Intelligent mobile terminal tourism information personalized pushing system based on user portraits

Abstract

The invention discloses an intelligent mobile terminal tourism information personalized pushing system based on a user portrait, and relates to the technical field of mobile terminal tourism pushing, and the system comprises a user portrait construction module which collects basic information, mines behavior data and generates an interest label; the tourism information processing module is used for collecting multi-source tourism data, carrying out structured processing and evaluating the quality; the personalized push decision module is used for matching user interests and situations and determining a push strategy; the interaction feedback module is used for tracking user operation and optimizing portraits; and the privacy protection module is used for desensitizing the data, managing the authority, encrypting and storing. According to the method, the pushing accuracy and timeliness are improved, the user experience is optimized, the information quality and privacy security are guaranteed, and the method is adaptive to the whole travel process demand.

Description

Technical Field

[0001] This invention relates to the field of mobile terminal tourism push technology, and in particular to a personalized tourism information push system for intelligent mobile terminals based on user profiles. Background Technology

[0002] With the deep integration of mobile internet and the tourism industry, users' ways of obtaining travel information have shifted from traditional travel agencies and paper guides to smart mobile terminals, with various travel apps and OTA platforms becoming the main channels for information acquisition. However, current travel information push notifications generally suffer from the contradiction of "information overload" and "insufficient accuracy": on the one hand, travel information on the internet is exploding, covering multiple dimensions such as scenic spot introductions, accommodation bookings, transportation guides, and food recommendations, requiring users to spend a lot of time filtering effective content from a massive amount of information; on the other hand, most push systems adopt a "one-size-fits-all" approach, broadly pushing information based on popularity or regional tags, ignoring individual user differences, resulting in low matching between pushed content and users' actual needs. For example, families with children frequently receive information related to extreme sports, while users who prefer budget accommodations continuously see recommendations for high-end hotels. This mismatched push not only reduces the user experience but may also cause users to resist the push notification function.

[0003] The application of existing personalized recommendation technologies in tourism scenarios has significant limitations. While some systems attempt simple tagging based on users' historical behavior, these tags are too narrow and lack depth in understanding users' underlying needs, failing to capture complex preferences such as "family trips requiring facilities for children" or "hiking enthusiasts focusing on route difficulty and weather." Furthermore, users' travel needs are dynamic, changing over time and in different contexts, while existing systems often lag behind in updating user profiles, resulting in recommendations based on outdated data. In addition, the timeliness and accuracy of tourism information are difficult to guarantee; some recommendations fail to consider real-time factors such as temporary park closures, traffic restrictions, and price changes, leading to users receiving invalid or incorrect information and impacting their travel plans.

[0004] Balancing user privacy protection with personalized needs is also a pressing issue. Personalized recommendations rely on the collection and analysis of user data, but existing systems suffer from over-collection and insufficient data anonymization, raising user concerns about privacy leaks. Many users choose to disable data authorization due to concerns about the misuse of their personal location and spending records, making it difficult for the system to build a complete user profile and further exacerbating the problem of insufficient recommendation accuracy. Furthermore, the ability to integrate travel information across platforms is weak; when users switch between different apps, preference information cannot be shared, requiring repeated filtering and reducing the efficiency of travel planning. These problems collectively constrain the level of intelligence in travel information services, necessitating a personalized recommendation solution that can accurately depict user needs, dynamically adapt to scenarios, and ensure information quality and privacy security. Summary of the Invention

[0005] The present invention proposes a personalized tourism information push system for intelligent mobile terminals based on user profiles to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A personalized tourism information push system for intelligent mobile terminals based on user profiles includes the following modules:

[0008] The user profile building module includes a basic information collection component, which obtains user information through mobile terminal authorization and uses data cleaning algorithms to remove invalid information; a behavioral data mining component, which captures users' travel behavior on the terminal and uses a sliding window algorithm to count the frequency of behavior; and an interest tag generation component, which maps the mined behavioral data into structured tags, with tag weights dynamically adjusted according to the frequency of behavior.

[0009] The tourism information processing module includes an information collection component that connects to the platform and official data sources via API interfaces to collect tourism content; a structured processing component that transforms unstructured data into a unified format, extracts key information using natural language processing technology, and indexes time-series data by timestamp; and a quality assessment component that filters out qualified information through dimensional scoring and puts it into the push pool.

[0010] The personalized push decision-making module includes an interest matching component, which calculates the relevance between travel information and user interest tags, prioritizing content with completely identical three-level tags; a context adaptation component, which adjusts the push content based on the user's real-time context; and a push strategy generation component, which determines the push format and frequency based on the user's terminal usage habits and network environment.

[0011] The interactive feedback module includes a behavior tracking component that records user actions on push notifications and associates the action time and path; a feedback analysis component that converts action behavior into feedback weights to adjust the weights of corresponding interest tags; and a model iteration component that updates the user profile every 48 hours based on feedback data, using an incremental learning algorithm to retain effective parameters of the historical model and only correcting tags with large deviations.

[0012] The privacy protection module includes a data anonymization component that anonymizes sensitive user information and retains location information only at the district / county level; a permission management component that uses a fine-grained authorization mechanism, allowing users to individually control various permission switches; and an encrypted storage component that uses the AES-128 algorithm to encrypt user profile data and uses the TLS 1.3 protocol for data transmission.

[0013] Furthermore, it also includes a dynamic profile update module and an interest decay processing component. The dynamic profile update module includes a short-term interest capture component and a long-term feature stabilization component. The short-term interest capture component tracks sudden interests through a sliding window, generates temporary labels, and assigns them a high initial weight. The long-term feature stabilization component strengthens the label weight of stable interests that have lasted for more than 30 days by 20%. The interest decay processing component is used to correct the label weight of interests that have not been interacted with for a long time, and the calculation formula is I(t)=I0×e (-λt) Where I(t) is the interest weight at time t, I0 is the initial weight, λ is the decay coefficient, and t is the number of days since the most recent relevant behavior. When I(t) ≤ 0.1, the interest tag is automatically removed.

[0014] Furthermore, it also includes a context-aware enhancement module and a push timing optimization component. The context-aware enhancement module includes a user status recognition component and an environmental feature extraction component. The user status recognition component determines the user's status through terminal sensors; the environmental feature extraction component collects real-time environmental data; and the push timing optimization component calculates the push suitability score using the formula S = α × C1 + β × C2 + γ × C3, where S is the push suitability score, C1 is the user's idle timeliness, C2 is the information timeliness, and C3 is the environmental matching degree. A push is triggered when S ≥ 6.

[0015] Furthermore, the interest tag generation component of the user profile construction module adopts a multi-level tag system. The core tags include travel type, budget range, and travel mode. The secondary tags are subdivisions of the core tags. The tag weight is calculated based on the depth and frequency of the behavior: search keyword weight × 1.2, browsing time > 60 seconds × 1.5, collection operation × 2.0, completed order × 3.0. The influence factors of each level of tag are determined by the hierarchical analysis method, and the weight is recalculated every 7 days based on new behavioral data. The weight value ranges from 0 to 1.0.

[0016] Furthermore, the information collection component of the tourism information processing module adopts a combination of dynamic crawling and API integration: for high-frequency changing information, it is synchronized every 30 minutes via an encrypted API interface; for medium-frequency changing information, it is crawled every 6 hours; for UGC content, it is crawled and updated daily between 2-4 AM, and anti-crawling mechanisms are avoided during the crawling process through User-Agent rotation and IP proxy pools, and third-party recognition interfaces are automatically called for pages containing CAPTCHAs; the structured processing component performs three-level parsing on unstructured data: first-level parsing extracts entities, second-level parsing extracts attributes, and third-level parsing extracts sentiment; for content containing time information, it automatically associates with the calendar system to generate reminder time nodes; and for multilingual information, it uses a pre-trained neural machine translation model to convert it into the default language set by the user terminal.

[0017] Furthermore, the interest matching component of the personalized push decision module adopts a hybrid similarity calculation mechanism: for text information, a 768-dimensional feature vector is extracted using the BERT model, and the cosine similarity with the user's interest tag vector is calculated; for image / video information, a 2048-dimensional visual feature vector is extracted using the ResNet50 model, and the Euclidean distance is calculated as the similarity. The matching process is divided into two steps: the first step is to filter information with a similarity ≥ 0.7 into the candidate pool; the second step is to introduce the user's implicit preferences to adjust the ranking. In addition to the basic context, the context adaptation component also dynamically adjusts based on the user's travel stage: if the user is in the planning stage, information containing transportation + accommodation combinations is prioritized; if the user is in the travel stage, destination weather, real-time traffic, and a must-pack list are prioritized; if the user is in the return stage, local souvenirs and return travel guides are pushed; the push strategy generation component adapts the push format according to the terminal model: for devices with a screen size ≥ 6.7 inches, a graphic card containing 3 thumbnails is pushed; for small-screen devices, a text summary + 1 main image is pushed; for tablet devices, landscape immersive push is supported; the push frequency adopts intelligent traffic limiting.

[0018] Furthermore, the feedback analysis component of the interactive feedback module constructs a multi-dimensional feedback system: operation behavior feedback, dwell time feedback, and secondary behavior feedback. For user-inputted text comments, sentiment tendencies are extracted through a sentiment analysis model. The comprehensive feedback value is the weighted sum of feedback from each dimension. When the comprehensive feedback value is ≥1.5, the weight of the corresponding interest tag increases by 10%; when it is ≤-1.0, the weight decreases by 15%. The model iteration component adopts a federated learning framework, where 80% of the model update calculation is completed locally on the terminal. Only the parameter update difference is encrypted and uploaded to the server, aggregated with update data from other users to generate a global model, and then distributed to the terminal.

[0019] Furthermore, the data anonymization component of the privacy protection module adopts a hierarchical anonymization strategy: partial masking is used for absolutely sensitive information; obfuscation is used for relatively sensitive information; and k-anonymization is used for behavioral data. The permission management component provides a permission control panel, allowing users to individually enable / disable 12 permissions, each with a clear purpose description. After a permission is disabled, the system immediately stops collecting related data, and historical data that has already been collected is automatically deleted after 7 days and does not participate in subsequent profile calculations. The encrypted storage component uses a hardware security module to generate and store AES-128 encryption keys, which never leave the terminal. User profile data is encrypted and stored in blocks according to "basic information + behavioral characteristics + interest tags," with each block using an independent initialization vector. During data transmission, in addition to TLS 1.3 encryption, sensitive fields are additionally encrypted end-to-end.

[0020] Furthermore, the short-term interest capture component of the dynamic profile update module adopts a mechanism combining sliding window and event triggering: within a 24-hour sliding window, if a user exhibits sudden behavior, a temporary tag is generated, with the tag naming format being "temporary-specific interest," an initial weight of 0.6, and this tag enjoys a 20% ranking priority in push matching; the long-term feature stabilization component performs trend analysis on stable behaviors lasting more than 30 days; the interest migration identification component calculates the change rate of interest tag weights over three consecutive 7-day windows: current window weight - previous window weight / previous window weight. When the change rate is ≥50% and the correlation between the newly appearing interest tag and the original core tag is <0.3, it is determined to be an interest migration, at which point a rapid adaptation mechanism is activated: the new interest tag weight is updated every 12 hours within 3 days, and calibration is accelerated based on the new behavior density; the decay coefficient λ of the old core tag is increased by 20%, but the "transition weight" for 30 days is retained.

[0021] Furthermore, the user state recognition component of the context awareness enhancement module integrates multi-source sensor data: it uses an accelerometer to determine motion intensity, a gyroscope to determine changes in motion direction, and combines screen illumination frequency and APP usage sequence to comprehensively identify user state; the environmental feature extraction component, in addition to basic environmental data, also connects to third-party services to obtain special context information: it connects to the scenic spot reservation system to obtain real-time reservation occupancy rates, connects to the transportation department to obtain road construction and temporary control information, and connects to the cultural and tourism bureau to obtain special event information; the push timing optimization component, based on S≥6, refines the push scenarios: when the user is "planning a trip," it prioritizes pushing information containing price comparison and booking links; when the user is "leisurely browsing," it pushes information containing travel guides and experience videos; when the user is "waiting," it pushes information about nearby short-term tours; the alternative recommendation mechanism calculates the multi-dimensional similarity of scenic spots, filters out 3 alternative options with a similarity ≥0.8, sorts them by priority, and appends them below the original push information, labeled "similar recommendation" and the reason for the recommendation.

[0022] Compared with existing technologies, the beneficial effects of this invention are:

[0023] By constructing multi-dimensional user profiles, the system can deeply explore users' travel preferences, encompassing not only explicit browsing and order records but also implicit interest associations, forming a comprehensive user tagging system. This refined profiling allows travel information pushes to go beyond superficial matching, reaching deeper user needs. For example, it prioritizes recommending scenic spots with children's facilities and family-themed hotels for families with children, and focuses on recommending hiking routes and budget hostels for backpackers, significantly improving the relevance of push content to user needs.

[0024] Context-aware functionality dynamically adjusts push notifications based on factors such as the user's location, weather conditions, and travel stage. For example, it can push transportation and accommodation combination information when the user is planning their trip, provide real-time traffic updates and tourist attraction crowd alerts during the journey, and recommend souvenir purchase guides upon return, making the notifications more timely and practical. Simultaneously, the push strategy fully considers user habits, such as pushing notifications during peak activity times, adjusting content format based on network conditions, and intelligently controlling push frequency to avoid information overload. This ensures users receive the content they need precisely when they need it, improving the efficiency and experience of travel planning.

[0025] The system uses multi-source information collection and a rigorous quality assessment mechanism to filter out authentic, timely, and complete tourism information, preventing users from being misled by outdated or false content. The privacy protection module employs tiered anonymization, fine-grained access control, and encrypted storage technologies. While acquiring necessary user data to build user profiles, it maximizes user privacy, alleviates user concerns about data leaks, and enhances user trust in the system.

[0026] Furthermore, the system possesses self-optimization capabilities, continuously iterating and updating user profiles and push strategies based on user feedback, thus gradually improving push effectiveness over time. Its cross-scenario adaptability and information integration capabilities also enable the system to meet the diverse needs of users throughout the entire travel process, providing a smarter and more considerate solution for smart mobile terminal tourism information services. Attached Figure Description

[0027] Figure 1 This is a schematic block diagram of the intelligent mobile terminal personalized tourism information push system based on user profile proposed in this invention.

[0028] Figure 2 A bar chart comparing the accuracy of push notifications across different systems;

[0029] Figure 3 A line graph showing the distribution of user interaction depth;

[0030] Figure 4 A line chart showing the dynamic changes in user profile tag weights. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0033] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0034] Reference Figures 1 to 4 A personalized tourism information push system for intelligent mobile terminals based on user profiles, comprising the following modules:

[0035] The user profile building module is used to establish a multi-dimensional travel characteristic model of users, including a basic information collection component, a behavioral data mining component, and an interest tag generation component. The basic information collection component obtains user registration information, device information, and account association information through mobile terminal authorization. Registration information includes age, gender, and place of residence; device information includes terminal model and operating system; and account association information includes travel-related activities on social media accounts. Data cleaning algorithms are used to remove invalid information, such as duplicate submissions of the same address. The behavioral data mining component captures users' travel-related behaviors on the terminal in real time, including browsing time in the APP, search keywords, collection and sharing operations, and order history. A single piece of information is considered valid if it stays for more than 3 seconds. Search keywords include "family-friendly homestay" and "hiking route." Order history includes hotel type and travel mode. A sliding window algorithm is used to count the frequency of behavior, with a window size of 7 days. The interest tag generation component maps the mined behavioral data into structured tags, organized in a three-level structure of "travel type - preference level - associated features," such as "natural landscape - high - mountain / spring" and "food experience - medium - local snacks / night market." The tag weights are dynamically adjusted according to the frequency of behavior.

[0036] The tourism information processing module is used to integrate and optimize multi-source tourism data. It includes an information collection component, a structured processing component, and a quality assessment component. The information collection component connects to tourism platforms, social media platforms, and official data sources via API interfaces. Tourism platforms include OTA websites and scenic spot official websites; social media platforms include travel guide communities; and official data sources include meteorological departments and transportation bureaus. The collected content covers scenic spot information, accommodation resources, transportation information, and folk activities. Scenic spot information includes opening hours, ticket prices, and visitor flow index; accommodation resources include room types, reviews, and price fluctuations; transportation information includes real-time traffic conditions and service schedules; and folk activities include event dates, etc. Participation Method: The structured processing component transforms unstructured data into a unified format. Unstructured data includes user travelogues and reviews. Natural language processing technology is used to extract key information. For example, if the word "quiet" appears three times in a review, it is marked with the "quiet environment" attribute. Time-series data, including ticket prices, is also indexed by timestamp. The quality assessment component scores information based on three dimensions: credibility of the information source, timeliness of updates, and completeness of content. Official data sources have a weight of 0.8, user-generated content has a weight of 0.5, updates within 24 hours are considered valid, and content containing more than 5 core attributes is considered complete. Finally, information with a score ≥ 0.7 is selected for the push pool.

[0037] The personalized push decision module is used to match the best travel information based on user profiles. It includes an interest matching component, a context adaptation component, and a push strategy generation component. The interest matching component calculates the degree of fit between travel information and user interest tags, prioritizing content with completely consistent three-level tags. For example, if the user's tag is "family travel - high-end - theme park", information containing "theme park + family facilities" will be prioritized. The context adaptation component adjusts the push content based on the user's real-time context. The current location is determined by GPS with an accuracy of ≤10 meters, the weather is determined by location and meteorological data, and the travel time is obtained through calendar authorization. For example, indoor attraction information will be prioritized on rainy days. The push strategy generation component determines the push format and frequency based on the user's terminal usage habits and network environment. If the user's active period is concentrated between 19:00 and 21:00, pushes will be made during this period. Rich media content including video will be pushed in WiFi environments, and text summaries will be pushed in mobile networks. Push formats include notification bar messages, in-app recommendation pages, and SMS reminders, with a push frequency of no more than 5 messages per day.

[0038] The interaction feedback module tracks user responses to push notifications and optimizes the system. It includes a behavior tracking component, a feedback analysis component, and a model iteration component. The behavior tracking component records user actions on push notifications, including clicks, ignores, blocks, and secondary sharing. It also associates the action time with the action path, such as whether the details page was viewed after clicking. The feedback analysis component converts actions into feedback weights, with clicks receiving +0.3, blocks receiving -0.5, and shares receiving +0.8, thereby adjusting the weights of corresponding interest tags. The model iteration component updates the user profile every 48 hours based on feedback data. It uses an incremental learning algorithm to retain 80% of the effective parameters of the historical model and only corrects tags with large deviations. Tags with large deviations are specifically those with a cumulative feedback weight ≤ -1.0.

[0039] The privacy protection module safeguards user data security and privacy, and includes a data anonymization component, a permission management component, and an encrypted storage component. The data anonymization component anonymizes sensitive user information (such as ID card number and precise address) by retaining the first 6 digits and the last 4 digits, replacing the rest with asterisks (*). Location information is only retained at the district / county level. The permission management component employs a fine-grained authorization mechanism, allowing users to individually enable / disable permissions such as "location information collection" and "consumption record analysis." Disabling these permissions prevents the updating of related profile tags. The encrypted storage component uses the AES-128 algorithm to encrypt user profile data. The key is generated locally on the terminal and is not uploaded. Data transmission uses the TLS 1.3 protocol to prevent information leakage.

[0040] This invention also includes a dynamic profile update module and an interest decay processing component. The dynamic profile update module comprises a short-term interest capture component and a long-term feature stabilization component. The short-term interest capture component tracks sudden interests through a sliding window, with the sliding window size being 24 hours. A sudden interest is, for example, a temporary search for "weekend getaways," and a temporary tag is generated and assigned a relatively high initial weight of 0.6. The long-term feature stabilization component strengthens the tag weight for stable interests that have persisted for more than 30 days. A stable interest is, for example, browsing "historical sites" at least twice a month, with the tag weight specifically increased by 20%. The interest decay processing component is used to correct the tag weights of interests that have not been interacted with for a long time, calculated using the formula I(t) = I0 × e (-λt) Where I(t) is the interest weight at time t, I0 is the initial weight (calculated based on historical behavior), λ is the decay coefficient (set according to interest type: λ = 0.02 / day for leisure, λ = 0.03 / day for adventure), and t is the number of days since the most recent relevant behavior (unit: days). When I(t) ≤ 0.1, the interest tag is automatically removed to ensure that the user profile is consistent with the current preferences.

[0041] This invention also includes a context-aware enhancement module and a push timing optimization component. The context-aware enhancement module comprises a user state recognition component and an environmental feature extraction component. The user state recognition component determines the user's state (stationary, walking, riding in a vehicle) using terminal sensors (accelerometer, gyroscope), with an accuracy rate ≥90%. The environmental feature extraction component collects real-time environmental data (such as holiday markers, regional population density, and epidemic risk levels). The push timing optimization component calculates the push suitability score using the formula S=α×C1+β×C2+γ×C3, where S is the push suitability score (0-10 points), C1 is the user's idle time (8-10 points for stationary status and non-working hours, 5-7 points for riding in a vehicle, and 1-3 points for working hours, weight α=0.4), C2 is the information timeliness (10 points for information valid within 2 hours, 7 points for information valid within 1 day, weight β=0.3), and C3 is the environmental matching degree (e.g., 8 points for matching rainy days with indoor attractions, weight γ=0.3). A push notification is triggered when S ≥ 6 minutes, ensuring that the information is delivered at the time when the user is most likely to pay attention.

[0042] In this invention, the interest tag generation component of the user profile construction module adopts a multi-level tag system. The core tags include travel type, budget range, and travel mode. Among them, the travel type includes 8 categories such as natural landscapes, historical sites, and theme parks. The budget range is divided into economy (<500 yuan / day), comfort (500-1000 yuan / day), and high-end (>1000 yuan / day). The travel mode covers self-driving, public transportation, and group tours. The secondary tags are subdivisions of the core tags. For example, "natural landscapes" includes 12 categories such as mountains, lakes, forests, and beaches. "Historical sites" includes 9 categories such as ancient towns, museums, and temples. The derived tags are scenario-based features. For example, "mountains" are associated with hiking difficulty (beginner, intermediate, advanced), best season (spring, summer, autumn, winter), and supporting facilities (parking lot, dining area), etc. Tag weight calculation combines behavioral depth and frequency: search keyword weight × 1.2, browsing time > 60 seconds × 1.5, collection action × 2.0, completed order × 3.0; the influence factor of each level of tag is determined by the hierarchical analysis method, with core tags at 0.5, secondary tags at 0.3, and derived tags at 0.2. The weights are recalculated every 7 days based on new behavioral data, with a weight value range of 0-1.0, where ≥ 0.6 is considered high preference, 0.3-0.5 is considered medium preference, and < 0.3 is considered low preference. This ensures that the tag system can reflect both the core needs of users and capture detailed preferences.

[0043] In this invention, the information collection component of the tourism information processing module adopts a combination of dynamic crawling and API integration: high-frequency changing information such as hotel prices and room status on OTA platforms is synchronized every 30 minutes via an encrypted API interface; medium-frequency changing information such as opening hours and event announcements on scenic spot official websites is crawled every 6 hours; and user-generated content such as travelogues and guides is crawled and updated daily between 2-4 AM (when network load is low). During the crawling process, anti-crawling mechanisms are avoided through User-Agent rotation and an IP proxy pool (containing 1000+ proxy IPs). For pages containing CAPTCHAs, a third-party recognition interface is automatically called (supporting slider, image and text, and click-based CAPTCHAs, with a recognition success rate of ≥95%). The structured processing component performs three levels of parsing on unstructured data: Level 1 parsing extracts entities (scenic spot names, hotel names, restaurant names, etc.); Level 2 parsing extracts attributes (such as "soundproofing" and "service attitude" dimensions from "hotel reviews"); and Level 3 parsing extracts sentiment (positive / negative / neutral). For content containing time information (such as "folk performances every Saturday night"), it automatically links it to the calendar system to generate reminder time nodes. For multilingual information, a pre-trained neural machine translation model (supporting 10 languages ​​including Chinese, English, Japanese, and Korean, with a BLEU value ≥ 0.85) is used to convert it into the default language set by the user terminal, ensuring the readability of cross-language information.

[0044] In this invention, the interest matching component of the personalized push decision module adopts a hybrid similarity calculation mechanism: for text-based information (guides, reviews), a 768-dimensional feature vector is extracted using the BERT model, and the cosine similarity with the user's interest tag vector is calculated; for image / video-based information (scenic spots, hotel environments), a 2048-dimensional visual feature vector is extracted using the ResNet50 model, and the Euclidean distance is calculated as the similarity. The matching process consists of two steps: the first step is to filter information with a similarity ≥ 0.7 into the candidate pool; the second step introduces the user's implicit preferences to adjust the ranking. For example, if a user ignores the "high-end hotel" push notification three times in the last three times, the ranking weight of similar information in the candidate pool is reduced by 30%; if a user clicks on "family-themed hotel" twice in the last two times, the ranking weight of similar information is increased by 50%. In addition to basic scenarios, the context-adaptive component dynamically adjusts based on the user's travel stage: if the user is in the "planning phase" (has not placed an order but has searched for the destination), information including transportation and accommodation packages is prioritized; if the user is in the "travel phase" (has placed an order and is ≤3 days away from the travel date), destination weather, real-time traffic, and a must-pack list are prioritized; if the user is in the "return phase" (travel date has passed), local souvenirs and return travel guides are pushed. The push strategy generation component adapts the push format according to the terminal model: for devices with a screen size ≥6.7 inches, a graphic card with 3 thumbnails is pushed; for smaller screen devices, a text summary + 1 main image is pushed; for tablet devices, landscape immersive push (videos automatically mute) is supported; the push frequency uses intelligent traffic limiting: ≤3 pushes per day on weekdays (concentrated at 8 am, 12 pm, and 8 pm), ≤5 pushes per day on weekends / holidays (adding the 3 pm time slot), and no more pushes are sent that day if the user ignores the push twice consecutively to avoid information harassment.

[0045] In this invention, the feedback analysis component of the interactive feedback module constructs a multi-dimensional feedback system: operation behavior feedback (click +0.3, view details +0.5, favorite +0.8, share +1.0, ignore -0.2, block -0.5, report -1.0); dwell time feedback (<3 seconds -0.1, 3-10 seconds +0.1, 10-30 seconds +0.2, >30 seconds +0.4); secondary behavior feedback (jump from push page to booking page +1.2, complete order +2.0, add to itinerary planning +0.6). For user-inputted text comments, sentiment analysis model (based on LSTM + attention mechanism, accuracy ≥88%) is used to extract sentiment tendency: positive comments (such as "very useful", "exactly what I wanted") +0.6, negative comments (such as "information error", "not interested") -0.8, and neutral comments are not assigned a value. The overall feedback value is a weighted sum of feedback from each dimension (40% for operational behavior, 20% for dwell time, 30% for secondary behavior, and 10% for text comments). When the overall feedback value is ≥1.5, the weight of the corresponding interest tag increases by 10%; when it is ≤-1.0, the weight decreases by 15%. The model iteration component adopts a federated learning framework, with 80% of the model update calculations completed locally on the terminal. Only the parameter update difference is encrypted and uploaded to the server, aggregated with update data from other users to generate a global model, and then distributed to the terminal. This ensures that user data is not leaked while improving the model's generalization ability.

[0046] In this invention, the data anonymization component of the privacy protection module adopts a hierarchical anonymization strategy: For absolutely sensitive information (ID card number, bank card number), partial masking is used (displaying the first 6 digits + the last 4 digits, with padding in between, e.g., ID card number displayed as 110101********1234); for relatively sensitive information (phone number, precise address), obfuscation is used (phone number displayed as 138***5678, address displayed as "Chaoyang District, Beijing"); and for behavioral data, k-anonymization is used (merging user behavior with at least k-1 similar user behaviors, k≥5). The permission management component provides a fine-grained permission control panel, allowing users to individually enable / disable 12 permissions, such as "Location Information (Always Allowed / Allowed Only When Used / Disallowed)," "Consumption Record Analysis," "Social Relationship Association," and "Device Information Collection." Each permission corresponds to a clear purpose description (e.g., "Location information is used to recommend nearby attractions"). After a permission is disabled, the system immediately stops collecting related data; historical data already collected is retained for 7 days and then automatically deleted, and does not participate in subsequent profile calculations. The encrypted storage component uses a hardware security module (TEE) to generate and store AES-128 encryption keys, which never leave the terminal; user profile data is encrypted and stored in blocks according to "basic information + behavioral characteristics + interest tags", with each block using an independent initialization vector (IV); during data transmission, in addition to TLS 1.3 encryption, sensitive fields (such as location coordinates) are additionally encrypted end-to-end (using the receiver's public key) to ensure dual security of the transmission link and server storage.

[0047] In this invention, the short-term interest capture component of the dynamic profile update module adopts a mechanism combining a sliding window and event triggering: within a 24-hour sliding window, if a user exhibits sudden behavior, such as searching for "hot spring hotel" three times consecutively within one hour or sharing two skiing guides, a temporary tag is generated. The tag naming format is "temporary-specific interest," for example, "temporary-hot spring hotel," with an initial weight of 0.6, and this tag enjoys a 20% ranking priority in push matching. The long-term feature stabilization component performs trend analysis on stable behaviors lasting more than 30 days: if an interest tag, such as "historical sites," has a monthly weight ≥0.6 for three consecutive months and a monthly behavior frequency ≥5 times, it is marked as a "core interest," with a 20% increase in weight and a reduced decay rate, with the decay coefficient λ lowered to 50% of its original value; if the weight fluctuates between 0.3 and 0.5 for two consecutive months, but the behavior frequency is stable, it is marked as a "stable interest," with the weight remaining unchanged. The interest migration identification component calculates the rate of change in interest tag weights over three consecutive 7-day windows. The rate of change is calculated by subtracting the weight of the previous window from the current window's weight and then dividing by the previous window's weight. When the rate of change is ≥50% and the correlation between the newly emerging interest tag (e.g., "camping gear") and the original core tag (e.g., "family travel") is <0.3 (the correlation is calculated using tag co-occurrence probability), it is determined to be an interest migration. At this point, a rapid adaptation mechanism is activated: the weights of new interest tags are updated every 12 hours within 3 days, and calibration is accelerated based on the new behavior density. The weight increases by 0.1 for each valid behavior. The decay coefficient λ of the old core tags is increased by 20%, but a 30-day "transition weight" is retained, meaning it does not immediately drop to 0, to avoid sudden changes in pushed content affecting the user experience.

[0048] In this invention, the user status recognition component of the context awareness enhancement module integrates multi-source sensor data: an accelerometer determines motion intensity (≤0.1g for stationary motion, 0.1-0.5g for walking, and 0.5-2g for riding in a vehicle); a gyroscope determines changes in motion direction (regular fluctuations during walking, more pronounced fluctuations during riding in a vehicle); and the user status is comprehensively identified by combining screen illumination frequency (typically once every 10-30 minutes during riding in a vehicle) and APP usage sequence (e.g., continuously opening a map app and a travel app indicates trip planning). The recognition accuracy is ≥92%. In addition to basic environmental data, the environmental feature extraction component also connects to third-party services to obtain special contextual information: it connects to the scenic area reservation system to obtain real-time reservation occupancy rates (e.g., a scenic area with an occupancy rate ≥90% is marked as "crowded"); it connects to the transportation department to obtain information on road construction and temporary traffic control; and it connects to the cultural and tourism bureau to obtain information on special events such as holiday activities and cultural festivals. The push notification timing optimization component, based on an S≥6 score, further refines the push scenarios: when users are "planning their trip" (opening a map app and searching for a destination), information containing price comparisons and booking links is prioritized; when users are "leisurely browsing" (opening social media but not searching for a specific destination), information containing travel guides and experience videos is pushed; when users are "waiting" (e.g., stationary for more than 30 minutes in a train station / airport WiFi environment), information on nearby short-term tours is pushed. The alternative recommendation mechanism calculates the multi-dimensional similarity of scenic spots: type matching accounts for 40%, user rating similarity accounts for 30%, distance proximity accounts for 20%, and price range similarity accounts for 10%. It selects three alternative options with a similarity ≥0.8, sorts them by priority, and appends them below the original push information, labeled "Similar Recommendation" and the reason for the recommendation (e.g., "It belongs to the same ancient town type as the scenic spot A you are interested in, and has fewer tourists").

[0049] The following two examples further illustrate the specific implementation of this system:

[0050] Example 1: Daily travel planning scenario (applicable to users' daily browsing and short trip plans)

[0051] When the user profile building module is running, the basic information collection component obtains the user's registration information through the authorized travel app: age 32, female, permanent residence in Hangzhou, device is an iOS system phone (model iPhone 13), and the associated social media account has posted 3 hiking photos around West Lake in the past 3 months. The data cleaning component removes two duplicate address entries submitted by the user and retains the most recent records. The behavioral data mining component uses a 7-day window to collect statistics on the user in the past week: browsing articles related to "ancient towns around Hangzhou" 4 times (each visit lasting 8-15 seconds), searching keywords "weekend family-friendly homestays" and "Fuchun River hiking route", collecting 2 "Tips for Hiking with Kids" guides, and no order records. The interest tag generation component maps these behaviors to three-level tags: "Natural Landscape - Medium - Hiking / Spring", "Ancient Town - High - Family-Friendly / Weekend", and "Homestay - Medium - Family-Friendly Facilities / Price ≤ 800 RMB". The tag weights are calculated as follows: Search behavior × 1.2, Favorites × 2.0, and Browsing × 1.0. Finally, the tag weights of "Ancient Town - High" are 0.7, "Natural Landscape - Medium" are 0.5, and "Homestay - Medium" are 0.6.

[0052] In the tourism information processing module, the information collection component synchronizes Hangzhou surrounding homestay prices from OTA platforms every 30 minutes. For example, the real-time price of "Fuyang Family Homestay" is 680 yuan / night. It crawls information from scenic spot websites every 6 hours, such as the opening hours of Tangqi Ancient Town (8:00-17:00). It also crawls user travelogues from tourism communities at 3 AM daily, extracting information such as "Tangqi Ancient Town is suitable for families with children" and "The Fuchun River hiking is easy." The structured processing component performs a three-level analysis on a user travelogue: Level 1 extracts entities "Tangqi Ancient Town" and "Guangji Bridge"; Level 2 extracts attributes such as "ample parking" and "children's play area"; Level 3 judges the sentiment as positive. The quality assessment component scores the collected information: the source credibility of information from the Tangqi Ancient Town official website is 0.8, update timeliness is 0.9, and completeness is 0.9, for a total score of 0.86, and it is included in the push pool; the source credibility of a personal blog's "Fuchun River Hiking Guide" is 0.4, for a total score of 0.65, and it is filtered out.

[0053] In the personalized push decision-making module, the interest matching component calculates the similarity between candidate information and user tags: the combination of "Tangqi Ancient Town + Family-Friendly Homestay" has a cosine similarity of 0.82 with the tag "Ancient Town - High-end - Family-Friendly," and enters the candidate pool; "West Lake Cruise" has a similarity of 0.45 with the user tag and is excluded. The context adaptation component confirms that the user is currently in Hangzhou city through GPS positioning (accuracy 8 meters), and combined with meteorological data, it knows that the next day will be cloudy (15-22℃), and the user's calendar shows that the weekend is free, so it prioritizes matching short-distance information that is "suitable for weekend travel." The push strategy generation component pushes information in the form of a picture and text card in the notification bar at 20:00 on Saturday night, based on the user's terminal (6.1-inch screen) and active time period (the highest frequency of opening the APP in the past 7 days is 19:30-21:00). The message includes a photo of the entrance to Tangqi Ancient Town + the title "Top 3 Family-Friendly Ancient Towns" + homestay price, and because the user is detected to be connected to WiFi, a 30-second short video of the ancient town tour is attached.

[0054] In the interactive feedback module, the user clicks on the push notification card and views the details page (staying for 22 seconds), then shares it to the family group. The behavior tracking component records these actions, and the feedback analysis component calculates the comprehensive feedback value: click +0.3 + view details +0.5 + share +1.0 + stay for 22 seconds +0.2 = 2.0. Because the comprehensive feedback value is ≥1.5, the weight of the "Ancient Town - High" tag increases by 10% to 0.77. After 48 hours, the model iteration component uses incremental learning to retain 80% of the original model's parameters, only slightly adjusting the weight of the "Homestay" tag to 0.65 (because the user did not click on the homestay link).

[0055] In the dynamic profile update module, the short-term interest capture component generates a temporary tag "Temporary - Tea Culture Experience" (initial weight 0.6) after a user suddenly searches for "Jingshan Tea Picking Experience" on Sunday, and prioritizes matching related information in the daily push notifications. The interest decay processing component calculates the decay for the "Island Tour" tag that the user viewed 3 months ago: I0 = 0.4, λ = 0.02 / day (leisure category), t = 90 days, I(t) = 0.4 × e (-0.02×90) =0.4×0.165=0.066≤0.1, automatically remove the tag.

[0056] The push timing optimization component calculates the suitability of pushing "Jingshan Tea Picking" information on Sunday: C1 = 9 points (user is in a static state, outside of working hours), C2 = 7 points (tea picking activity starts the next day), C3 = 8 points (environmental matching degree, spring is suitable for tea picking), S = 0.4×9 + 0.3×7 + 0.3×8 = 3.6 + 2.1 + 2.4 = 8.1 ≥ 6 points, trigger push.

[0057] Example 2: Holiday travel scenario (applicable to long-distance travel during Golden Week and short holidays)

[0058] The user profile building module targeted a user (28-year-old male, residing in Shanghai, using an Android tablet) planning a National Day trip. Basic information showed that he had three inter-provincial group tour bookings in the past year and had shared content related to "Xi'an Terracotta Warriors" and "Chengdu Hot Pot" on his social media accounts. The behavioral data mining component (7-day window) found that he searched "National Day Beijing independent travel guide" 6 times, browsed "Forbidden City reservation tips" and "Beijing roast duck restaurant rankings", saved "7-day 6-night Beijing family route", and placed an order for "a hotel in Beijing (price 1200 yuan / night)". The interest tag generation component generated tags: "Cultural and Historical Sites - High - History / National Day", "Food - High - Local Specialties / Average Price ≤ 200 yuan", and "Hotels - High - Four-star and above / Available for National Day". Among them, order behavior × 3.0, search × 1.2, and the tag weights are all ≥ 0.7.

[0059] In the tourism information processing module, the information collection component synchronizes Beijing hotel prices every 30 minutes (detecting that the price of a hotel booked by a user has increased to 1500 yuan / night), crawls the Palace Museum's official website every 6 hours (confirming the National Day opening hours as 8:30-17:00, requiring reservations 7 days in advance), and connects with the Beijing Municipal Bureau of Culture and Tourism to obtain information on "National Day Cultural Activities on Qianmen Street." The structured processing component parses an article titled "Beijing National Day Crowd Forecast" into: entities "Palace Museum" and "Great Wall," attributes "estimated 80,000 visitors to the Palace Museum on October 1st," sentiment neutral, and generates a reminder for October 1st by associating it with a calendar. The quality assessment component scores "Palace Museum Official Website Reservation Information" at 0.9 (source credibility 0.8 + timeliness 0.9 + completeness 0.9) and includes it in the push pool.

[0060] In the personalized push decision module, the interest matching component calculates the similarity of the combination "In-depth tour of the Forbidden City + Four Seasons Minfu Roast Duck Restaurant". The text feature vector is extracted using BERT. The cosine similarity between this combination and the user's tag vector is 0.85, resulting in a final similarity of 0.85, which is then entered into the candidate pool. The context adaptation component detects that the user is in a "travel period", which is 7 days before the National Day holiday and the user has already booked a hotel. Therefore, it prioritizes pushing "Forbidden City reservation countdown (3 days left)" and "crowd warning for subway stations near the hotel". The push strategy generation component recognizes that the user is using a tablet with a screen size of 10.9 inches. Based on this, it pushes landscape immersive content, including a map of the Forbidden City, a reservation portal, and the queue time for the roast duck restaurant (the queue time is real-time data). The push time is selected during the user's peak tablet activity time in the past 3 days, which is 21:00-22:00.

[0061] In the interactive feedback module, when a user comments "Very timely, already booked" on the push notification, the sentiment analysis model identifies it as positive (+0.6). The overall feedback value (click +0.3 + comment +0.6 + 40-second dwell time +0.4) = 1.3, increasing the weight of the "Cultural Heritage" tag by 10%. The model iteration component uses federated learning, with 80% of the parameter updates completed locally on the user's tablet. Only the minor 0.05 difference in the "Food" tag weight is encrypted and uploaded.

[0062] In the privacy protection module, the data anonymization component displays the user's mobile phone number as 139****7890, and only retains the location information "Jing'an District, Shanghai"; the permission management component allows users to enable "Get location only when using the APP" and disable "Social relationship association", the system immediately stops collecting social data, and deletes historical social association records after 7 days; the encrypted storage component generates an AES-128 key in the tablet TEE and encrypts the user profile in blocks (basic information block IV=0x123, behavioral feature block IV=0x456).

[0063] The performance comparison data is shown in the table below:

[0064]

[0065] Table 1

[0066] Table 1 shows that this system significantly outperforms traditional systems in all metrics. The push accuracy rate has increased to over 89%, thanks to the multi-level tagging and hybrid similarity matching mechanism of user profiles. In particular, the accuracy rate reached 92% in holiday scenarios in Example 2 due to more precise contextual awareness. The user click-through rate increased nearly threefold, benefiting from optimized push timing (such as the 20:00 active time slot push in Example 1) and content format adaptation (immersive display in landscape mode on tablets). The information timeliness compliance rate exceeded 96%, attributed to the dynamic crawler strategy (high-frequency synchronization of prices and reservation information) and the quality assessment mechanism. The privacy complaint rate was almost zero, reflecting the effectiveness of hierarchical anonymization and access control, and increasing user trust in data security. Overall, the data validates the system's comprehensive advantages in personalization, timeliness, and security.

[0067] refer to Figure 2The figure clearly demonstrates the significant advantage of this system in push notification accuracy. Traditional user profile-less systems rely solely on basic information such as location, achieving an accuracy rate of less than 50%; while basic tagging systems have been improved, their tag dimensions are limited, resulting in an accuracy rate of around 60%. This system, through a multi-level tagging system and context-aware technology, achieves an accuracy rate of 89% in Implementation Example 1, 92% in Implementation Example 2 due to precise adaptation to holiday scenarios, and further improves to 94% in the optimized version after incorporating more user feedback data. The data differences intuitively reflect the synergistic effect of the user profile construction module and the personalized push notification decision-making module, verifying the effectiveness of the three-level tagging structure and hybrid similarity calculation mechanism in improving push notification accuracy.

[0068] refer to Figure 3 This figure reflects the improvement in user interaction quality brought about by this system. In traditional systems, 60% of users choose to ignore push notifications, and deep interactions (favorites, shares, orders) account for less than 13%. This system, through precise push notifications, reduces the ignore rate to 20%, increases the click-through rate to 45%, and achieves a deep interaction rate of 35%. The data changes demonstrate the effectiveness of the interest matching component and the push strategy generation component: precisely matched content reduces the cost of information filtering for users, push formats adapted to different devices improve the reading experience, and the intelligent traffic limiting mechanism reduces user resistance. In particular, the order conversion rate increased from 5% to 15%, directly reflecting the system's auxiliary role in tourism consumption decisions and verifying the actual effect of continuous optimization of the interaction feedback module.

[0069] refer to Figure 4 The chart illustrates the dynamic evolution of user profiles. The "Ancient Town" tag saw its weight increase from 0.5 to 0.8 due to multiple user views and shares; the "Family" tag saw its weight increase from 0.3 to 0.6 within 10 days due to increased searches for family-friendly accommodations; and the "Island" tag, due to lack of interaction, saw its weight decrease according to the formula I(t)=I0×e. (-λt) The calculation (I0 = 0.4, λ = 0.02 / day, t = 15) shows that tags with a weight reduced to 0.05 are removed. The data demonstrates the role of the dynamic profile update module: the short-term interest capture component quickly responds to sudden behaviors (such as the "parenting" tag), the interest decay processing component promptly removes invalid tags (such as the "island" tag), and the long-term feature stability component maintains the weight of core tags (such as the "food" tag, which fluctuates less). This dynamic adjustment ensures that the user profile always aligns with current preferences, providing a reliable foundation for accurate recommendations.

[0070] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A personalized tourism information push system for intelligent mobile terminals based on user profiles, characterized in that, Includes the following modules: The user profile building module includes a basic information collection component, which obtains user information through mobile terminal authorization and uses data cleaning algorithms to remove invalid information; a behavioral data mining component, which captures users' travel behavior on the terminal and uses a sliding window algorithm to count the frequency of behavior; and an interest tag generation component, which maps the mined behavioral data into structured tags, with tag weights dynamically adjusted according to the frequency of behavior. The tourism information processing module includes an information collection component that connects to the platform and official data sources via API interfaces to collect tourism content. The structured processing component transforms unstructured data into a unified format, extracts key information using natural language processing technology, and indexes time-series data by timestamp. The quality assessment component filters out qualified information based on dimensional scoring and puts it into the push pool. The personalized push decision module includes an interest matching component, which calculates the relevance between travel information and user interest tags, prioritizing content with completely identical three-level tags; and a context adaptation component, which adjusts the push content based on the user's real-time context. The push strategy generation component determines the push format and frequency based on user terminal usage habits and network environment. The interactive feedback module includes a behavior tracking component that records user interactions with push notifications and associates the interaction time and path. The feedback analysis component converts user actions into feedback weights, which are used to adjust the weights of corresponding interest tags. The model iteration component updates the user profile every 48 hours based on feedback data, uses an incremental learning algorithm to retain effective parameters of the historical model, and only corrects tags with large deviations. The privacy protection module includes a data anonymization component that anonymizes sensitive user information and retains location information only at the district / county level; a permission management component that uses a fine-grained authorization mechanism, allowing users to individually control various permission switches; and an encrypted storage component that uses the AES-128 algorithm to encrypt user profile data and uses the TLS 1.3 protocol for data transmission.

2. The personalized tourism information push system for intelligent mobile terminals based on user profiles according to claim 1, characterized in that, It also includes a dynamic profile update module and an interest decay processing component. The dynamic profile update module includes a short-term interest capture component and a long-term feature stabilization component. The short-term interest capture component tracks bursting interests through a sliding window, generates temporary labels, and assigns them high initial weights. The long-term feature stabilization component strengthens the weight of stable interest tags that have remained stable for more than 30 days by 20%; the interest decay processing component is used to correct the weight of interest tags that have not been interacted with for a long time, and the calculation formula is I(t) = I0 × e (-λt) Where I(t) is the interest weight at time t, I0 is the initial weight, λ is the decay coefficient, and t is the number of days since the most recent relevant behavior. When I(t) ≤ 0.1, the interest tag is automatically removed.

3. The personalized tourism information push system for intelligent mobile terminals based on user profiles according to claim 1, characterized in that, It also includes a context awareness enhancement module and a push timing optimization component. The context awareness enhancement module includes a user status recognition component and an environmental feature extraction component. The user status recognition component determines the user status through terminal sensors. The environmental feature extraction component collects real-time environmental data; the push timing optimization component calculates the push suitability score, with the formula S=α×C1+β×C2+γ×C3, where S is the push suitability score, C1 is the user's idle time, C2 is the information timeliness, and C3 is the environmental matching degree. Push is triggered when S≥6.

4. The personalized tourism information push system for intelligent mobile terminals based on user profiles according to claim 1, characterized in that, The interest tag generation component of the user profile construction module adopts a multi-level tag system. The core tags include travel type, budget range, and travel mode. The secondary tags are subdivisions of the core tags. The tag weight is calculated based on the depth and frequency of the behavior: search keyword weight × 1.2, browsing time > 60 seconds × 1.5, collection operation × 2.0, completed order × 3.

0. The influence factors of each level of tag are determined by the hierarchical analysis method, and the weight is recalculated every 7 days based on new behavior data. The weight value ranges from 0 to 1.

0.

5. The personalized tourism information push system for intelligent mobile terminals based on user profiles according to claim 1, characterized in that, The tourism information processing module's information collection component employs a combination of dynamic web crawling and API integration: for frequently changing information, it synchronizes every 30 minutes via an encrypted API interface; for medium-frequency changing information, it crawls every 6 hours; for UGC content, it crawls and updates daily between 2-4 AM, using User-Agent rotation and IP proxy pools to circumvent anti-crawling mechanisms, and automatically calls third-party recognition interfaces for pages containing CAPTCHAs; the structured processing component performs three levels of parsing on unstructured data: first-level parsing extracts entities, second-level parsing extracts attributes, and third-level parsing extracts sentiment; for content containing time information, it automatically associates it with the calendar system to generate reminder time nodes; and for multilingual information, it uses a pre-trained neural machine translation model to convert it into the default language set by the user terminal.

6. The personalized tourism information push system for intelligent mobile terminals based on user profiles according to claim 1, characterized in that, The personalized push decision module's interest matching component employs a hybrid similarity calculation mechanism: for text-based information, a 768-dimensional feature vector is extracted using the BERT model, and the cosine similarity with the user's interest tag vector is calculated; for image / video-based information, a 2048-dimensional visual feature vector is extracted using the ResNet50 model, and the Euclidean distance is calculated as the similarity. The matching process consists of two steps: the first step filters information with a similarity ≥ 0.7 into the candidate pool; the second step introduces the user's implicit preferences to adjust the ranking. In addition to basic scenarios, the context adaptation component also dynamically adjusts based on the user's travel stage: if the user is in the planning stage, information containing transportation + accommodation combinations is prioritized; if the user is in the travel stage, destination weather, real-time traffic, and a must-pack list are prioritized; if the user is in the return stage, local souvenirs and return travel guides are pushed. The push strategy generation component adapts the push format according to the terminal model: for devices with a screen size ≥ 6.7 inches, a graphic card containing 3 thumbnails is pushed; for small-screen devices, a text summary + 1 main image is pushed; for tablet devices, landscape immersive push is supported; the push frequency uses intelligent rate limiting.

7. The personalized tourism information push system for intelligent mobile terminals based on user profiles according to claim 1, characterized in that, The feedback analysis component of the interactive feedback module constructs a multi-dimensional feedback system: operation behavior feedback, dwell time feedback, and secondary behavior feedback. For user-inputted text comments, sentiment tendencies are extracted through a sentiment analysis model. The comprehensive feedback value is a weighted sum of feedback from each dimension. When the comprehensive feedback value is ≥1.5, the weight of the corresponding interest tag increases by 10%; when it is ≤-1.0, the weight decreases by 15%. The model iteration component adopts a federated learning framework, where 80% of the model update calculation is completed locally on the terminal. Only the parameter update difference is encrypted and uploaded to the server, aggregated with update data from other users, to generate a global model, which is then distributed to the terminal.

8. The personalized tourism information push system for intelligent mobile terminals based on user profiles according to claim 1, characterized in that, The data desensitization component of the privacy protection module adopts a hierarchical desensitization strategy: partial masking is used for absolutely sensitive information; Relatively sensitive information is blurred. Behavioral data is k-anonymized; the permission management component provides a permission control panel, allowing users to individually enable / disable 12 permissions, each with a clear purpose description; after a permission is disabled, the system immediately stops collecting related data, and historical data already collected is automatically deleted after 7 days and does not participate in subsequent profile calculations; the encrypted storage component uses a hardware security module to generate and store AES-128 encryption keys, and the keys never leave the terminal; user profile data is encrypted and stored in blocks according to "basic information + behavioral characteristics + interest tags", with each block using an independent initialization vector; During data transmission, in addition to TLS 1.3 encryption, sensitive fields are additionally encrypted end-to-end.

9. The personalized tourism information push system for intelligent mobile terminals based on user profiles according to claim 2, characterized in that, The short-term interest capture component of the dynamic profile update module adopts a mechanism that combines sliding window and event triggering: within the 24-hour sliding window, if a user exhibits sudden behavior, a temporary tag is generated. The tag naming format is "temporary-specific interest", with an initial weight of 0.6, and the tag enjoys a 20% ranking priority in push matching. The long-term feature stability component performs trend analysis on stable behaviors that have lasted for more than 30 days; the interest migration identification component calculates the change rate of interest tag weights over three consecutive 7-day windows: current window weight - previous window weight / previous window weight. When the change rate is ≥50% and the correlation between the newly emerging interest tag and the original core tag is <0.3, it is determined to be an interest migration. At this time, a rapid adaptation mechanism is activated: the new interest tag weight is updated every 12 hours within 3 days, and calibration is accelerated based on the new behavior density; the decay coefficient λ of the old core tag is increased by 20%, but the 30-day "transition weight" is retained.

10. The personalized tourism information push system for intelligent mobile terminals based on user profiles according to claim 3, characterized in that, The user state recognition component of the context awareness enhancement module integrates multi-source sensor data: it uses an accelerometer to determine the intensity of motion, a gyroscope to determine changes in the direction of motion, and combines screen illumination frequency and APP usage sequence to comprehensively identify the user state. In addition to basic environmental data, the environmental feature extraction component also integrates with third-party services to obtain special contextual information: it connects with scenic spot reservation systems to obtain real-time reservation occupancy rates, connects with transportation departments to obtain information on road construction and temporary traffic control, and connects with cultural and tourism bureaus to obtain information on special events. The push timing optimization component, based on an S score of ≥6, refines the push scenarios: when the user is "planning their trip," it prioritizes pushing information containing price comparisons and booking links; when the user is "leisurely browsing," it pushes information containing travel guides and experience videos; when the user is "waiting," it pushes information on nearby short-term tours. The alternative recommendation mechanism calculates the multi-dimensional similarity of scenic spots, selects three alternative options with a similarity score ≥0.8, sorts them by priority, and appends them below the original push information, labeled "similar recommendations" and the reasons for the recommendations.

Images