Travel strategy automatic generation method based on artificial intelligence

Abstract

The invention discloses an artificial intelligence-based tourism strategy automatic generation method, which belongs to the technical field of tourism strategy generation, and comprises the following steps: S1, tourism demand analysis, S2, dynamic physical strength and emotion prediction, S3, elastic scenic spot sorting and alternative generation, S4, cross-platform transaction, and S5, preference and strategy optimization. According to the method, a multi-modal model is utilized to analyze hidden information in texts, pictures and voices, for example, demands for barrier-free facilities are recognized from playing with the elderly, meanwhile, physiological data of intelligent equipment are accessed, and factors such as weather are combined to predict daily physical strength and emotion, so that the user experience is improved. The generated strategy is accurately matched with the real demand and the physical state of the user, the stroke discomfort caused by the demand understanding deviation is avoided, and the integrating degree of the playing experience of the user is improved.

Description

[0001] This invention relates to the field of travel guide generation technology, and more specifically, to an artificial intelligence-based method for automatically generating travel guides. Background Technology

[0002] Automatic travel guide generation refers to the use of artificial intelligence technology to automatically generate detailed travel plans and suggestions for travelers, including itinerary arrangements, restaurant recommendations, accommodation options, etc., to help travelers better plan and manage their trips.

[0003] The prior art patent document with authorization announcement number CN119398021B discloses a "method for automatically generating travel guides based on artificial intelligence". By establishing a "vacation-emotion" model, based on the user's basic information and the number of days of travel, the "vacation-emotion" model automatically predicts the user's daily emotional index data during the travel period. After the user enters the city of travel destination, the system automatically pre-screens the attractions in that city and analyzes them in real time to obtain the recommendation level of the attractions. Then, based on the user's daily emotional index data, it matches attractions with corresponding recommendation levels. Based on the matched attractions and nearby hotels, it finally grabs the purchase links for attraction tickets and the booking links for recommended hotels. Using an AI generation model, it generates a text description of the travel itinerary based on the departure date, return date, and target attractions, and generates a paid summary order based on the number of travelers.

[0004] While existing technologies are time-saving, labor-saving, simple, and quick, enabling users to generate travel guides and book attractions and hotels with a single click, they suffer from shortcomings such as superficial semantic parsing, lack of flexible mechanisms, and a single ranking dimension. They rely on a pre-set "holiday-emotion" model and do not use deep learning models to understand the implicit needs of natural language. For example, they cannot parse "attractions requiring barrier-free facilities" from "traveling with elderly people." They also do not have an attraction alternative library. When the preferred attraction is closed due to weather or other reasons, users need to manually adjust their itinerary. For example, they cannot automatically replace indoor attractions on rainy days. Moreover, they only match attractions based on emotion index and do not take into account factors such as real-time visitor flow and transportation convenience, which may lead to congestion and time-consuming recommendations. Summary of the Invention

[0005] This invention provides an artificial intelligence-based method for automatically generating travel guides, which can solve the problems of insufficient semantic parsing, lack of flexible mechanisms, and single sorting method mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for automatically generating travel guides based on artificial intelligence, comprising:

[0007] First, the system receives user travel requests via text, images, or voice input. A multimodal model is used to analyze the implicit information, generating precise interest feature vectors to improve the comprehensiveness of request capture. Next, physiological data from the user's smart device is integrated with weather and travel duration factors to dynamically predict daily physical and emotional states, providing a scientific basis for subsequent itinerary planning and making the travel guide more tailored to the user's actual condition. Then, real-time visitor flow and traffic conditions at attractions are considered, and a ranking algorithm is used to match attractions and provide similar alternatives for each main attraction. When the main attraction's visitor flow exceeds the limit, it is automatically replaced with an alternative attraction, effectively enhancing the trip's resilience and handling unforeseen circumstances. Afterward, ticket and hotel information is scraped from across platforms and compared, with order details pre-filled based on user scenarios, creating a closed-loop service from guide generation to payment, eliminating the need for users to switch between multiple platforms and improving operational convenience. Finally, by analyzing users' historical travel preferences, the recommendation strategy is dynamically optimized, making the travel guide recommendations increasingly accurate with repeated use and continuously improving the user experience.

[0008] Furthermore, the physiological data accessed from the user's smart device is specifically authorized and connected to the wearable device through the OAuth2.0 protocol to obtain the user's heart rate variability, daily steps, and sleep quality physiological indicators in real time, and transmits them to the server for storage and analysis through an encrypted channel to ensure data security and real-time performance.

[0009] Furthermore, when the dynamic model predicts daily physical and emotional states, it uses a sliding window algorithm to smooth physiological data from the past 3-7 days, constructs a decay function based on the number of days of travel, and makes corrections based on the environmental data of the day to improve prediction accuracy.

[0010] Furthermore, when the multimodal model parses the implicit information, it first uses a convolutional neural network to perform scene recognition on the user-uploaded images, extracting the visual features of the attraction type and facilities, and then combines natural language processing to perform sentiment analysis on the keywords in the text and speech, generating a multidimensional interest feature vector.

[0011] Furthermore, when equipping each main attraction with similar alternative attractions, a network of attraction category associations is constructed based on a knowledge graph, and the alternative attractions are filtered according to their real-time ratings and travel time from the main attraction to ensure their suitability and accessibility.

[0012] Furthermore, when crawling ticket and hotel information across platforms, a distributed crawler simultaneously accesses at least three mainstream travel platforms, uses regular expressions to extract price, inventory, and user review information, and generates the optimal price combination scheme through a dynamic programming algorithm.

[0013] Furthermore, when analyzing users' historical travel preferences, a Bayesian personalized ranking algorithm is used to calculate the weights of the types of attractions, hotel classes, and spending amounts in the user's past travels.

[0014] Furthermore, when pre-filling order details based on user scenarios, the rule engine identifies key information in the user's itinerary and pre-fills the payment method according to the user's historical payment habits, reducing the number of steps required for user input.

[0015] Furthermore, when dynamically optimizing the recommendation strategy, an evaluation questionnaire is pushed to the user after each trip. The feedback data is integrated into the preference model through a supervised learning algorithm to achieve incremental updates of the recommendation strategy.

[0016] The beneficial effects of the artificial intelligence-based automatic travel guide generation method of the present invention are as follows:

[0017] By using multimodal intent parsing and dynamic state prediction, this method deeply understands user needs and matches them with actual conditions. It uses multimodal models to analyze implicit information in text, images, and voice, such as identifying the need for barrier-free facilities from "taking the elderly on a trip". At the same time, it integrates physiological data from smart devices with factors such as weather to predict daily physical strength and mood, so that the generated guides accurately match the user's real needs and physical condition, avoiding discomfort caused by misunderstanding of needs, and improving the fit of the user's travel experience.

[0018] By leveraging flexible itinerary generation, intelligent transaction loops, and preference learning, this method enhances the resilience of travel itineraries and optimizes service convenience. It integrates dynamic factors such as real-time pedestrian flow and traffic to match attractions and provides a backup library. In case of unforeseen circumstances at the main attraction, it automatically replaces it. It also enables one-click service by comparing prices across platforms and pre-filling orders. At the same time, it dynamically optimizes recommendations based on historical itineraries, making itineraries more flexible in dealing with unforeseen circumstances such as weather and pedestrian flow. The service process is simpler, and the accuracy of recommendations gradually improves with use, continuously providing users with worry-free and personalized travel solutions. Attached Figure Description

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0020] Figure 1 This is a schematic diagram of the method flow for an automatic travel guide generation method based on artificial intelligence according to the present invention. Detailed Implementation

[0021] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1

[0023] like Figure 1As shown, a technical solution is provided: a method for automatically generating travel guides based on artificial intelligence, including the following steps:

[0024] Step 1: Analysis of Tourism Demand

[0025] First, the system receives users' travel requests via text, images, or voice input. It then uses a multimodal model to analyze the implicit information within these requests, generating accurate interest feature vectors to improve the comprehensiveness of request capture. Specifically, it uses a convolutional neural network (CNN) to perform scene recognition on user-uploaded images, such as extracting visual features like "ancient town stone bridge" and "barrier-free access" from the images. Simultaneously, it combines natural language processing (NLP) to perform sentiment analysis on keywords in the text and voice, such as extracting demand weights like "barrier-free facilities" and "short walking distance" from "traveling with elderly people." Finally, it generates a multidimensional interest feature vector containing attraction type, facility preferences, and sentiment tendencies, further enhancing the comprehensiveness of request capture.

[0026] Convolutional Neural Networks (CNNs) achieve semantic understanding of tourism images through multi-level feature extraction. When processing user-uploaded images, convolutional layers use convolutional kernels of different sizes (such as 3×3 and 5×5) to perform sliding scans on the images, capturing local visual features such as "ancient building eaves" and "barrier-free ramps". Pooling layers retain key information through dimensionality reduction operations and filter out irrelevant backgrounds such as vegetation and crowds. Finally, fully connected layers map multi-level features into feature vectors of attraction types (such as "temples") and facility identifiers (such as "elevator entrances").

[0027] Natural Language Processing (NLP) uses the Transformer architecture to analyze tourism needs in text and speech. For text input, pre-trained models such as BERT first break down "traveling with elderly people requires less walking" into tokens such as "elderly" and "walking". Then, they capture word associations (such as the strong association between "elderly" and "less walking") through a self-attention mechanism to generate semantic vectors containing emotional tendencies (such as "less walking" corresponding to positive needs). For speech input, it is first converted into text by models such as DeepSpeech, and then the demand weights are adjusted by combining prosodic features (such as the intonation changes when emphasizing "elderly"). In addition, by expanding the tourism domain dictionary to include "hiking intensity" and "altitude", the model can accurately identify professional needs, such as extracting altitude restrictions from "requires attractions below 3000 meters in altitude".

[0028] Finally, the Convolutional Neural Network (CNN) and Natural Language Processing (NLP) generate multi-dimensional interest feature vectors through feature fusion. A mapping function projects visual features (such as the "accessibility facilities" vector output by the CNN) and semantic features (such as the "elderly" keyword vector extracted by NLP) into a unified space, and then uses a function:

[0029] F = Wu ·V+W t ·T

[0030] To achieve cross-modal alignment, where W u W represents the weights of visual features. t V represents the weights of the semantic features, V is the visual feature vector extracted by the convolutional neural network (CNN), and T is the semantic feature vector generated by natural language processing (NLP).

[0031] Step 2, Dynamic Physical Strength and Mood Prediction:

[0032] By accessing physiological data from users' smart devices and combining it with factors such as weather and number of days of travel, a dynamic model is used to predict daily physical and emotional states, thereby providing a scientific basis for subsequent itinerary planning and making travel guides more closely aligned with users' actual conditions.

[0033] The physiological data is obtained in real time through authorization and connection with wearable devices (smartwatches, smart bracelets) via the OAuth2.0 protocol to acquire indicators such as heart rate variability (HRV), daily steps, and sleep quality. Simultaneously, it is transmitted to the server via an encrypted channel, and a sliding window algorithm is used to smooth the physiological data from the past 3-7 days to construct a physical fatigue function.

[0034] F(t) = F0 × e -Kt

[0035] Where F0 is the initial physical strength value, K is the decay coefficient, t is the number of days of play, and the mood index is corrected by combining weather data (such as a rainy day correction coefficient of -0.2) and ambient temperature to generate a daily physical strength threshold and mood state curve.

[0036] Data transmission uses the TLS 1.3 encryption protocol to build a secure channel and perform end-to-end encryption on physiological data: the sending end negotiates a session key using an asymmetric encryption algorithm (such as RSA) and implements a periodic session key rotation mechanism (e.g., updating every 24 hours), then uses a symmetric encryption algorithm (such as AES-256) to encrypt the data in batches, and generates a message authentication code (MAC) using a hash algorithm (such as SHA-256) to ensure that the data is not tampered with or eavesdropped on during transmission. The encrypted data is transmitted to the server via the HTTPS protocol. After receiving the data, the server decrypts it using the corresponding key and verifies the data integrity, and finally stores it in an encrypted database.

[0037] Furthermore, the emotional index is dynamically adjusted by combining weather data and ambient temperature. This involves pre-setting emotional impact coefficients for different weather types (e.g., -0.2 for rainy days, +0.1 for sunny days) and constructing a temperature-emotional mapping function.

[0038]

[0039] Where E T The temperature-corrected mood index coefficient ranges from 0.5 (at extreme temperatures) to 1 (at 25°C), where T is the ambient temperature in °C. The mood index decays linearly when the temperature deviates from 25°C. Based on this, a decay factor based on the number of days of travel is introduced to construct a complete mood prediction model.

[0040] E = E0 × (1 + W) t )×E T ×(1-0.1×d)

[0041] Where E0 is the basic sentiment index based on the "holiday-mood" model, and W... t Here, 'd' represents the weather correction factor, and 'd' represents the cumulative number of days of travel (the decay factor increases after 3 days). For example, the mood index on the 3rd day of a 3-day trip is corrected as follows:

[0042] E = E0 × (1 + W) t )×E T ×0.7

[0043] By multiplying the weather correction factor, temperature correction function, and day decay factor, a comprehensive mood index is obtained, enabling dynamic prediction of users' daily mood status.

[0044] Step 3: Flexible attraction sorting and alternative generation:

[0045] Based on dynamic factors such as real-time visitor flow at tourist attractions (obtained from the scenic area API) and traffic conditions (map API), a ranking algorithm is applied:

[0046] S=αI+βE+γP+δR

[0047] The system matches attractions, where I represents interest level, E represents emotional compatibility, P represents physical fitness level, and R represents real-time risk index. α-δ is dynamically adjusted through reinforcement learning. Simultaneously, a network of attraction category associations is constructed based on a knowledge graph (e.g., the "Famous Mountains" category is associated with "Putuo Mountain" and "Jiuhua Mountain"). Each main attraction is equipped with similar alternatives. When the main attraction's visitor capacity exceeds 80%, it is automatically replaced with an alternative attraction, effectively enhancing the trip's resilience and responding to unforeseen circumstances.

[0048] The α-δ weights are dynamically adjusted through reinforcement learning. User feedback on the itinerary (such as satisfaction or evaluation after the actual experience) is used as a reward signal. For example, if a user gives a "very satisfied" rating when the itinerary is recommended according to the current α-δ weights, the reinforcement learning algorithm will increase the reward value of the current weight combination. If the user feedback is poor, such as due to congestion at the attractions, the reward value of the corresponding weight will be reduced. The algorithm continuously iterates and, based on the Markov decision process, dynamically optimizes α (interest weight), β (emotional matching weight), γ (physical fitness weight), and δ (real-time risk index weight) by constantly exploring (trying different α-δ combinations) and utilizing (using high-quality weight combinations). This ensures that the ranking algorithm always fits the user's real needs and changes in the travel scenario, and that the accuracy of the attraction ranking continues to improve with use.

[0049] The knowledge graph-based construction of a network linking scenic spot categories begins by sorting through massive amounts of tourism data and extracting relationships such as "scenic spot-category" and "category-related categories." For example, the "famous mountain" category is linked to "Putuo Mountain" and "Jiuhua Mountain" by analyzing cultural attributes and tourist search associations. During construction, a graph database is used to store nodes (scenic spots, categories) and edges (relationships). Based on a graph traversal algorithm, when the main scenic spot needs to be replaced due to exceeding the visitor limit (reaching 80%), alternative scenic spots belonging to the "famous mountain" category and with suitable real-time status (visitor flow, traffic, etc.) can be quickly selected along the category association edges.

[0050] Step 4, Cross-platform trading:

[0051] By simultaneously accessing at least three mainstream travel platforms (such as Ctrip, Fliggy, and Meituan) through distributed web crawlers, information such as ticket prices, inventory, and user reviews is extracted using regular expressions. Dynamic programming algorithms are then used to generate optimal price combination solutions. For example, when a "Forbidden City + Jingshan Park" package is 15% cheaper on Fliggy than on the official website and includes a tour guide service, it is automatically recommended first. At the same time, a rule engine identifies key itinerary information (such as the date of visit and the number of people), and pre-fills the payment method based on users' historical payment habits (such as 60% choosing Alipay), generating a paid summary order containing links to purchase tickets / book rooms.

[0052] Distributed web crawling technology breaks down the data crawling task of tourism platforms into multiple sub-tasks, which are executed collaboratively by multiple computer nodes. Facing multiple mainstream tourism platforms such as Ctrip, Fliggy, and Meituan, it uses a task allocation mechanism to allow different nodes to access the corresponding platforms and crawl information such as ticket prices, inventory, and user reviews in parallel. Compared with single-machine crawling, it can overcome the performance limitations of a single machine, improve data crawling efficiency and stability, shorten crawling time in large-scale tourism data collection scenarios, ensure the timeliness and comprehensiveness of data acquisition from multiple platforms, and provide sufficient data support for subsequent price comparison and solution generation.

[0053] Regular expressions are tools for matching string patterns, composed of ordinary and special characters. In cross-platform transactions, regular expressions, with their character classes, predefined character classes, quantifiers, boundaries, grouping, and other syntax rules, can accurately match specific formats of content on travel platform pages, such as ticket prices (e.g., ticket prices presented as "Ticket Price: 120 yuan", "Special Offer Tickets from 99.9 yuan", "Ticket ￥150"; to extract the numerical and possibly decimal prices, the regular expressions \d+\.?\d* can be used), inventory figures, and user review text. This allows for efficient extraction of key information from massive amounts of web page data, quickly filtering out the required ticket prices, inventory, and other transaction-related data. This provides a clear and structured data foundation for subsequent dynamic programming algorithm analysis and order generation, simplifying the data processing flow.

[0054] The core of dynamic programming algorithms is to break down the complex problem of generating the optimal combination of tourism products into multiple sub-problems (e.g., defining the state as "the optimal price combination of the i-th attraction on the j-th platform", where i is the attraction index and j is the platform index). By analyzing the dependencies between sub-problems, such as the price superposition of different attraction combinations and the adaptation of preferential policies, the algorithm gradually derives the optimal solution. Taking the "Forbidden City + Jingshan Park" package as an example, the algorithm compares the price, discounts and supporting services of the package on different platforms, and combines the costs of other parts of the trip to select the combination with the lowest total cost and the best service.

[0055] Step 5: Preference and Strategy Optimization:

[0056] The Bayesian Personalized Ranking (BPR) algorithm is used to calculate weights for data such as the type of attractions (e.g., "ancient town" appears 3 times), hotel class (average price of 500 yuan / night), and consumption amount in the user's past trips, and to construct a dynamic preference vector. After each trip, an evaluation questionnaire is pushed to the user. The feedback data (e.g., ratings of a certain attraction) is integrated into the preference model through a supervised learning algorithm to achieve incremental updates of the recommendation strategy. For example, after selecting "historical and cultural" attractions multiple times, the recommendation weight of the same type is automatically increased by 20%.

[0057] The Bayesian Personalized Ranking (BPR) algorithm focuses on uncovering users' implicit preference ranking relationships for items. Based on users' past travel data, such as attraction type, hotel class, and spending amount, the BPR algorithm constructs a "user-item" interaction matrix. Its core idea is to maximize the posterior probability of a user's "preferred item" compared to "unpreferred item," as shown in the formula below:

[0058] p(θ│>u)∝p(>u│θ)p(θ)

[0059] Where θ represents the model parameters, >u represents the preference ranking relationship of user u, and p(>u│θ) uses logistic regression, the function of which is:

[0060]

[0061] Where x u,i Given the feature vectors of user u and item i, parameters are learned using methods such as gradient descent to obtain the user's preference weights for different tourism elements (attractions, hotels, etc.), thus constructing a dynamic preference vector to accurately characterize the user's preferences in tourism choices and provide a basis for subsequent recommendation strategy adjustments.

[0062] Supervised learning algorithms, specifically linear regression algorithms, optimize preference models by learning the mapping relationship between input features (such as attraction attributes and user basic information) and output labels (ratings). Assuming the input X is the number of historical visits and type features of attractions, and the output is the user's rating of the attraction, the formula is:

[0063] y = W T X+b+∈

[0064] Where W is the feature weight vector, b is the bias term, and ∈ is the error term. By minimizing the error between the actual rating and the predicted rating (such as the least squares method), the model parameters W and b are adjusted so that the model can predict the degree of user preference based on the new tourism element features. Then, the feedback data is integrated into the preference model to achieve incremental updates of the recommendation strategy.

[0065] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for automatically generating travel guides based on artificial intelligence, characterized in that: First, the system receives user travel requests via text, images, or voice input. A multimodal model is used to analyze the implicit information, generating precise interest feature vectors to improve the comprehensiveness of request capture. Next, physiological data from the user's smart device is integrated with weather and travel duration factors to dynamically predict daily physical and emotional states, providing a scientific basis for subsequent itinerary planning and making the travel guide more tailored to the user's actual condition. Then, real-time visitor flow and traffic conditions at attractions are considered, and a ranking algorithm is used to match attractions and provide similar alternatives for each main attraction. When the main attraction's visitor flow exceeds the limit, it is automatically replaced with an alternative attraction, effectively enhancing the trip's resilience and handling unforeseen circumstances. Afterward, ticket and hotel information is scraped from across platforms and compared, with order details pre-filled based on user scenarios, creating a closed-loop service from guide generation to payment, eliminating the need for users to switch between multiple platforms and improving operational convenience. Finally, by analyzing users' historical travel preferences, the recommendation strategy is dynamically optimized, making the travel guide recommendations increasingly accurate with repeated use and continuously improving the user experience.

2. The method for automatically generating travel guides based on artificial intelligence according to claim 1, characterized in that: The physiological data accessed by the user's smart device is specifically authorized and connected to the wearable device through the OAuth2.0 protocol. The user's heart rate variability, daily steps, and sleep quality physiological indicators are obtained in real time and transmitted to the server for storage and analysis through an encrypted channel to ensure data security and real-time performance.

3. The method for automatically generating travel guides based on artificial intelligence according to claim 1, characterized in that: When predicting daily physical and emotional states, the dynamic model uses a sliding window algorithm to smooth physiological data from the past 3-7 days, constructs a decay function based on the number of days spent on the trip, and makes corrections based on the environmental data of the day to improve prediction accuracy.

4. The method for automatically generating travel guides based on artificial intelligence according to claim 1, characterized in that: When the multimodal model parses the implicit information, it first uses a convolutional neural network to perform scene recognition on the user-uploaded images, extracting the visual features of the attraction type and facilities, and then combines natural language processing to perform sentiment analysis on the keywords in the text and speech, generating a multidimensional interest feature vector.

5. The method for automatically generating travel guides based on artificial intelligence according to claim 1, characterized in that: When providing alternative attractions of the same type for each main attraction, a network of attraction category associations is constructed based on a knowledge graph, and the alternative attractions are filtered according to their real-time ratings and travel time from the main attraction to ensure the suitability and accessibility of the alternative attractions.

6. The method for automatically generating travel guides based on artificial intelligence according to claim 1, characterized in that: When crawling ticket and hotel information across platforms, a distributed crawler simultaneously accesses at least three mainstream travel platforms, uses regular expressions to extract price, inventory, and user review information, and generates the optimal price combination scheme through a dynamic programming algorithm.

7. The method for automatically generating travel guides based on artificial intelligence according to claim 1, characterized in that: When analyzing users' historical travel preferences, a Bayesian personalized ranking algorithm is used to calculate the weights of the types of attractions, hotel classes, and spending amounts in the user's past travels.

8. The method for automatically generating travel guides based on artificial intelligence according to claim 1, characterized in that: When pre-filling order details based on user scenarios, the rule engine identifies key information in the user's journey and pre-fills the payment method according to the user's historical payment habits, reducing the number of steps the user needs to input.

9. The method for automatically generating travel guides based on artificial intelligence according to claim 1, characterized in that: When dynamically optimizing the recommendation strategy, an evaluation questionnaire is pushed to the user after each trip. The feedback data is integrated into the preference model through a supervised learning algorithm to achieve incremental updates of the recommendation strategy.

Images