Travel route recommendation method based on LLM user portrait and multi-dimensional feature optimization
By constructing user profiles using LLM and BERT models, and combining multidimensional feature optimization with real-time traffic data, the problems of insufficient identification of implicit user preferences and path planning bias in existing technologies are solved, generating efficient and personalized travel routes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-25
- Publication Date
- 2026-05-15
AI Technical Summary
Existing travel route recommendation technologies cannot deeply understand users' implicit preferences, struggle to accurately characterize points of interest, and fail to integrate traffic data in real time during route planning, leading to biased recommendation results.
We use LLM user profiling to deeply analyze user preferences, combine it with BERT model to extract interest point sentiment scores, generate travel routes through multi-dimensional feature optimization and path planning algorithms, and optimize the itinerary using real-time traffic data.
It achieves a deep understanding of user preferences and a detailed description of points of interest, generating efficient and personalized travel route plans that conform to actual traffic conditions.
Smart Images
Figure CN122047677A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of tourism route recommendation technology, and relates to a tourism route recommendation method based on LLM user profiling and multi-dimensional feature optimization. Background Technology
[0002] With the rapid development of the global tourism industry and the increasing demand for personalized travel experiences, intelligent travel route recommendation technology has become crucial for improving travel planning efficiency and experience satisfaction. Traditional recommendation methods typically rely on collaborative filtering, tag-based content filtering, or fixed-rule-based route planning. While these methods can meet basic needs to some extent, their effectiveness is limited by insufficient identification of users' implicit preferences and weak ability to characterize the dynamic features of attractions, making them difficult to adapt to the complex and real-time experience-focused modern tourism scenarios.
[0003] Existing technologies already include travel recommendation solutions that combine user profiles with attraction features. For example, Chinese invention patent CN118780454A discloses a personalized travel route planning system based on big data. This system constructs user profiles using clustering algorithms, extracts textual features of attractions, calculates the similarity between the two to filter attractions, and finally uses the TOPSIS algorithm to optimize the order of attraction visits. While this solution introduces a matching mechanism between user features and attraction features to some extent, it still has two fundamental limitations: First, existing methods rely heavily on clustering or statistics of structured historical behavioral data when constructing user profiles. They cannot perform in-depth semantic analysis of the complex needs expressed by users in language, nor can they infer their implicit preferences, resulting in superficial user profiles. Second, in terms of attraction feature extraction, traditional methods such as keywords, tags, or shallow text vectors are often used, failing to fully utilize the thematic information and emotional tendencies contained in unstructured text. This results in insufficiently refined characterization of points of interest, making it difficult to accurately reflect the user's true perception of those points.
[0004] Second, existing methods rely solely on objective data distribution or user subjective preferences when determining the importance of each feature of a point of interest, failing to organically combine the two. The single-perspective weight determination method cannot balance the universality and rationality of the recommendation results with the personal accuracy. At the same time, route planning is mostly based on static distance estimation and fails to access real-time traffic data, which may result in a large deviation between the generated trip and the actual travel time, affecting the user experience.
[0005] Therefore, there is an urgent need for a recommendation method that can integrate deep semantic understanding, dynamic feature optimization, and multi-objective decision-making in the design and planning stage to improve the intelligence level and user experience satisfaction of tourism recommendation systems. Summary of the Invention
[0006] In view of this, in order to solve the problems mentioned in the background technology, a travel route recommendation method based on LLM user profiling and multi-dimensional feature optimization is proposed.
[0007] The objective of this invention can be achieved through the following technical solution: This invention provides a travel route recommendation method based on LLM user profiling and multi-dimensional feature optimization, including: based on the user's input travel demand text and historical search records, calling LLM to perform deep semantic parsing and implicit preference inference, extracting destination information, and constructing a user profile.
[0008] Based on destination information, we obtain structured basic information and associated unstructured text data of each tourist point of interest within the target area. We then perform thematic analysis on the unstructured text data associated with each point of interest to determine the thematic probability of each point of interest.
[0009] Based on the topic probability of each point of interest, and combined with the structured basic information, a static feature vector for each point of interest is generated.
[0010] Calculate the matching degree between the user profile and the dynamic feature vectors of each point of interest, and filter the candidate point of interest set based on the matching degree.
[0011] Based on the geographical coordinates of each point of interest in the candidate point of interest set, the estimated travel time, and traffic conditions, a travel route plan is generated through a route planning algorithm with the goal of minimizing the total travel time and maximizing the overall experience satisfaction.
[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention uses LLM to perform deep semantic analysis and implicit preference reasoning on user travel demand text and history, and constructs a user profile that integrates explicit and implicit preferences, breaking through the limitations of relying on behavior clustering and being unable to reason about implicit preferences, and realizing a profound characterization of user preferences.
[0013] (2) This invention extracts the topic distribution probability of interest points from the user's unstructured text data, uses the BERT model to extract the user's sentiment score, and uses topic information and sentiment information as a supplement to structured features to jointly constitute the original feature vector of interest points, so that the interest point features can reflect the user's actual experience perception.
[0014] (3) This invention uses the entropy weight method to calculate the objective weight based on the degree of numerical variation of features in the interest point set, which reflects the distinguishability of the features themselves. It obtains the subjective weight of each feature dimension from the user profile, which reflects the user's personalized preferences. The objective weight and subjective weight are linearly combined to obtain the comprehensive weight, so that the recommendation results can reflect the importance of the features and meet the user's personalized needs.
[0015] (4) This invention generates a set of recommended points of interest, obtains real-time travel time by calling a third-party map API, and combines the estimated travel time and matching degree of the points of interest to perform route planning with the goal of minimizing the total travel time and maximizing the overall experience satisfaction. This makes the generated travel route plan more in line with the actual traffic conditions and improves the feasibility of the trip. Attached Figure Description
[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram showing the connections between the steps of the method of the present invention.
[0018] Figure 2 This is a schematic diagram illustrating the user profile construction steps of the present invention.
[0019] Figure 3 This is a schematic diagram showing the connection steps of the travel route plan generation process of the present invention. Detailed Implementation
[0020] 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.
[0021] This invention constructs user profiles that integrate explicit and implicit preferences, generates static feature vectors of points of interest that integrate multi-source semantic information, and performs multi-objective path optimization based on these features to achieve travel route recommendations tailored to user needs during the design and planning phase. Specifically, the method first uses a Large Language Model (LLM) to deeply analyze users' explicit preferences and infer implicit preferences; then, it extracts the topic probabilities of points of interest from users' unstructured text data and uses a BERT model to extract sentiment scores, combining this with structured basic information to form static feature vectors; subsequently, it calculates the matching degree between the user profile and the feature vectors of points of interest, and performs candidate set screening; finally, with the goal of minimizing travel time and maximizing experience satisfaction, it completes the daily allocation of points of interest and itinerary serialization through a path planning algorithm that incorporates real-time traffic information.
[0022] Please see Figure 1As shown, the present invention provides a travel route recommendation method based on LLM user profile and multi-dimensional feature optimization, including the following steps S1 to S5.
[0023] S1. Constructing user profiles aims to deeply understand users' explicit and implicit travel preferences, forming a calculable user preference vector to provide a basis for subsequent point-of-interest matching. Please refer to [link / reference]. Figure 2 As shown, the specific steps include: S1-1, Identifying explicit preference elements: Analyzing travel demand text using LLM to identify and extract the user's explicit preference elements and text descriptions. These explicit preference elements include, but are not limited to: explicitly mentioned interest topics, travel dates and durations, traveler composition, and total budget range.
[0024] S1-2. Reasoning and Quantifying Implicit Preferences: Based on each explicit preference element and historical search records, collaborative reasoning is performed using LLM to generate textual descriptions and implicit weights for each implicit preference element, and to calculate the explicit weights of each explicit preference element. The implicit preference parameters include: preferences for the depth of exploration and level of explanation of points of interest, inferred from the explicit interest topic; preference thresholds for the compactness of daily itineraries and intensity of physical exertion, inferred from the composition and number of travelers; specific sub-disciplines under the explicit interest topic, inferred from the explicit interest topic; potential requirements for convenient and comfortable transportation connections, inferred from the composition of travelers and / or total budget range; and preferences for accommodation conditions and the surrounding environment, inferred from the travel date or duration, composition of travelers, and total budget range.
[0025] Optionally, as a specific implementation, the step of generating text descriptions and implicit weights of each implicit preference element and calculating explicit weights of each explicit preference element can be further refined to include: first, counting the number of explicit preference elements and then allocating initial basic weights to each explicit preference element on an average basis.
[0026] Next, semantic similarity calculation and weight adjustment: For each explicit preference element, the following operations are performed: Text representation: The text description of the explicit preference element and all relevant text fragments extracted from historical search records are converted into real-number vectors using a pre-trained text embedding model (e.g., Sentence-BERT, text-embedding model). This model is a well-known technique in the field of natural language processing; its function is to map the semantic information of text into fixed-dimensional real-number vectors, thereby transforming text into a computable data form. Through this step, explicit preference elements and a set of historical text vectors are obtained.
[0027] Similarity calculation: Calculate the cosine similarity between explicit preference elements and each historical text vector.
[0028] Aggregation and Correction: The average of all cosine similarities is taken as the overall semantic relevance of each explicit preference element. This relevance is then used to correct the initial base weights, for example, through the formula: In the formula, As the initial explicit weights, As the initial base weights, For overall semantic relevance, This is a preset adjustment coefficient, such as 0.5, used to control the strength of the weight correction based on historical correlation.
[0029] Secondly, perform LLM reasoning and quantification of implicit preferences: use explicit preference elements and historical search records as context input to LLM, and construct specific task instruction-type prompts to guide LLM to infer textual descriptions of users' possible implicit preferences that are not explicitly stated (such as potential tendencies in the depth of visits to points of interest, daily itinerary pace, and comfort of transportation connections) based on this context, and assign initial implicit weights to each implicit preference element.
[0030] Specifically, the generation of textual descriptions and initial implicit weights for each implicit preference element through LLM inference is achieved by constructing structured prompts. These prompts typically include the following: role definition (e.g., you are a senior travel planning analyst), task context (containing a summary of the user's explicit needs and historical search records), specific task requirements (e.g., based on the above information, infer the user's potentially unexpressed deep preferences, including at least inclinations towards tour depth, daily itinerary pace, and transportation comfort, and assign an initial importance weight between 0 and 1 to each inferred preference), and output format requirements (e.g., output in JSON format, including two fields: preference description and initial weight).
[0031] Finally, the initial explicit weights and initial implicit weights are normalized to generate the explicit weights of each explicit preference element and the implicit weights of each implicit preference element.
[0032] S1-3, Vectorized Fusion to Construct User Profiles: Based on the text descriptions and explicit weights of each explicit preference element, and the text descriptions and implicit weights of each implicit preference element, user profile vectors are constructed through the following steps: S1-3-1, Element Vectorization: The text descriptions of each explicit preference element and each implicit preference element are converted into corresponding real-number vectors using a pre-trained text embedding model, resulting in explicit preference vectors and implicit preference vectors. This step maps unstructured text preferences to a computable semantic vector space. Optionally, the text embedding model is the same as the model used in step S1-2 to calculate semantic similarity, ensuring that all preference vectors are in the same semantic vector space.
[0033] S1-3-2, Weighting and Concatenation: As a specific implementation method of vectorized fusion, a weighted concatenation method is adopted. First, each explicit preference vector is multiplied by its corresponding explicit weight to obtain a weighted explicit vector. Similarly, each implicit preference vector is multiplied by its implicit weight to obtain a weighted implicit vector. Then, all the weighted explicit and implicit vectors are concatenated in a preset order, for example, all explicit vectors first, then all implicit vectors, to form a comprehensive vector, which serves as the final user profile.
[0034] S2. Determine the topic probability for each point of interest.
[0035] This step aims to mine the topic distribution of interest points from users' unstructured text data, providing topic dimension information for the subsequent construction of static feature vectors. Specifically, it includes the following sub-steps: S2-1, Text Preprocessing: Preprocessing the unstructured text data associated with each interest point, including: using Chinese word segmentation tools for word segmentation, removing stop words, and performing text cleaning to obtain the processed text term set.
[0036] S2-2, Topic Analysis: This step uses the LDA topic model to perform topic analysis on the preprocessed unstructured text. The LDA model is a generative probabilistic model that can represent a document as a mixture of multiple topics.
[0037] The specific implementation method is as follows: First, preset the number of topics, take the set of unstructured text terms of all points of interest as input, train the LDA model, and obtain the term distribution of each topic.
[0038] Then, for each point of interest, the trained LDA model is used to infer the topic label to which each term in the text term set of that point of interest belongs. Based on this, the number of terms belonging to each preset topic in that point of interest is counted. Specifically, for each point of interest, the number of terms assigned to each preset topic by the model is counted out of all its terms.
[0039] It should be noted that the core of this step lies in obtaining the term counts belonging to each topic within each point of interest. The LDA topic model mentioned is only one preferred implementation method. Any method that can classify terms to preset topics and count the number of terms for each topic can be used to implement this step.
[0040] S2-3. Calculate topic distribution probability: For point of interest i, calculate the proportion of the number of terms belonging to each topic to the total number of terms of the point of interest, and obtain the topic distribution probability vector of the point of interest.
[0041] S2-4. Generate topic probability vectors: Use the topic distribution probability vectors as topic probability vectors for each point of interest, which will be used to construct subsequent static feature vectors.
[0042] S3. Generate static feature vectors for each point of interest.
[0043] This step aims to integrate the structured information, sentiment information, and topic information of points of interest to construct a basic vector representation that reflects the characteristics of points of interest, and to determine the final weights of each feature dimension through a combination of subjective and objective weights. Specifically, it includes the following sub-steps: S3-1, Constructing the first feature vector: Organize the structured basic information of each point of interest into a structured table, extract the descriptive features of each point of interest from the structured table, and then convert them into real number vectors as the first feature vector.
[0044] The descriptive features include categorical features and numerical features. For categorical features (such as attraction type, location, etc.), they are mapped to real number vectors according to preset encoding rules (such as one-hot encoding or label encoding). For numerical features (such as ticket price, tour area, altitude, etc.), they are normalized to map feature values of different dimensions to the [0,1] interval to obtain real number vectors. The two types of vectors are concatenated to form the first feature vector.
[0045] S3-2. Constructing the Second Feature Vector: Input the unstructured text data associated with each point of interest into the BERT model. The BERT model has been fine-tuned for sentiment analysis tasks and can output a sentiment score as a real-valued vector. Specifically, all unstructured texts associated with each point of interest are input into the BERT model to obtain the sentiment score for each text. The average value is then taken as the overall sentiment score for that point of interest, and this sentiment score is used as the second feature vector.
[0046] S3-3. Constructing the original feature vectors: Connect the first feature vector, the second feature vector, and the topic probability vector in sequence to form the original feature vectors of each interest point.
[0047] S3-4. Calculating Objective Weights Using the Entropy Weight Method: Based on the original feature vectors of all interest points, the objective weights of each feature dimension are calculated using the entropy weight method. Points of interest and Construct the original feature matrix using 1 feature dimension. ,in Indicates the first The point of interest is at the first The values can be taken in each feature dimension.
[0048] The original feature matrix is normalized to obtain the normalized matrix. The normalization method can be extreme value normalization.
[0049] Calculate the first Information entropy of each feature dimension : ,when When, define Information entropy The range of values is The smaller the value, the greater the difference between points of interest for that feature dimension, and the more information it provides.
[0050] Calculate the first Objective weights of each feature dimension : ,in, For the summation index, from 1 to Traverse all feature dimensions, denominator This represents the sum of information redundancy across all feature dimensions, used to normalize the weights so that the sum of the objective weights of all feature dimensions is 1.
[0051] S3-5. Obtaining Subjective Weights from User Profiles: Based on the feature dimensions corresponding to each preference element in the user profile, obtain the subjective weights of each feature dimension, wherein: the user profile includes multiple preference elements, and each preference element corresponds to at least one feature dimension; for each feature dimension, if its corresponding preference element is an explicit preference element, then the subjective weight of the feature dimension is the explicit weight of the explicit preference element; if its corresponding preference element is a implicit preference element, then the subjective weight of the feature dimension is the implicit weight of the implicit preference element; if a feature dimension corresponds to multiple preference elements simultaneously, then its subjective weight is the weighted sum of the corresponding explicit weights and implicit weights.
[0052] S3-6. Combining to Obtain the Comprehensive Weight: The objective and subjective weights of each feature dimension are linearly combined to obtain the comprehensive weight of each feature dimension; the comprehensive weight of each feature dimension... In the formula For the first Subjective weights for each feature dimension and The preset harmonic coefficient, and This is used to balance the relative importance of objective information and subjective preferences. For example, it can be taken as... , .
[0053] S3-7. Generate static feature vectors: The original feature vectors of each interest point are weighted and summed with their corresponding comprehensive weights to generate static feature vectors for each interest point.
[0054] S4. Filter the candidate set of points of interest.
[0055] This step aims to select a candidate set of points of interest that meet both the user's explicit constraints and have a high degree of matching in terms of implicit preferences. Specifically, it includes the following steps: S4-1, Preliminary screening based on constraints: Based on the constraints parsed from the user's explicit preference elements, and combined with the matching of the opening time of the points of interest with the user's travel date obtained from the descriptive feature set of each point of interest, a preliminary screening is performed on all points of interest in the target area to obtain a set of candidate points of interest.
[0056] Specifically, the constraints mainly include two categories: exclusionary constraints: interest types or specific names explicitly excluded by the user from explicit preference elements; and inclusion constraints: interest types or specific names that the user explicitly requires to be included. During the initial screening, all interest points subject to exclusionary constraints are first filtered out, while ensuring that interest points specified by inclusion constraints are included. Subsequently, it is checked whether the opening hours of the remaining interest points match the user's travel date; for example, if the travel date is Monday, interest points closed on Mondays are excluded.
[0057] S4-2. Calculate the matching degree: Calculate the matching degree between the user profile and the static feature vector of each candidate point of interest, and use it as the matching degree of each candidate point of interest.
[0058] The matching degree specifically employs a multi-index decision-making method based on positive and negative ideal solutions: the user profile vector is taken as the positive ideal solution, and the minimum value of each feature dimension is taken as the negative ideal solution. The Euclidean distance between each point of interest and the positive and negative ideal solutions is calculated separately, and then the relative proximity formula is used. Calculate the matching degree, where, As candidate point of interest numbers, , For the first The matching degree of the candidate points of interest; the higher the value, the more the points of interest match the user's preferences. and They represent the first The Euclidean distances between each candidate point of interest and the positive and negative ideal solutions.
[0059] This invention employs a multi-index decision-making method based on positive and negative ideal solutions to calculate the matching degree between user profiles and static feature vectors of points of interest. The user profile vector is the positive ideal solution, and the minimum feature dimension is the negative ideal solution. By measuring the matching degree through relative proximity, this method more accurately reflects the degree of fit between points of interest and user preferences compared to traditional similarity calculation methods. This further solves the problem of low recommendation accuracy and improves the personalization of recommendation results.
[0060] S4-3. Generate a candidate set: Sort all candidate points of interest in descending order of matching degree to generate a candidate point of interest set.
[0061] S5 generates travel route plans, designed to transform candidate points of interest into time-efficient, high-quality, and actionable daily itineraries. See also... Figure 3 As shown, the specific steps include: S5-1, generating a set of recommended points of interest: based on the expected number of travel days parsed from the user's explicit preference elements, the expected play time and matching degree of each point of interest in the candidate point of interest set, a set of recommended points of interest is selected from the candidate point of interest set that matches the total expected play time with the user's expected number of travel days and maximizes the overall matching degree of the points of interest.
[0062] Furthermore, the process of selecting the recommended set of interest points includes: S5-1-1, calculating the total estimated playtime of all interest points based on the estimated playtime of each interest point in the candidate set of interest points.
[0063] S5-1-2. Based on the user's expected number of travel days and the preset average daily playtime, calculate the total available playtime, where the preset average daily playtime is, for example, 8 hours. This value represents a relatively reasonable daily playtime in terms of physical strength and time allocation when planning itineraries for most tourists.
[0064] S5-1-3. If the total estimated playtime does not exceed the total available playtime, then the set of candidate points of interest will be used as the set of recommended points of interest.
[0065] S5-1-4. If the total estimated playtime exceeds the total available playtime, then with the goal of maximizing the sum of interest point matching degrees, sort the interest points in the candidate interest point set in descending order of matching degree.
[0066] S5-1-5. Based on the sorting results, add points of interest to the recommended point of interest set from front to back. Before adding a point of interest, determine whether the sum of the cumulative play time of the added points of interest and the estimated play time of the current point of interest exceeds the total available play time. If it does not exceed the total available play time, add the current point of interest and update the cumulative play time. If it exceeds the total available play time, stop adding points of interest and use the set of added points of interest as the recommended point of interest set.
[0067] S5-2. Allocate points of interest to travel days: Based on the geographical coordinates of each point of interest in the recommended point of interest set, obtain the real-time travel time by calling a third-party map API, and combine the estimated play time and matching degree of each point of interest. With the goal of minimizing the total travel time and maximizing the overall experience satisfaction of the trip, under the constraint of the opening time of the points of interest, the route planning algorithm allocates each point of interest in the recommended point of interest set to each travel day.
[0068] Specifically, the method for obtaining the real-time travel time is as follows: For any two points of interest in the recommended point of interest set, based on their geographical coordinates, call the path planning interface of a third-party map API (such as Baidu Map API or Gaode Map API), input the latitude and longitude of the starting point and the destination, and obtain the estimated travel time (including travel time and possible congestion delay) at the current time or during the user's travel period, as the basic data for path planning.
[0069] The objective of minimizing the total travel time and maximizing the overall travel experience satisfaction is achieved by constructing a comprehensive optimization objective function. The total travel time is defined as... Overall satisfaction level was (For example, defined as the sum of interest point matching degrees assigned daily). The optimization objective can be transformed into minimizing the objective function. , ,in and These are preset weighting coefficients used to balance the relative importance of time efficiency and experience quality.
[0070] Furthermore, the method of allocating each point of interest in the recommended point of interest set to each travel day through the path planning algorithm includes: S5-2-1, using the geographical coordinates of each point of interest as input, and employing a spatial clustering algorithm to divide them into Q geographical clusters, where Q equals the number of days the user expects to travel.
[0071] S5-2-2 Calculate the sum and variance of the matching degree of all points of interest within the geographic cluster, which are used as the expected total score of daily experience and the balance degree of daily experience for the geographic cluster, respectively.
[0072] S5-2-3. Sequentially check the daily experience balance of each geographic cluster. If it exceeds the preset balance threshold, initiate optimization adjustments until the balance of all clusters meets the standard or the number of optimization adjustments reaches the preset maximum number of iterations. The balance threshold is dynamically set based on the overall distribution of the matching degree of all points of interest in the recommended point of interest set. For example, the balance threshold can be set as a preset proportion of the total variance of the matching degree of all points of interest, with the preset proportion being an adjustment coefficient less than 1, such as 0.5. The preset maximum number of iterations ranges from 20 to 50 times and can be adjusted according to the number of geographic clusters; the more clusters, the greater the number of iterations.
[0073] Furthermore, the optimization adjustment includes a combination of one or more of the following operations: S5-2-3-1, High-score point migration: From the cluster with the highest daily experience balance, select an interest point with the highest matching degree and an appropriate expected play time, and migrate it to the neighboring cluster with the lowest expected total daily experience score.
[0074] S5-2-3-2, Low-score point migration: From the cluster with the highest expected total score for daily experience, select the point of interest with the lowest matching degree and migrate it to another cluster that is geographically closest and whose total play time after migration does not exceed the limit.
[0075] S5-2-3-3, Inter-cluster point swap: Between the two clusters with the highest and lowest expected total score for daily experience, swap two points of interest that are geographically close, have similar play time, but have the greatest difference in matching degree.
[0076] S5-2-4. For each geographic cluster after optimization, calculate the sum of the estimated total playtime for all points of interest (POIs) within it and the internally estimated travel time to obtain the estimated total playtime for each geographic cluster. The internally estimated travel time refers to the total travel time required to visit all POIs within the cluster in a specific visit order. As a specific estimation implementation method, this can be based on the following steps: Estimating travel time between POIs: For any two POIs within a geographic cluster, calculate the road distance based on their geographic coordinates, and divide this distance by a preset average travel speed to obtain the estimated travel time between the POIs. The preset average travel speed is obtained by querying the traffic conditions within the target area and obtaining the average speed of vehicles in the current time period as the preset average speed.
[0077] Calculating the total traffic duration within a cluster: To quickly estimate the internal traffic time of a geographic cluster during the daily allocation phase, an approximate algorithm based on geographic distribution can be used. For example, calculate the average of the pairwise Euclidean distances between all points of interest, multiply it by (number of points of interest - 1) to obtain a distance estimate, and divide the distance estimate by a preset average travel speed to obtain the estimated internal traffic duration of the geographic cluster.
[0078] S5-2-5. Check if there are any geographic clusters whose total expected duration exceeds the maximum available time per day. If so, initiate the time limit adjustment process: remove the point of interest with the lowest matching degree from the geographic cluster that exceeds the limit, and record the point of interest as a point of interest to be migrated.
[0079] The process involves attempting to migrate the points of interest to other geographic clusters. Specifically, this includes: sequentially checking each other geographic cluster in order of distance from the point of interest to be migrated to the center point of each other geographic cluster, from closest to furthest; adding the estimated play time and traffic increment of the point of interest to be migrated to the checked cluster; if the new estimated total duration is less than or equal to the maximum available time per day, then the geographic cluster is the migration geographic cluster, and the point of interest to be migrated is added to it.
[0080] If no suitable geographic cluster for migration is found after traversing all other geographic clusters, the point of interest to be migrated is discarded from the set of recommended points of interest.
[0081] After completing a migration or discard, recalculate the estimated total duration of the original over-limit cluster and the migrated cluster (if migration occurs), and repeat the above time over-limit adjustment steps until the estimated total duration of all geographical clusters does not exceed the maximum daily available time, or the number of adjustments reaches the preset maximum number of iterations.
[0082] During point-of-interest (POI) migration, if all removed POIs fail to migrate to other clusters in a single check, the migration process will be stopped. The system will record the POIs that cannot be accommodated at present and will prompt the user or planner: Under the current time constraints, the selected set of recommended POIs may be too dense, and it is recommended to adjust the maximum available time per day or reduce the number of POIs. At the same time, it is necessary to ensure that at least one POI is retained for each travel day.
[0083] The maximum daily available time refers to the upper limit of the total activity time allowed to be arranged when the system plans an itinerary for a single travel day. This total time includes the estimated time spent visiting all points of interest on that day and the travel time between points of interest, for example, 12 hours.
[0084] S5-2-6. Allocate the final obtained geographic clusters to a travel day to form a daily allocation scheme for recommended points of interest.
[0085] S5-3. Travel Route Plan Generation: Based on all points of interest on each travel day, determine the order of visiting each point of interest on each travel day, generate the travel plan for each travel day, and form a travel route plan accordingly.
[0086] Furthermore, the process of generating travel plans for each travel day includes: S5-3-1, for all points of interest (POIs) within each travel day, generating multiple candidate access sequences using a path planning algorithm (e.g., a genetic algorithm), provided that the opening times of each POI are met. When generating sequences, the path planning algorithm must ensure that each POI is accessible within its opening time window.
[0087] S5-3-2. Calculate the total travel time for each candidate access sequence, and select the candidate access sequence with the shortest total travel time as the access order for each point of interest on each travel day.
[0088] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.
[0089] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0090] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.
[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
[0092] Finally, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A travel route recommendation method based on LLM user profiling and multidimensional feature optimization, characterized in that: The method includes: Based on user-input travel needs text and historical search records, LLM is invoked to perform deep semantic parsing and implicit preference inference, extract destination information, and build user profiles; Based on destination information, we obtain structured basic information and associated unstructured text data of each tourist point of interest within the target area, perform thematic analysis on the unstructured text data associated with each point of interest, and determine the thematic probability of each point of interest. Based on the topic probability of each point of interest, and combined with the structured basic information, a static feature vector of each point of interest is generated. Calculate the matching degree between the user profile and the dynamic feature vectors of each point of interest, and filter the candidate point of interest set based on the matching degree; Based on the geographical coordinates of each point of interest in the candidate point of interest set, the estimated travel time, and traffic conditions, a travel route plan is generated through a route planning algorithm with the goal of minimizing the total travel time and maximizing the overall experience satisfaction.
2. The travel route recommendation method based on LLM user profiling and multidimensional feature optimization according to claim 1, characterized in that: The construction of user profiles includes: By analyzing travel demand text using LLM, we can identify and extract users' explicit preference elements and textual descriptions. Based on each explicit preference element and historical search records, collaborative reasoning is performed through LLM to generate textual descriptions and implicit weights of each implicit preference element, and to calculate the explicit weights of each explicit preference element. User profiles are constructed using vector concatenation methods, based on the textual descriptions and explicit weights of each explicit preference element, and the textual descriptions and implicit weights of each implicit preference element.
3. The travel route recommendation method based on LLM user profiling and multidimensional feature optimization according to claim 2, characterized in that: The process of generating textual descriptions and their implicit weights for each implicit preference element, and calculating the explicit weights for each explicit preference element, includes: The number of explicit preference elements is counted, and an initial basic weight is assigned to each explicit preference element on an average basis. Extract text related to each explicit preference element from historical travel search records, calculate the semantic similarity between the text description of each explicit preference element and the related text, and obtain the initial explicit weight by combining the initial basic weight; Based on explicit preference elements and historical search records, textual descriptions of each implicit preference element are generated through LLM inference, and initial implicit weights are assigned. The initial explicit weights and initial implicit weights are normalized to generate the explicit weights of each explicit preference element and the implicit weights of each implicit preference element.
4. The travel route recommendation method based on LLM user profiling and multidimensional feature optimization according to claim 1, characterized in that: The determination of the topic probability for each point of interest includes: The unstructured text data associated with each point of interest is preprocessed to obtain a processed set of text terms; The number of terms belonging to each preset topic in the text term set of each point of interest is counted, and the proportion of the number of terms of each topic to the total number of terms of the point of interest is calculated to obtain the topic distribution probability vector of the point of interest under the preset number of topics. The topic distribution probability vector is used as the topic probability vector for each point of interest.
5. The travel route recommendation method based on LLM user profiling and multidimensional feature optimization according to claim 4, characterized in that: The generation of static feature vectors for each point of interest includes: The structured basic information of each point of interest is organized into a structured table, and the descriptive features of each point of interest are extracted from the structured table and then transformed into a real number vector as the first feature vector. The unstructured text data associated with each point of interest is input into the BERT model, and the sentiment score is output as a real number vector, which is also used as the second feature vector. The first feature vector, the second feature vector, and the topic probability vector are connected in sequence to form the original feature vector of each point of interest; Based on the original feature vectors of all points of interest, the objective weights of each feature dimension are calculated using the entropy weight method. The subjective weights of each feature dimension are obtained based on the feature dimensions corresponding to each preference element in the user profile. The objective and subjective weights of each feature dimension are linearly combined to obtain the comprehensive weight of each feature dimension. The original feature vectors of each interest point are weighted and summed with their corresponding comprehensive weights to generate the static feature vectors of each interest point.
6. The travel route recommendation method based on LLM user profiling and multidimensional feature optimization according to claim 1, characterized in that: The set of candidate points of interest includes: Based on the constraints parsed from the explicit user preference elements, and combined with the matching between the opening time of the points of interest and the user's travel date obtained from the descriptive feature set of each point of interest, a preliminary screening of all points of interest in the target area is performed to obtain a set of candidate points of interest. Calculate the matching degree between the user profile and the static feature vector of each candidate point of interest, and use it as the matching degree of each candidate point of interest; All candidate points of interest are sorted in descending order of matching degree to generate a set of candidate points of interest.
7. The travel route recommendation method based on LLM user profiling and multidimensional feature optimization according to claim 1, characterized in that: The generated travel route plan includes: Based on the expected number of travel days, the expected play time and matching degree of each interest point in the candidate interest point set, which are parsed from the user's explicit preference elements, a recommended interest point set is selected from the candidate interest point set that matches the total expected play time with the user's expected number of travel days and maximizes the overall matching degree of the interest points; Based on the geographical coordinates of each point of interest in the recommended point of interest set, the real-time travel time is obtained by calling a third-party map API. Combined with the estimated play time and matching degree of each point of interest, with the goal of minimizing the total travel time and maximizing the overall satisfaction of the trip, the route planning algorithm allocates each point of interest in the recommended point of interest set to each travel day under the constraint of the opening time of the points of interest. Based on all points of interest on each travel day, the order in which to visit each point of interest on each travel day is determined, and a travel plan for each travel day is generated, thereby forming a travel route plan.
8. The travel route recommendation method based on LLM user profiling and multidimensional feature optimization according to claim 7, characterized in that: The selected set of recommended points of interest includes: Based on the estimated play time of each interest point in the candidate interest point set, calculate the total estimated play time of all interest points; The total available playtime is calculated based on the user's expected number of travel days and the preset average daily playtime. If the total estimated playtime does not exceed the total available playtime, then the set of candidate points of interest will be used as the set of recommended points of interest. If the total estimated playtime exceeds the total available playtime, then the interest points in the candidate interest point set are sorted from largest to smallest according to the matching degree, with the goal of maximizing the sum of interest point matching degrees. Based on the sorting results, points of interest are added to the recommended point of interest set in order from front to back. Before adding a point of interest, it is determined whether the sum of the cumulative play time of the added points of interest and the estimated play time of the current point of interest exceeds the total available play time. If it does not exceed the total available play time, the current point of interest is added and the cumulative play time is updated. If it exceeds the total available play time, the addition is stopped and the set of added points of interest is used as the recommended point of interest set.
9. The travel route recommendation method based on LLM user profiling and multidimensional feature optimization according to claim 7, characterized in that: The process of allocating each point of interest in the recommended point of interest set to each travel day using a route planning algorithm includes: Using the geographic coordinates of each point of interest as input, a spatial clustering algorithm is employed to divide them into Q geographic clusters; The sum and variance of the matching degree of all points of interest within a geographic cluster are calculated and used as the expected total score of daily experience and the balance of daily experience for the geographic cluster, respectively. Check the daily experience balance of each geographic cluster in turn. If it exceeds the preset balance threshold, start optimization and adjustment until the balance of all clusters reaches the standard or the number of optimization and adjustment reaches the preset maximum number of iterations. For each geographic cluster after optimization and adjustment, calculate the sum of the estimated total play time for all points of interest within it and the estimated internal travel time to obtain the estimated total play time for each geographic cluster. Check if there are any geographic clusters whose total expected duration exceeds the maximum available time per day. If so, remove the point of interest with the lowest matching degree from the over-limit geographic cluster and try to migrate it to other geographic clusters. If it cannot be migrated to any other geographic cluster, discard the point of interest from the recommended point of interest set. Repeat the above steps until the total expected duration of all geographic clusters does not exceed the maximum available time per day or the preset maximum number of iterations is reached. The resulting geographic clusters are then assigned to a travel day to form a daily allocation scheme for recommended points of interest.
10. The travel route recommendation method based on LLM user profiling and multidimensional feature optimization according to claim 7, characterized in that: The process of generating travel plans for each travel day includes: For all points of interest on each travel day, multiple candidate visit sequences are generated using a path planning algorithm, provided that the opening hours of each point of interest are met. Calculate the total travel time for each candidate access sequence, and select the candidate access sequence with the shortest total travel time as the access order for each point of interest on each travel day.