user's session behavior (e.g., the number of times the user viewed a certain type of weather information, the frequency of search, etc.);

[0197] (hat{y}_i) is the retelling value corresponding to the (i)th behavior in the user's session behavior retelling log;

[0198] The above formula calculates the mean absolute error between the user's actual behavior and the system's retelling of the behavior, which is used to measure the accuracy of the system's description of the user's behavior.

[0199] 2) Confidence coefficient thermal error variable (confidence coefficient thermal training loss)

[0200] The confidence coefficient thermal error variable reflects the system's confidence in the characteristics of the user's key subscription interest points. The example formula of the confidence coefficient thermal error variable is:

[0201] [text{Confidence coefficient thermal error variable}=frac{1}{M}sum_{j=1}^{M}(1-text{Conf}_j)times|text{Feat}_j-hat{text{Feat}}_j|];

[0202] in:

[0203] (M) is the total number of key subscription POI features;

[0204] (text{Conf}_j) is the confidence coefficient of the system for the feature of the (j)th key subscription point of interest, ranging from 0 to 1;

[0205] (text{Feat}_j) is the user's actual key subscription interest point feature value (for example, the user's attention and preference for a certain type of weather information, etc.);

[0206] (hat{text{Feat}}_j) is the feature value of the (j)th key subscription interest point predicted by the system.

[0207] This formula comprehensively considers the system's confidence and prediction accuracy for the features of each key subscription interest point, and is used to measure the system's ability to grasp the user's key subscription interest points.

[0208] Through the calculation of the above error variables, the performance of the user data mining system can be quantitatively evaluated, and the parts with larger errors can be optimized, thereby improving the system's responsiveness and service quality to users' meteorological service needs.

[0209] In an alternative embodiment, the training and debugging of the to-be-debugged discrete structured behavior semantic embedding network and the to-be-debugged discrete structured push preference parsing network based on the subscription interest error variable, the behavior description derived error variable and the confidence coefficient thermal error variable to obtain the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network includes: based on the subscription interest error variable, the behavior description derived error variable and the confidence coefficient thermal error variable, the training and debugging of the to-be-debugged discrete structured behavior semantic embedding network and the to-be-debugged discrete structured push preference parsing network to obtain the to-be-verified discrete structured behavior semantic embedding network and the to-be-verified discrete structured push preference parsing network; obtaining user session behavior verification data and the prior push label of the user session behavior verification data; and converting the user session behavior verification data into a discrete structured behavior semantic embedding network. Data is input into the discrete structured behavior semantic embedding network to be verified and the discrete structured push preference parsing network to be verified to obtain a meteorological service push verification label; consistency judgment is performed on the meteorological service push verification label and the prior push label to obtain a consistency judgment weight, as well as a first evaluation index and a second evaluation index of the probability model of the meteorological service push verification label and the prior push label; a discrete evaluation index is determined based on the first evaluation index and the second evaluation index; a global network quality evaluation value is obtained based on the consistency judgment weight and the discrete evaluation index, and when the global network quality evaluation value is higher than the set evaluation value, the discrete structured behavior semantic embedding network to be verified is used as a discrete structured behavior semantic embedding network, and the discrete structured push preference parsing network to be verified is used as a discrete structured push preference parsing network.

[0210] In the meteorological service application scenario, the user data mining system trains and optimizes its internal network model through a series of complex steps to ensure that the meteorological service push provided is more accurate and personalized.

[0211] First, the user data mining system conducts preliminary training and debugging of the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged based on the subscription interest error variable, the behavior description derivative error variable and the confidence coefficient thermal error variable. The purpose of this step is to use these error variables to guide the learning and optimization of the network so that it can better understand and predict user behavior and preferences. After preliminary training, the user data mining system obtains the discrete structured behavior semantic embedding network to be verified and the discrete structured push preference parsing network to be verified.

[0212] Next, to verify the performance of these networks, the user data mining system obtains user session behavior verification data and the prior push labels of these data. Prior push labels are service push labels that have been previously determined based on user behavior and preferences, and they will serve as a benchmark for verifying the accuracy of network predictions.

[0213] Then, the user data mining system inputs the user session behavior verification data into the network to be verified, and obtains the weather service push verification tags. These verification tags are the service push results predicted by the network based on the input data.

[0214] Next, the user data mining system performs consistency judgment on the verification label pushed by the weather service and the prior pushed label. The purpose of this process is to measure the degree of consistency between the network predicted label and the true label. Through consistency judgment, the user data mining system obtains a consistency judgment weight, which reflects the similarity or difference between the verification label and the prior label.

[0215] At the same time, the user data mining system also calculates the first evaluation index and the second evaluation index of the probability model of the meteorological service push verification label and the prior push label. These evaluation indicators may be accuracy, recall rate, F1 score, etc., which are used to quantitatively evaluate the prediction performance of the network.

[0216] According to the first evaluation index and the second evaluation index, the user data mining system determines a discrete evaluation index, which is an overall performance evaluation index obtained by comprehensively considering multiple evaluation factors.

[0217] Finally, the user data mining system obtains a global network quality evaluation value based on the consistency judgment weight and discrete evaluation index. This evaluation value is a comprehensive measure of the performance of the entire network. When the global network quality evaluation value is higher than the set evaluation value, it means that the performance of the network has reached the expected standard. At this time, the user data mining system determines the discrete structured behavior semantic embedding network to be verified as the final discrete structured behavior semantic embedding network, and determines the discrete structured push preference parsing network to be verified as the final discrete structured push preference parsing network.

[0218] In this way, through a series of refined training and verification steps, the user data mining system can ensure that its internal network model has a higher degree of accuracy and personalization in weather service push. This not only improves the user experience, but also enables the user data mining system to more accurately meet the user's weather service needs.

[0219] In combination with the above content, the diverse customized push text includes a page push message; the method also includes: performing page push message identification processing based on the meteorological service push tag to generate a page push message identification result; and issuing the page push message identification result.

[0220] In the meteorological service application scenario, the user data mining system will not only push relevant meteorological information based on the user's behavior and preferences, but also push it in a diversified and customized way to meet the different information receiving habits and needs of users. One of the important push forms is page push messages.

[0221] When the user data mining system determines the weather service push tags, these tags are used not only to guide what kind of weather information to push, but also to determine how to push this information. In this scenario, the user data mining system will identify and process the page push messages based on the weather service push tags.

[0222] For example, the user data mining system will first analyze the information in the weather service push tags, which may include multiple dimensions such as weather conditions, temperature range, wind force level, air quality, etc. The user data mining system will identify which information is most suitable for presentation in the form of page push messages based on the specific content in these tags. For example, if the tag contains keywords such as "rainstorm warning", the user data mining system may identify that this is a message that needs to be pushed urgently.

[0223] Next, the user data mining system will generate a page push message recognition result. This result is actually a customized push plan, which includes multiple factors such as the specific content to be pushed, the timing of the push, and the method of pushing. The user data mining system will comprehensively consider factors such as the user's historical behavior data, current weather conditions, and user needs and preferences to generate this recognition result.

[0224] Finally, the user data mining system will send the page push message identification results. During the sending process, the user data mining system will ensure that the message can be accurately delivered to the user's device and presented in a way that is easy for the user to receive and understand. This may include a notification containing key weather information popping up directly on the user's mobile phone or computer screen, or automatically jumping to a page containing detailed weather information when the user opens an application.

[0225] In this way, the user data mining system can achieve more accurate and personalized weather service push. Users can not only receive weather information related to their needs in a timely manner, but also choose the way and time to receive information according to their preferences and habits. This undoubtedly greatly improves the user experience and makes weather services closer to the actual needs of users. It can be seen that by identifying and processing page push messages based on weather service push tags and issuing identification results, the user data mining system successfully combines weather information with users' personalized needs, and realizes efficient, accurate and diversified weather service push. This not only improves users' satisfaction and dependence on weather services, but also lays a solid foundation for the further development and innovation of weather services.

[0226] In addition, in some independent embodiments, the method further includes: combining the meteorological service push tag and the meteorological user conversation behavior data to generate a target meteorological user portrait knowledge graph.

[0227] In detail, the step of generating a target meteorological user portrait knowledge graph by combining the meteorological service push tag and the meteorological user session behavior data includes:

[0228] Data integration step: Match and integrate the weather service push tags with the corresponding weather user session behavior data to form a user behavior-tag data set, where each user behavior-tag data contains the content, timestamp, behavior type and corresponding weather service push tag of the user session behavior;

[0229] Feature extraction step: extracting features from the data in the user behavior-label data set to identify key features of user behavior, including: session duration, session frequency, information click type, information focus, and feedback behavior; associating the key features of user behavior with corresponding weather service push labels;

[0230] Knowledge graph construction step: based on the user behavior-label data set and the key features of the user behavior, generate an initial meteorological user portrait knowledge graph, wherein the nodes in the initial meteorological user portrait knowledge graph represent user features or labels, and the edges represent the connections between features or labels;

[0231] Optimization and verification steps: optimizing the initial meteorological user portrait knowledge graph using a preset machine learning algorithm to obtain a meteorological user portrait optimized knowledge graph; mining the potential relationships between the nodes through the meteorological user portrait optimized knowledge graph; verifying the meteorological user portrait optimized knowledge graph based on the potential relationships and a preset verification data set;

[0232] Update and maintenance steps: When the meteorological user portrait optimization knowledge graph passes the verification, the meteorological user portrait optimization knowledge graph is regularly updated based on real-time user session behavior data.

[0233] In the meteorological service application scenario, the user data mining system not only focuses on how to accurately push meteorological information, but also strives to deepen the understanding of user behavior in order to provide more personalized services. To this end, the user data mining system implements an important step: combining meteorological service push tags and meteorological user session behavior data to generate a target meteorological user portrait knowledge graph. First, enter the data integration step, the user data mining system will match and integrate each meteorological service push tag with the corresponding meteorological user session behavior data. This process forms a user behavior-tag dataset, which records in detail the specific content of the user session behavior, the timestamp of occurrence, the behavior type (such as browsing, clicking, commenting, etc.), and the meteorological service push tags related to it. For example, after a user receives a push with the "heavy rain warning" tag, he clicks on the behavior of viewing detailed information, and this data will be integrated into the dataset. The next step is the feature extraction step, and the user data mining system will conduct an in-depth analysis of the integrated data to extract the key features of user behavior. These features include, but are not limited to, session duration (how long a user stays in a session), session frequency (how often a user conducts a session), information click type (which type of weather information a user is more inclined to click), information focus (content that a user pays special attention to in weather information, such as temperature, wind speed, etc.), and feedback behavior (user response to pushed information, such as likes, comments, or shares). These key features are closely related to the corresponding weather service push tags, providing a rich data foundation for the subsequent knowledge graph construction. Entering the knowledge graph construction step, the user data mining system will generate an initial weather user portrait knowledge graph based on the above user behavior-label dataset and user behavior key features. In this graph, each node represents a user feature or label, and the edge represents the connection between these features or labels. For example, a user who "often pays attention to rainstorm warnings" may be labeled as "sensitive to extreme weather" and connected to the feature node "frequently checks weather information". This is followed by the optimization and verification step, in which the user data mining system optimizes the initial weather user portrait knowledge graph using a preset machine learning algorithm. During this process, the user data mining system will continuously explore the potential relationships between nodes to make the knowledge graph more accurate and complete. After the optimization is completed, the user data mining system will use the preset verification data set to verify the effectiveness of the optimized knowledge graph. For example, by comparing the user behavior in the verification data set and the user characteristics in the knowledge graph, it can be tested whether the graph can accurately reflect the user's actual behavior and preferences. The last step is the update and maintenance step. Once the meteorological user portrait optimization knowledge graph passes the verification, the user data mining system will regularly update the graph based on real-time user session behavior data. This ensures that the graph can be continuously adjusted and optimized as user behavior changes, and always maintains an accurate grasp of user needs.Through this series of steps, the user data mining system successfully combined the weather service push tags and weather user conversation behavior data to generate a comprehensive, dynamic and accurate target weather user portrait knowledge graph. This graph not only helps the user data mining system to understand user needs more deeply, but also provides strong data support for the push of personalized weather services.

[0234] In this way, by generating a knowledge graph of target meteorological user portraits, the user data mining system can more accurately grasp user needs and preferences, thereby providing users with more personalized meteorological services that meet actual needs. This not only improves user experience, but also further enhances the pertinence and practicality of meteorological services.

[0235] Further, Figure 2 Schematic diagram of the structure of a user data mining system 200 provided in an embodiment of the present application. Figure 2 The user data mining system 200 shown includes a processor 210, which can call and run a computer program from a memory to implement the method in the embodiment of the present application.

[0236] Alternatively, if Figure 2 As shown, the user data mining system 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method in the embodiment of the present application.

[0237] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0238] Alternatively, if Figure 2 As shown, the user data mining system 200 may further include a transceiver 220. The processor 210 may control the transceiver 220 to interact with other devices. Specifically, the transceiver 220 may send information or data to other devices, or receive information or data sent by other devices.

[0239] Optionally, the user data mining system 200 may implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the embodiments of the present application, which will not be described here for the sake of brevity.

[0240] It should be understood that the processor of the embodiment of the present application may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method embodiment can be completed by the hardware integrated logic circuit or software instructions in the processor. The above processor can be a general processor, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to perform, or the hardware and software modules in the decoding processor are combined and performed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0241] It can be understood that the memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct RAM bus random access memory (DR RAM). It should be noted that the memory of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0242] It should be understood that the above-mentioned memory is exemplary but not restrictive. For example, the memory in the embodiments of the present application may also be static random access memory (static RAM, SRAM), dynamic random access memory (dynamic RAM, DRAM), synchronous dynamic random access memory (synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (double data rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (enhanced SDRAM, ESDRAM), synchronous link dynamic random access memory (synch link DRAM, SLDRAM) and direct memory bus random access memory (Direct Rambus RAM, DR RAM), etc. That is to say, the memory in the embodiments of the present application is intended to include but not limited to these and any other suitable types of memory.

[0243] Based on the above, a computer-readable storage medium is provided, on which a computer program is stored, and the computer program implements the above method when running.

[0244] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations.

Claims

1. A user data mining method for diversified customized meteorological services, characterized in that: The method is applied to a user data mining system, and the method comprises: Obtain the diversified customized push texts and meteorological user conversation behavior data of digital meteorological services; The diverse customized push text is used as a user demand identification indication, and the meteorological user conversation behavior data is subjected to behavioral semantic embedding mining through a discrete structured behavioral semantic embedding network, and personalized demand feedback elements are configured for the output of the behavioral semantic embedding mining to obtain a number of push preference decision knowledge features of the meteorological user conversation behavior data; The multi-modal customized push text is used as a user demand identification indication, and the push preference decision knowledge features are analyzed through a discrete structured push preference analysis network to obtain a number of meteorological content push preference viewpoints; the discrete structured behavior semantic embedding network and the discrete structured push preference analysis network are obtained by debugging behavior variation based on past user session behavior data; According to the meteorological content push preference viewpoint, a meteorological service push decision is made on the meteorological user conversation behavior data through a multivariate decision algorithm to obtain a meteorological service push label corresponding to each meteorological content push preference viewpoint; The method further comprises: Obtain several data category labels of the user's past session behavior data; According to the data category labels, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are debugged for behavior variation, so as to obtain the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network; According to the data category label, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are debugged for behavior variation, and the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network are obtained, including: Sequentially configuring different levels of subscription interest point features for each of the data category tags to obtain subscription interest description text for each data category tag under the influence of the subscription interest point features at different levels; According to the subscription interest description text under the influence of the subscription interest point features at different levels of each data category label, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are sequentially subjected to behavior description derivation to obtain the subscription interest error variables, behavior description derived error variables and confidence coefficient thermal error variables corresponding to the subscription interest point features at different levels; Based on the subscription interest error variable, the behavior description derived error variable and the confidence coefficient thermal error variable, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are trained and debugged to obtain the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network.

2. The method according to claim 1, characterized in that The method uses the diverse customized push text as a user demand identification indication, performs behavioral semantic embedding mining on the meteorological user conversation behavior data through a discrete structured behavioral semantic embedding network, and configures personalized demand feedback elements for the output of the behavioral semantic embedding mining, and obtains several push preference decision knowledge features of the meteorological user conversation behavior data, including: The diversified customized push text is used as a user demand identification indication, and a discrete structured behavior semantic embedding network is used to extract multi-level behavior semantics from the meteorological user session behavior data to obtain multi-level user session behavior semantics; Retrieving the associated session behavior semantics of the multi-level user session behavior semantics in the push mapping feature coordinate system through a preset matching rule; Configuring personalized demand feedback elements for the associated session behavior semantics to obtain a plurality of session behavior push associated semantics; Generate push preference decision knowledge features corresponding to each session behavior push association semantics.

3. The method according to claim 2, characterized in that The configuring of personalized demand feedback elements for the associated session behavior semantics to obtain a plurality of session behavior push associated semantics includes: Downsampling the associated session behavior semantics by a dynamic semantic downsampling algorithm to obtain page operation downsampling behavior semantics; Performing deep residual processing on the semantics of the downsampled behavior of the page operation, and obtaining feature influence coefficients of different attention dimensions in the semantics of the downsampled behavior of the page operation based on the output of the deep residual processing; Based on the discrete thermal analysis strategy, the feature influence coefficients of the different attention dimensions are configured with demand feedback definition features to obtain semantic definition coefficients of the different attention dimensions; According to the semantic definition coefficients of the different attention dimensions and the page operation downsampling behavior semantics, a plurality of session behavior push associated semantics are obtained.

4. The method according to claim 1, characterized in that The multi-modal customized push text is used as a user demand identification indication, and the push preference decision knowledge features are analyzed through a discrete structured push preference analysis network to obtain several meteorological content push preference viewpoints including: Performing feature projection on a plurality of the push preference decision knowledge features to obtain a plurality of push preference viewpoint scene projection features; The diverse customized push text is used as an indication of user demand identification, and a preference viewpoint identification operation is performed on each push preference viewpoint scene projection feature through a discrete structured push preference analysis network to determine the meteorological content push preference viewpoint corresponding to each push preference viewpoint scene projection feature.

5. The method according to claim 1, characterized in that According to the subscription interest description text of each data category label under the influence of the subscription interest point features at different levels, the behavior description is derived from the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged in turn, and the subscription interest error variables corresponding to the subscription interest point features at different levels are obtained, including: Combining the core subscription interest description text of each data category tag with the past data of the user session behavior to obtain the session subscription description combination text of each data category tag, wherein the core subscription interest description text is the subscription interest description text configured with the highest-level subscription interest point feature; The session subscription description combination text is subjected to behavioral semantic embedding mining through the discrete structured behavioral semantic embedding network to be debugged, and personalized demand feedback elements are configured for the output of the behavioral semantic embedding mining to obtain the push preference decision knowledge features corresponding to each session subscription description combination text; Performing a preference viewpoint recognition operation on the push preference decision knowledge features through the discrete structured push preference parsing network to be debugged, thereby obtaining the identified key subscription interest point features; Based on the core subscription interest description text and the key subscription interest point features of each data category label, generating a subscription interest description prediction text set for each data category label; Determine a subscription interest error variable of key subscription interest point features based on the subscription interest description prediction text set of each data category label and the subscription interest description priori text set of each data category label; The step of performing a preference viewpoint recognition operation on the push preference decision knowledge feature through the discrete structured push preference parsing network to be debugged to obtain the identified key subscription interest point features includes: performing a session behavior log retelling on the past data of the user session behavior through the discrete structured push preference parsing network to be debugged to obtain the user session behavior retelling log; mapping the push preference decision knowledge feature to the session behavior retelling feature coordinate system of the user session behavior retelling log to obtain the identified key subscription interest point features; Among them, the method also includes: comparing the past data of the user session behavior and the user session behavior retelling log to obtain the behavior description derived error variable corresponding to the key subscription interest point features; obtaining the feature influence coefficient and semantic definition coefficient of the push preference decision knowledge feature in different attention dimensions to obtain the confidence coefficient thermal error variable corresponding to the key subscription interest point features.

6. The method according to claim 1, characterized in that The method of training and debugging the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged based on the subscription interest error variable, the behavior description derived error variable and the confidence coefficient thermal error variable to obtain the discrete structured behavior semantic embedding network and the discrete structured push preference parsing network includes: Based on the subscription interest error variable, the behavior description derived error variable and the confidence coefficient thermal error variable, the discrete structured behavior semantic embedding network to be debugged and the discrete structured push preference parsing network to be debugged are trained and debugged to obtain the discrete structured behavior semantic embedding network to be verified and the discrete structured push preference parsing network to be verified; Obtaining user session behavior verification data and a priori push tags for the user session behavior verification data; Inputting the user session behavior verification data into the discrete structured behavior semantic embedding network to be verified and the discrete structured push preference parsing network to be verified to obtain a weather service push verification tag; Perform consistency judgment on the meteorological service push verification tag and the prior push tag to obtain a consistency judgment weight, and a first evaluation index and a second evaluation index of a probability model of the meteorological service push verification tag and the prior push tag; Determining a discrete evaluation index according to the first evaluation index and the second evaluation index; A global network quality evaluation value is obtained according to the consistency judgment weight and the discrete evaluation index. When the global network quality evaluation value is higher than the set evaluation value, the discrete structured behavior semantics to be verified is embedded in the network as a discrete structured behavior semantics embedding network, and the discrete structured push preference parsing network to be verified is used as a discrete structured push preference parsing network.

7. The method according to any one of claims 1 to 6, characterized in that: The diverse customized push text includes a page push message; the method further includes: performing page push message identification processing according to the meteorological service push tag to generate a page push message identification result; and issuing the page push message identification result.

8. A user data mining system, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 7.

Images