Artificial intelligence-based travel itinerary planning method

Abstract

The application discloses a tourism route planning method based on artificial intelligence and belongs to the technical field of route planning, and specifically comprises the following steps: obtaining historical route sequences of a group and corresponding route evaluation data, and constructing a route-evaluation mapping library; identifying a specific scenic spot based on the mapping library and extracting common behavior deviation characteristic values of a user after visiting the scenic spot; clustering the behavior characteristic values to obtain behavior characteristic clusters, and screening out subsequent scenic spots that bring higher experience feedback to each behavior characteristic cluster through statistical testing to form an association rule library; generating an initial route plan according to a current user request and the rule library; in the execution of the route, determining the deviation category to which the user belongs according to real-time collected user behavior data, querying the rule library according to the category to obtain an experience optimization decision set, and dynamically adjusting the remaining route. The application realizes the change from static resource matching to dynamic experience management, and effectively improves the personalization and satisfaction of tourism route planning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of travel planning, in particular to a travel planning method based on artificial intelligence. BACKGROUND

[0002] In the field of digital tourism, travel planning services based on artificial intelligence have become an important tool for improving the efficiency and personalization of tourists' travel. Existing intelligent travel planning methods usually combine weather conditions with the popularity of scenic spots, user explicit preferences, geographical proximity, and historical access data to achieve static optimization of multi-dimensional environmental adaptation. By incorporating meteorological factors, it can to some extent avoid the direct obstacles and explicit risks caused by weather to the trip, such as recommending outdoor sightseeing on sunny days or guiding users to indoor venues on rainy and snowy days, thereby improving the basic feasibility and safety of the trip. While meeting the user's interest preferences and geographical location efficiency, it also meets the environmental conditions of the day, making the recommended results more comprehensive in multiple static dimensions and improving the user's immediate activity selection experience in specific weather conditions.

[0003] However, such mainstream methods generally have a significant technical limitation: their planning logic is essentially static and based on resource tag matching, which seriously ignores the dynamic characteristics of tourism activities as a high-intensity physical and cognitive load experience. Specifically, existing technical solutions lack the ability to model and respond to the key factor of "cumulative physiological fatigue of sightseeing activities" and the resulting "degradation of subsequent experience quality". It is impossible to predict or perceive how the significant changes in a user's physiological state (such as fatigue level) after completing a high-intensity scenic spot (such as a mountainous scenic area that requires long-distance hiking) will fundamentally affect their experience satisfaction with subsequent activities.

[0004] Current solutions mostly stay at the level of post-feedback or simple filtering. For example, some allow users to manually set the pace of travel (such as "easy" or "intense"), but this is a static, subjective a priori setting that cannot adapt to the real-time changes in physical state during the trip. Another solution monitors heart rate and other indicators through wearable devices, but its application is mostly limited to health warnings or exercise analysis, and does not deeply and intelligently couple with the content of the trip (such as the type and intensity of the next scenic spot) and dynamically reorganize it. More importantly, existing technologies lack the ability to use massive amounts of anonymous historical data to scientifically analyze and verify "what kind of fatigue state should choose what kind of subsequent activity to obtain a better experience". They cannot answer the core question: when many tourists have leg pain after visiting the Huangshan Mountain, historical data shows that they went to the Hongcun Ancient Town or the indoor museum the next day to ultimately achieve higher overall satisfaction?

[0005] Therefore, the prior art leads to a common user experience dilemma: a recommended route may be optimal in terms of sightseeing popularity and geographical logic, but the user experiences a decline in subsequent travel due to fatigue accumulation in actual execution, forming a contradiction between "planning seems perfect and experience is actually poor". This is due to the prior art, which breaks the closed-loop data correlation between "travel planning", "behavior state perception" and "experience result feedback" in the technical architecture. SUMMARY

[0006] The purpose of the present application is to provide an artificial intelligence-based travel itinerary planning method to solve the following technical problems:

[0007] The intelligent itinerary planning method based on static interest matching and popularity sorting in the prior art cannot dynamically adapt to the real-time physiological fatigue state of the user during the trip, leading to a mismatch between subsequent recommendations and the user's actual experience ability, causing the overall quality of the travel experience to decline.

[0008] The purpose of the present application can be achieved by the following technical solutions:

[0009] The artificial intelligence-based travel itinerary planning method comprises the following steps:

[0010] S1, obtaining historical travel sequences of an anonymous user group in a target area and corresponding travel evaluation data, and constructing a travel-evaluation mapping library;

[0011] S2, performing group behavior analysis on the travel-evaluation mapping library, identifying common behavior deviation characteristic values after visiting a specific sightseeing spot, clustering sightseeing spot visit records after the specific sightseeing spot in the historical travel sequences based on the common behavior deviation characteristic values, and forming an association rule library of behavior deviation characteristic values and subsequent sightseeing spot selection clusters;

[0012] S3, generating an initial travel plan containing a sightseeing spot visit sequence based on the association rule library in response to a planning request of a current user;

[0013] S4, obtaining a positioning sequence and a stay duration sequence of the current user in real time and calculating a behavior feature vector during execution of the initial travel plan by the current user, inputting the behavior feature vector into a preset behavior deviation classification model, and outputting a deviation category to which the real-time behavior of the current user belongs;

[0014] S5, obtaining a corresponding target sightseeing spot selection cluster from the association rule library according to the deviation category, dynamically adjusting the remaining travel plan of the current user based on the target sightseeing spot selection cluster, and synchronizing to the terminal device of the current user.

[0015] As a further scheme of the present application: in S1, the specific construction process of the travel-evaluation mapping library is:

[0016] obtain a historical travel record set and a historical evaluation record set of an anonymous user group; parse a scenic spot visiting sequence indexed by a travel identifier from the historical travel record set, the scenic spot visiting sequence containing scenic spot identifiers arranged in time sequence;

[0017] parse evaluation data corresponding to each travel identifier from the historical evaluation record set, the evaluation data containing numerical overall score data and text comment data; perform sentiment analysis on the text comment data to extract sentiment labels and intensity values representing experience tendencies, and generate a sentiment vector;

[0018] splice and combine the numerical overall score data and the sentiment vector to form an experience feedback vector corresponding to the travel identifier; bind the scenic spot visiting sequence and the experience feedback vector corresponding to the travel identifier with the travel identifier as an associated key, and generate a travel-evaluation mapping record; and collect all travel-evaluation mapping records corresponding to all travel identifiers to generate a travel-evaluation mapping library.

[0019] As a further scheme of the present application, in S2, the specific obtaining process of the specific scenic spot is:

[0020] extract all scenic spot visiting sequences from the travel-evaluation mapping library, and collect scenic spot identifiers therein to form an initial scenic spot set; for each scenic spot identifier in the initial scenic spot set, retrieve all travel-evaluation mapping records containing the identifier from the travel-evaluation mapping library to form an associated record set;

[0021] for each record in the associated record set, locate the visiting time point of the scenic spot identifier in the scenic spot visiting sequence, and divide the sequence into a preceding sub-sequence and a subsequent sub-sequence according to the time point; obtain local experience feedback averages corresponding to the preceding sub-sequence and the subsequent sub-sequence respectively, and calculate the difference between the two as the individual experience feedback attenuation degree of the record;

[0022] statistically analyze the individual experience feedback attenuation degrees of all records in the associated record set to obtain the group experience feedback attenuation degree of the scenic spot identifier and its statistical distribution parameters; after traversing the initial scenic spot set, filter based on the group experience feedback attenuation degree and its statistical distribution parameters of each scenic spot identifier through statistical significance test;

[0023] mark the scenic spot identifier with a negative group experience feedback attenuation degree and a statistical significance exceeding a preset threshold as a specific scenic spot to form a specific scenic spot set.

[0024] As a further scheme of the present application, in S2, the specific process of the common behavior deviation feature value is:

[0025] For each specific scenic spot in the specific scenic spot set, a corresponding set of associated records is extracted from the trip-evaluation mapping library; records with a negative individual experience feedback attenuation degree and an absolute value exceeding a first preset threshold are filtered from the set of associated records to form a high-attenuation record subset;

[0026] For each record in the high-attenuation record subset, trajectory data within a fixed time window from the access time point of the specific scenic spot in the scenic spot access sequence of the record is intercepted; based on the trajectory data, a moving speed change rate and a stay event frequency change rate of the record within the fixed time window compared to a control time window before the access time point are calculated;

[0027] The moving speed change rate and the stay event frequency change rate of all records in the high-attenuation record subset are collected, and the arithmetic mean and the standard deviation of the two are calculated respectively; the combination of the arithmetic mean and the standard deviation of the moving speed change rate and the arithmetic mean and the standard deviation of the stay event frequency change rate are taken as the common behavior deviation characteristic value corresponding to the specific scenic spot.

[0028] As a further scheme of the application: in S2, the specific construction process of the association rule library is:

[0029] For each specific scenic spot in the specific scenic spot set, a corresponding set of associated records is extracted from the trip-evaluation mapping library; records with a negative individual experience feedback attenuation degree and an absolute value exceeding a first preset threshold are filtered from the set of associated records to form a high-attenuation record subset;

[0030] For each record in the high-attenuation record subset, the identifier of the first subsequent scenic spot after the specific scenic spot in the scenic spot access sequence of the record and the corresponding experience feedback vector are extracted;

[0031] The subsequent scenic spot identifier and the experience feedback vector are assigned to the corresponding behavior characteristic cluster; for each behavior characteristic cluster, the records within it are grouped according to the subsequent scenic spot identifier to form multiple candidate option subsets;

[0032] The mean of the experience feedback vectors corresponding to all records in each candidate option subset is calculated; the experience feedback vector means of different candidate option subsets within the same behavior characteristic cluster are subjected to statistical significance test;

[0033] In the test results, the subsequent scenic spot identifier corresponding to the candidate option subset with an experience feedback vector mean significantly higher than other options is extracted as the experience optimization decision set of the behavior characteristic cluster;

[0034] establishing a mapping relationship between the behavior feature cluster of the specific scenic spot and the experience optimization decision set, and integrating all mapping relationships of all specific scenic spots to construct the association rule library.

[0035] As a further scheme of the present application, in the S3, the specific process of the initial travel planning is:

[0036] parsing a current user request to obtain a target scenic spot identification list and generating a to-be-planned scenic spot set, matching the to-be-planned scenic spot set with the specific scenic spot set, and marking an identification belonging to the specific scenic spot set as a to-be-optimized node;

[0037] For each to-be-optimized node, querying a mapped experience optimization decision set from the association rule library, extracting a scenic spot identification contained in the set, adding the extracted scenic spot identification to the to-be-planned scenic spot set to form an expanded scenic spot set;

[0038] generating an initial scenic spot access sequence based on the expanded scenic spot set and space-time constraint data between scenic spots, and traversing the initial scenic spot access sequence to identify a to-be-optimized node arranged as a continuous access node;

[0039] When a continuous to-be-optimized node is identified, obtaining an experience optimization decision set mapped by a first to-be-optimized node from the association rule library;

[0040] judging whether a second to-be-optimized node belongs to a scenic spot identification in the experience optimization decision set, and if not, selecting a scenic spot identification from the experience optimization decision set to replace the second to-be-optimized node in the sequence or insert it before the second to-be-optimized node;

[0041] combining the adjusted scenic spot access sequence with traffic and time consumption data to generate the initial travel planning.

[0042] As a further scheme of the present application, in the S4, the specific calculation process of the behavior feature vector is:

[0043] obtaining a real-time positioning sequence and a stay duration sequence uploaded by a current user mobile terminal, extracting a leaving time point of a scenic spot most recently visited by the current user from the real-time positioning sequence, and taking the time point as a starting point to cut a subsequent real-time positioning sequence in a fixed time window;

[0044] calculating a moving speed sequence of the fixed time window based on the subsequent real-time positioning sequence in the fixed time window, and calculating a change rate of the moving speed sequence compared with a moving speed sequence in a control time window before the leaving time point;

[0045] statistically count the frequency of the stay events in the fixed time window based on the stay duration sequence, and calculate the change rate of the frequency compared with the stay event frequency in the control time window;

[0046] combine the change rate of the moving speed sequence and the change rate of the stay event frequency to generate a behavior feature vector of the current user.

[0047] As a further scheme of the present application: in S4, the specific construction process of the preset behavior deviation classification model is:

[0048] extract the clustering center feature vector of each behavior feature cluster corresponding to each specific scenic spot from the association rule base; take each clustering center feature vector as a model training sample, and take the identification of the behavior feature cluster from which the clustering center feature vector is derived as the class label of the training sample;

[0049] collect all the training samples and class labels corresponding to all the specific scenic spots to construct a labeled training sample set; and use the labeled training sample set to supervise the training of a multi-classification machine learning model with a preset basic structure;

[0050] save the multi-classification machine learning model that has completed training and reached the preset precision index as the behavior deviation classification model.

[0051] As a further scheme of the present application: in S5, the specific process of dynamically adjusting the remaining travel plan of the current user is:

[0052] obtain the deviation category to which the real-time behavior of the current user belongs from the behavior deviation classification model; and obtain the corresponding experience optimization decision set from the association rule base according to the deviation category;

[0053] obtain the real-time geographic location information of the current user and the sequence of scenic spots to be visited in the remaining travel plan; generate a subsequent travel segment with the real-time geographic location as the starting point and the experience optimization decision set as the candidate destination constraint, in combination with real-time traffic data and scenic spot opening time data;

[0054] fuse and compare the generated subsequent travel segment with the original scenic spot sequence in the remaining travel plan that has not been executed, and replace the scenic spot identification that does not belong to the experience optimization decision set with the scenic spot identification that belongs to the experience optimization decision set in the subsequent travel segment when the scenic spot identification in the original scenic spot sequence does not belong to the experience optimization decision set;

[0055] combine the new travel plan sequence after the fusion and replacement processing with the real-time updated traffic moving mode and the predicted time consumption data to obtain the adjusted dynamic travel plan.

[0056] Advantages of the present application:

[0057] 1) By establishing a verifiable correlation between behavior patterns and experience results, it realizes the fundamental change of the trip planning from static resource matching to dynamic experience management. It can be understood that the traditional method relies on fixed scenic spot tags and user interests for combination, and its essence is to regard tourism as a static resource scheduling problem. The present application breaks through this limitation, and its core principle lies in building a data analysis framework that can interpret the "experience traces" of the group. First, the objective behavior trajectory of tourists (such as moving speed, stay mode) is accurately correlated and quantitatively fused with their subjective experience feedback (score, emotional tendency) to form "experience-behavior" coupled data. Further, using statistical hypothesis testing methods, those scenic spots that will lead to significant attenuation of subsequent experience are identified from massive data, and the common behavior deviation characteristics of tourists after these scenic spots are accurately described. Further, within the same behavior pattern group, through rigorous statistical comparison, the subsequent activity options that can bring significantly higher experience evaluation to the group are screened out. This makes the final decision rule not based on subjective guess or simple frequency, but derived from the empirical findings of the "behavior-selection-result" causal chain. Therefore, the planning scheme generated by this method has its core being the optimized strategy verified by the historical group experience, realizing the paradigm shift from arranging "scenic spots" to managing "experience flow".

[0058] 2) Realize the prospective avoidance of potential experience risks, and generate personalized initial schemes with inherent resilience. The present application injects the experience risk patterns and optimization strategies mined from history into the generation logic of new trips in reverse. When arranging the initial route for the user, it actively identifies whether there are known high-intensity risk points in the draft of the trip, and predicts the behavior state that the user may enter after visiting the point. Based on this prediction, the experience optimization decision set bound to this state and verified by history in the associated rule library is immediately called to automatically reconstruct the subsequent trip— for example, inserting adaptive buffer items between high-intensity activities, or replacing them with more suitable scenic spots. The essence of this process is to convert the "lessons learned" of countless tourists in history into a set of preset optimization algorithms, providing avoidance solutions in advance before the user has personally experienced fatigue. Therefore, the initial planning delivered at the end is a "resilient scheme" that has built-in anti-risk structure, fundamentally improving the scientificity and reliability of the plan, and significantly reducing the probability of experience collapse in the later stage of the trip due to improper planning.

[0059] 3) Constructed the "real-time perception-state diagnosis-dynamic optimization" online intelligent closed loop, and provided the ability of sustainable evolution and personalized service. The application extends the trip planning from one-time pre-trip service to a full-process intelligent agent covering pre-travel and in-travel. It can be understood that the application collects the behavior data of the user in real time through the mobile terminal, and uses the pre-trained classification model to quickly match it with the typical behavior pattern library deposited in history, so as to realize the objective diagnosis of the real-time fatigue state of the user. Then, according to the diagnosed state category, the corresponding optimization decision set is retrieved and called, and combined with real-time traffic, location and other information, an immediate trip adjustment scheme is generated and pushed to the user. The core value of this closed loop is not only to realize the "dynamic individualization", but also to build a self-evolution channel. Each real-time diagnosis, user feedback after adopting the suggestion, will be used as new "experience-behavior" association data, and will be fed back to the database after anonymization, for regular retraining and verification of the rule base and classification model. This makes the decision knowledge continuously updated and evolved, forming a virtuous cycle of "data-driven optimization, optimization generating new data", and continuously enhancing the precision and adaptability of its service. BRIEF DESCRIPTION OF DRAWINGS

[0060] The application will be further described below with reference to the drawings.

[0061] Figure 1 It is a flowchart of the tourism trip planning method based on artificial intelligence of the application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0063] Please refer to Figure 1 The application is a tourism trip planning method based on artificial intelligence, which comprises the following steps:

[0064] S1, obtaining the historical trip sequence and corresponding trip evaluation data of an anonymous user group in a target area, and constructing a trip-evaluation mapping library;

[0065] First, the application program interface is used to call the background database of the tourism service platform to obtain all anonymous user itinerary records in a target geographic area (such as Huangshan City, Anhui Province) in a past statistical period (such as 24 months). Each record is indexed by a unique itinerary identifier and includes a sequence of point-of-interest visits sorted by timestamp, such as "Itinerary ID-1001: [Timestamp1: Hongcun, Timestamp2: Huangshan Scenic Area, Timestamp3: Xidi]". At the same time, the evaluation data corresponding to the above itinerary identifier is retrieved from the evaluation database, which includes a numerical overall rating (such as 4.5 / 5.0) and a text comment (such as "Climbing Huangshan made my legs weak, but the scenery was breathtaking").

[0066] Subsequently, the text comment is subjected to automated sentiment analysis. A pre-trained natural language processing model is used to identify and extract keywords and their sentiment intensity from the comment, such as identifying "weak legs" as a negative physical state label with an intensity value of 0.8 (range 0-1) and "breathtaking" as a positive landscape evaluation label with an intensity value of 0.9. These discrete labels and intensity values are converted into a multi-dimensional sentiment vector.

[0067] Finally, the numerical overall rating and the generated sentiment vector are concatenated and normalized to form a comprehensive experience feedback vector. The itinerary identifier is used as a unique key to bind the sequence of point-of-interest visits to the experience feedback vector for that itinerary, generating a complete itinerary-evaluation mapping record. By collecting a large number of such records, a journey-evaluation mapping library is constructed for deep analysis.

[0068] S2, group behavior analysis is performed on the journey-evaluation mapping library to identify common behavior deviation characteristic values after visiting a specific point of interest, and based on the common behavior deviation characteristic values, the point-of-interest visit records after the specific point of interest in the historical itinerary sequence are clustered to form an association rule library of behavior deviation characteristic values and subsequent point-of-interest selection clusters.

[0069] First, the specific point of interest is identified by traversing the journey-evaluation mapping library. For each point of interest, the following calculations are performed: for all itinerary records containing the point of interest, the visit time of the point of interest is used as a dividing point to divide the itinerary into a "presequence" and a "postsequence". The mean values of the experience feedback vectors corresponding to the presequence and postsequence are calculated, and their difference is calculated to obtain the "individual experience feedback attenuation degree" of the point of interest in a single itinerary. Next, statistical analysis (such as calculating the mean and variance) is performed on the attenuation degrees of all related itineraries, and t-test or other methods are used to determine whether the attenuation is statistically significant (p-value less than 0.05). Points of interest that cause statistically significant negative attenuation of experience feedback are marked as specific points of interest (i.e., high experience risk points), such as "Huangshan Scenic Area".

[0070] Secondly, the common behavior shift feature values are calculated. For each specific scenic spot, a subset of high-decay records is selected, which has the most significant "individual experience feedback decay degree". For each record in the subset, the GPS trajectory data of the user within a fixed time window (such as the next 24 hours) after leaving the specific scenic spot is extracted. By calculating the average moving speed, the number of stops per unit distance, and other indicators within the time window, and comparing them with the baseline values of the previous control time window before visiting the scenic spot, the behavior shift indicators such as "moving speed change rate" and "stop frequency change rate" are obtained. The mean and standard deviation of these indicators are aggregated in all records to calculate the common behavior shift feature values that describe the group behavior pattern after the specific scenic spot, such as a vector (speed decay rate: -30% ± 5%, stop growth rate: +50% ± 10%).

[0071] Subsequently, the construction of the association rule base is carried out. Based on the calculated common behavior shift feature values, clustering analysis (such as using the K-means algorithm) is performed on the records in the high-decay record subset, and users with similar behavior patterns are grouped into different behavior feature clusters, such as "Cluster A: significant deceleration and frequent rest" and "Cluster B: moderate deceleration".

[0072] For each behavior feature cluster, the next scenic spot that is visited immediately after the specific scenic spot by all records in the cluster is analyzed. Then, the key step is to compare whether there is a statistically significant difference in the mean value of the experience feedback vector of the sub-population that selects different subsequent scenic spots (such as "Hongcun" or "Huizhou Ancient City"). Through hypothesis testing, the subsequent scenic spots that can bring statistically significantly higher experience feedback for the behavior feature cluster are selected and classified as the experience optimization decision set for the cluster.

[0073] Finally, mapping rules such as "specific scenic spot (Huangshan) + behavior feature cluster (Cluster A) → experience optimization decision set ({Huizhou Ancient City, Hot Spring Hotel})" are established, and all rules are integrated to form the association rule base.

[0074] S3, in response to the planning request of the current user, and based on the association rule base, an initial itinerary plan containing a sequence of scenic spots is generated;

[0075] When receiving the target scenic spot list submitted by the current user (such as "Huangshan, Hongcun, Xidi"), first, it is compared with the specific scenic spot set to identify the high-risk points as optimization nodes (such as "Huangshan"). According to the optimization node, the association rule base is queried to pre-obtain the candidate scenic spots (such as "Huizhou Ancient City") in the experience optimization decision set, and add them to the candidate list to form an expanded scenic spot set.

[0076] Then, based on the extended set of attractions, combined with the distance between attractions, opening hours, regular tour length, etc. Temporal constraints, use path planning algorithm (such as considering the traveling salesman problem algorithm) to generate a preliminary sequence of attraction visits.

[0077] After generation, the sequence will be checked for "risk simulation": scan the sequence to see if there are consecutive arrangements of nodes to be optimized (such as "Huangshan" followed by "Hongcun"). If so, call the experience optimization decision set corresponding to the first node (Huangshan) in the rule base, check if the second node (Hongcun) is in it. If not, automatically select a scenic spot (such as Huizhou Ancient City) from the experience optimization decision set, insert it between the two high-risk points as a buffer, or replace the second node, so as to generate an initial itinerary plan that has embedded fatigue risk avoidance strategies.

[0078] S4, in the process of the current user executing the initial itinerary plan, real-time acquisition of its positioning sequence and stay time sequence and calculation of behavior feature vector, input of the behavior feature vector into the preset behavior offset classification model, output of the current user's real-time behavior belonging to the offset category;

[0079] In the itinerary execution phase, the user-authorized mobile terminal application continuously collects encrypted GPS positioning point sequence and stay time at each location.

[0080] When the user ends the tour of a specific attraction (such as Huangshan), take the departure time as the starting point, and intercept the positioning and stay data in the subsequent fixed time window (such as 6 hours). Based on this data, calculate the average moving speed, stay event frequency and other indicators in the time window, and compare them with the baseline values of the user's tour of the attraction in the previous control time window, and calculate the "speed change rate" and "stay frequency change rate" in real time, so as to construct the behavior feature vector describing the current state.

[0081] Prior to this, a multi-class behavior offset classification model (such as a support vector machine or neural network classifier) has been trained based on the clustering center vectors of all behavior feature clusters in the association rule base. At this time, the real-time calculated behavior feature vector is input into the model, which will output which historical behavior feature cluster (such as "Cluster A: significant speed reduction and frequent rest") it best matches, i.e. determine the user's real-time behavior offset category.

[0082] S5, according to the offset category, obtain the corresponding target attraction selection cluster from the association rule base, and dynamically adjust the remaining itinerary plan of the current user based on the target attraction selection cluster, and synchronize to the terminal device of the current user.

[0083] According to the offset category determined by S4, the experience optimization decision set mapped by the category (such as {Hui Zhou Ancient City, Hot Spring Hotel} for "Cluster A") is immediately queried from the associated rule base. At the same time, the real-time geographical position of the user and the remaining itinerary planning not yet performed are obtained. Taking the current position of the user as the starting point and the scenic spots in the experience optimization decision set as candidate destinations, one or more optimal "subsequent itinerary segments" are re-planned and generated in combination with real-time traffic conditions, scenic spot opening status and other information.

[0084] The newly generated "subsequent itinerary segment" is compared with the original plan. If the scenic spot immediately to be visited in the original plan is not in the experience optimization decision set, the scenic spot in the new segment is used to replace or insert. For example, the original plan is to visit "Hongcun" tomorrow, but according to the real-time state determination and rule base, it is recommended to visit "Hui Zhou Ancient City". Finally, the newly adjusted itinerary is pushed to the terminal device interface of the user in real time, together with the updated traffic guide and time suggestion, to complete a personalized itinerary optimization based on real-time physiological state feedback.

[0085] In a preferred embodiment of the present application, the specific construction process of the itinerary-evaluation mapping library in S1 is as follows:

[0086] The historical itinerary record set and the historical evaluation record set of an anonymous user group are obtained. The scenic spot visit sequence indexed by the itinerary identifier is parsed from the historical itinerary record set, and the scenic spot visit sequence contains the scenic spot identifier arranged in chronological order.

[0087] The evaluation data corresponding to each itinerary identifier is parsed from the historical evaluation record set, and the evaluation data contains numerical overall score data and text comment data. The text comment data is subjected to sentiment analysis processing to extract sentiment labels and intensity values representing experience tendencies, and a sentiment vector is generated.

[0088] The numerical overall score data and the sentiment vector are spliced and combined to form an experience feedback vector corresponding to the itinerary identifier. The scenic spot visit sequence and the experience feedback vector corresponding to the itinerary identifier are bound to generate an itinerary-evaluation mapping record with the itinerary identifier as the association key. All itinerary-evaluation mapping records corresponding to the itinerary identifier are collected to generate an itinerary-evaluation mapping library.

[0089] First, two basic data sets need to be obtained: one is the journey record set recording the past journey footprints of countless tourists, and the other is the evaluation record set storing their feelings about these journeys. This is done because only by putting objective behavior and subjective feedback into the same analysis framework can the rules behind the experience be understood. For example, from the journey record, a series of records indexed by a unique identifier such as "Journey ID-1001" are parsed, and each record contains a series of accurately arranged scenic spot identifiers in chronological order, such as "08:00 ticket office in Hongcun, 10:30 Huangshan Cloud Valley cableway, 14:00 Welcoming Pine … …", which restores the tourist's movement track in the space and time dimension. At the same time, from the evaluation record, the evaluation data corresponding to the same "Journey ID-1001" is located, which includes not only the overall numerical score such as "4.5 points", but also text comments such as "Huangshan scenery is breathtaking but climbing is really too tired, legs are not their own". The reason why it is necessary to parse and associate the two is that a single score cannot reveal the specific reason for "tiredness", and isolated comments are also detached from the specific journey context. Combining the two can lay the foundation for subsequent analysis.

[0090] Then, the unstructured text comments need to be deeply interpreted, i.e. sentiment analysis processing. The principle is to let the computer understand the emotional color and evaluation object in human language through natural language processing technology. For example, analyzing the above comment can identify that "shocking" is a positive emotion and the associated object is "scenery"; "tired" is a negative physiological feeling and the associated object is "climbing". Further, these identified emotions are given intensity values, and these scattered "emotion-object-intensity" information is integrated and coded into a multi-dimensional emotion vector. The purpose of this is to convert abstract, difficult-to-directly-calculate subjective feelings into structured, quantifiable digital features, so that "experience" becomes measurable and analyzable.

[0091] Then, the numerical score representing the overall satisfaction is spliced and fused with the above-mentioned emotional vector depicting the detailed feelings to form a comprehensive and multi-dimensional experience feedback vector. This vector combines the overall satisfaction and the emotional dimension of fine granularity. The reason for combination is that the overall score is the final result of the experience, and the emotional vector reveals the various internal factors that constitute this result, and the combination of the two can more finely represent the quality of the experience. After this step, the previously parsed point of access time sequence is bound to the unique experience feedback vector of the trip with "trip ID-1001" as the non-repetitive key association key, generating a complete record containing "behavior trajectory" and "experience image", that is, a trip-evaluation mapping record. When a large number of such records are collected, a trip-evaluation mapping library for deep mining is built. By creating a knowledge base that can make large-scale and computable association between "behavior patterns" of a large number of user groups and "experience results", it makes it no longer isolated to look at the popularity of scenic spots or user scores, but can study the core problem of "how do tourists actually feel under a specific tour sequence and rhythm". For example, the library may contain a trip pattern of "visiting Huangshan the next day after visiting Hongcun", and the "fatigue" dimension in the corresponding experience feedback vector is significantly high; while the pattern of "arranging a hot spring hotel for rest after visiting Huangshan" has more prominent "relaxation" and "overall satisfaction" dimensions. The establishment of this library provides indispensable data raw materials and verification basis for subsequent identification of scenic spots prone to fatigue and discovery of trip arrangement optimization strategies that can improve satisfaction, and is the bottom layer data pillar for the whole scheme to realize the transition from "planning a trip" to "managing an experience".

[0092] In another preferred embodiment of the present application, the specific acquisition process of the specific scenic spot in S2 is:

[0093] Extract all scenic spot access sequences from the trip-evaluation mapping library, and collect the scenic spot identifiers therein to form an initial scenic spot set; for each scenic spot identifier in the initial scenic spot set, retrieve all trip-evaluation mapping records containing the identifier from the trip-evaluation mapping library to form an associated record set;

[0094] For each record in the associated record set, locate the access time point of the scenic spot identifier in its scenic spot access sequence, and divide the sequence into a pre-sequence subsequence and a subsequent subsequence according to the time point; obtain the local experience feedback mean values corresponding to the pre-sequence subsequence and the subsequent subsequence respectively, and calculate the difference value as the individual experience feedback attenuation degree of the record;

[0095] statistical analysis is performed on the individual experience feedback attenuation degrees of all records in the association record set to obtain a group experience feedback attenuation degree of the scenic spot identifier and statistical distribution parameters thereof; after traversing the initial scenic spot set, the group experience feedback attenuation degree of each scenic spot identifier and the statistical distribution parameters thereof are filtered through statistical significance testing;

[0096] The scenic spot identifier with a negative group experience feedback attenuation degree and a statistical significance exceeding a preset threshold is marked as a specific scenic spot to form a specific scenic spot set.

[0097] All recorded journey routes are extracted therefrom, and all scenic spot names appearing in each route are collected together to form a large "initial scenic spot set", for example, the set can include hundreds of places such as "Huangshan", "Hongcun", "Xidi", and "Ancient Huizhou City". This is done because it needs to independently and fairly evaluate each scenic spot in a unified framework without assuming any properties of the scenic spots in advance. Next, for each scenic spot in the set, for example, "Huangshan", a search is performed in the entire mapping library to find all records of the tourist journey containing "Huangshan", and these records constitute the "association record set" of "Huangshan". The principle of this step is that to scientifically evaluate the influence of a scenic spot, it is necessary to be based on a sufficient number of actual cases containing the scenic spot to avoid misjudgment due to individual accidents.

[0098] Subsequently, each personal journey record in the "association record set" is analyzed in depth. Taking a tourist record containing "Huangshan" as an example, first, the specific time point when "Huangshan" is visited is accurately located on the timeline of the journey, and then the entire journey is divided into a "pre-Huangshan part" and a "post-Huangshan part", i.e., a pre-sequence and a post-sequence, based on this time point. After that, the "experience feedback vectors" associated with the two subsequences are read respectively. Since the "experience feedback vector" is a comprehensive numerical representation containing an overall score and multiple emotional dimensions, by calculating the average value of all dimensions, a value representing the "average experience level before Huangshan" and a value representing the "average experience level after Huangshan" can be obtained. Subtracting the "before" value from the "after" value gives a difference value, which is defined as the "individual experience feedback attenuation degree" of "Huangshan" in this record. If the difference value is negative, it means that the experience level of this tourist's journey has decreased after visiting Huangshan. The principle is that by comparing the state changes of the same individual before and after a specific event, the influence of the event itself can be more effectively stripped off, and the interference caused by the differences in the basic experience of different individual tourists can be reduced.

[0099] After calculating the individual attenuation degree of all the associated records of "Huangshan", all these positive or negative individual attenuation degree values are put together to calculate their average value (i.e. the group experience feedback attenuation degree) and the fluctuation range (statistical distribution parameters). The average value reflects the overall trend of the impact of "Huangshan" on the tourist experience, while the fluctuation range indicates whether this impact is stable and consistent. After traversing all the scenic spots in the "initial scenic spot set", finally, a statistical significance test method is used to screen the list. For example, for "Huangshan", a statistical test is performed to determine whether the calculated negative average attenuation degree is large enough to be unlikely caused by random fluctuations; at the same time, "Huizhou Ancient City" may also show a weak negative attenuation, but its fluctuations may be large and the statistical test cannot confirm its significance. Only those scenic spot identifiers like "Huangshan", whose group experience feedback attenuation degree is not only negative (meaning it generally leads to a decline in experience), but also this negative impact is confirmed as "significant" through statistical tests, will be finally marked as "specific scenic spots" and form the "specific scenic spot set". The fundamental principle of doing this is to introduce scientific statistical inference to convert the subjective "feeling tired of a certain scenic spot" into objective, repeatable data conclusions, so as to ensure that the identified risk points are real and universally exist, rather than accidental or imagined.

[0100] The traditional method may rely on operating experience or sporadic complaints to guess which scenic spots are intense, but this is neither comprehensive nor accurate. However, the present application can answer the key question "which scenic spots indeed and universally lower the subsequent experience of tourists" through the above steps. The advantage of doing this is that it provides accurate, data-driven "target points" for all subsequent optimization strategies. Identifying "Huangshan" as a "specific scenic spot" and putting it into the "specific scenic spot set" means that a key knowledge is obtained: special care and optimization design are needed when arranging "Huangshan" in the itinerary. Because only by knowing where the "problem" is most likely to occur can preventive layout be effectively carried out in the planning stage, such as avoiding the close arrangement of multiple "specific scenic spots" or inserting buffer points between them. Therefore, this step is an indispensable bridge between "historical data phenomenon discovery" and "future itinerary active optimization", which enables the entire method to jump from passive feedback recording to active risk prediction and management.

[0101] In another preferred embodiment of the present application, the specific process of the common behavior deviation characteristic value in S2 is:

[0102] For each specific scenic spot in the specific scenic spot set, extract its corresponding associated record set from the itinerary-evaluation mapping library; from the associated record set, select records with a negative individual experience feedback attenuation degree and an absolute value exceeding a first preset threshold to form a high attenuation record subset;

[0103] For each record in the high-decay record subset, the trajectory data within a fixed time window from the visiting time point of the specific attraction in the attraction visiting sequence of the record is intercepted; based on the trajectory data, the moving speed change rate and the stay event frequency change rate of the record within the fixed time window compared to a control time window before the visiting time point are calculated;

[0104] The moving speed change rate and the stay event frequency change rate of all records in the high-decay record subset are aggregated, and the arithmetic mean and the standard deviation of the two are calculated respectively; the combination of the arithmetic mean and the standard deviation of the moving speed change rate and the arithmetic mean and the standard deviation of the stay event frequency change rate are taken as the common behavior deviation characteristic value corresponding to the specific attraction.

[0105] First, for each "specific attraction" that has been identified, such as "Huangshan", a secondary screening is performed from its large "associated record set". Instead of analyzing all tourists who have visited Huangshan, those records with a "individual experience feedback decay degree" of negative and large absolute value are intentionally screened out, that is, those tourists who have experienced a significant decline in experience after visiting Huangshan, constitute a "high-decay record subset". By analyzing the group with the most obvious experience damage, the external behavior that may have common characteristics that lead to a decline in experience can be more clearly captured, and the interference of individual cases that have not been significantly affected or even improved by visiting Huangshan is excluded, so that the subsequent search for behavior patterns is more representative and targeted.

[0106] Next, micro-level behavioral analysis is performed on each record in the subset. Take a filtered visitor record as an example, starting from the exact time point of leaving "Huangshan", a fixed and statistically meaningful time window (e.g. the next 24 hours) is extracted from the visitor's sequence of visits. At the same time, in order to establish a comparable baseline, a "control time window" of the same length before visiting "Huangshan" is also extracted. Based on the continuous location coordinates (such as GPS track points) and timestamps within these two time windows, the following calculations are performed: it calculates the average moving speed by analyzing the distance of coordinate change per unit time, and defines a "stay event" by identifying the long-time aggregation of track points in one place. Then, the average moving speed in the fixed time window is divided by the average moving speed in the control time window, and 1 is subtracted, to get the "moving speed change rate". Similarly, the number of stay events per hour in the fixed time window is subtracted from the control time window, and then divided by the control value, to get the "stay event frequency change rate". It can be understood that the decline in moving speed can directly map the decline in physical strength and the decrease in action willingness, while the abnormal increase in stay frequency may reflect the need for more rest in physiology or fatigue and observation in psychology. By calculating the change rate rather than the absolute value, the difference in the walking habits of different visitors can be eliminated, and the relative behavior change of the same individual before and after a specific event (visiting Huangshan) can be focused on.

[0107] After completing the above calculations for all visitors in the high-degradation record subset, the induction and abstraction phase is entered. It collects all the "moving speed change rates" calculated by the visitors, calculates their arithmetic mean and standard deviation. The mean represents the average reduction in moving speed of the group of visitors whose experience has significantly decreased after visiting Huangshan, revealing the general strength of fatigue; while the standard deviation reflects the fluctuation range of the speed reduction, indicating the dispersion degree of fatigue performance. The same calculation is performed on the "stay event frequency change rate". Finally, the combination of the mean and standard deviation of the two indicators is used as the "common behavior deviation characteristic value" corresponding to the specific scenic spot "Huangshan". For example, a possible value is "speed change rate: mean -30%, standard deviation 5%; stay frequency change rate: mean +50%, standard deviation 15%". The principle of this characteristic value is that it is no longer a isolated number, but a statistical description containing central tendency and dispersion, which can more robustly and comprehensively define a class of behavior patterns. It means that a typical visitor whose experience is impaired by visiting Huangshan has a high probability of falling within the range defined by this statistical description in subsequent behavior.

[0108] By strongly correlating the subjective "experience drop" with the objective, measurable "behavior signals", a clear "fingerprint" or "template" is provided for the machine to recognize. Knowing that "Huangshan" is not just a place that statistically leads to a score drop, but more specifically, that this drop is usually accompanied by a quantifiable behavior pattern of "a general speed drop of about 30%, and a rest frequency increase of about 500%". The benefit of doing this is twofold: first, it enables an indirect, objective and scalable measurement of the "fatigue" or "experience impairment" state. Instead of asking the user "are you tired", it probabilistically infers by analyzing their anonymous trajectory data. When a new visitor is touring Huangshan, it can be determined whether this visitor is currently heading into a similar state of experience drop as most people in history by calculating their actual behavior indicators and comparing them with this pre-prepared "common behavior shift feature value". The key knowledge transformation from discovering the problem (a specific attraction), to characterizing the problem (a behavior shift feature), to making the problem diagnosable in real time is completed. Without this step, it would not be possible to evolve from "knowing that Huangshan can make people tired" to "being able to identify a visitor who is getting tired", and all subsequent personalized dynamic interventions would be out of the question. The solution would remain at the level of post-mortem analysis.

[0109] In another preferred embodiment of the present application, the specific construction process of the association rule base in S2 is as follows:

[0110] For each specific attraction in the specific attraction set, a high attenuation record subset is obtained; based on the common behavior shift feature values of each record in the high attenuation record subset, clustering analysis is performed to form a plurality of behavior feature clusters;

[0111] For each record in the high attenuation record subset, the first subsequent attraction identifier after the specific attraction in the attraction visiting sequence and its corresponding experience feedback vector are extracted;

[0112] The subsequent attraction identifier and the experience feedback vector are assigned to the corresponding behavior feature cluster; for each behavior feature cluster, the records within it are grouped according to the subsequent attraction identifier to form a plurality of candidate option subsets;

[0113] The mean value of the experience feedback vectors corresponding to all records in each candidate option subset is calculated; the experience feedback vector mean values of different candidate option subsets in the same behavior feature cluster are subjected to statistical significance test;

[0114] The subsequent attraction identifier corresponding to the candidate option subset whose experience feedback vector mean value is significantly higher than that of other options in the test result is extracted as the experience optimization decision set of the behavior feature cluster;

[0115] establishing a mapping relationship between the behavior feature cluster of the specific scenic spot and the experience optimization decision set; integrating all mapping relationships of all specific scenic spots to construct the association rule base.

[0116] First, for each labeled "specific scenic spot", such as "Huangshan", its corresponding "high attenuation record subset" is retrieved, that is, those records of tourists whose experience dropped significantly after visiting Huangshan. Analyze the "common behavior deviation feature values" (such as a set of values including speed and stay changes) carried by each of these records, and use clustering analysis method to automatically divide all records into several internally consistent groups, i.e. "behavior feature clusters", according to the similarity of these values. For example, it may be found that among these tourists with poor experience, a part of people's behavior features show "sharp decline in moving speed and very frequent rest", and are classified into "high fatigue cluster A"; while another part of people show "moderate decline in speed and slight increase in rest", and are classified into "medium fatigue cluster B". It can be understood that even if the experience drops, the degree of fatigue and the performance pattern behind it may differ, and refined clustering helps to provide more accurate and hierarchical optimization suggestions in the future, rather than giving general solutions to all cases.

[0117] After clustering is completed, the choices actually made by tourists in each cluster in the fatigue state and their results are explored. It traverses each record in the "high attenuation record subset", locates the next place the tourist goes to after leaving the "specific scenic spot" (Huangshan), and extracts the identification of this place, such as "Hongcun" or "Huizhou Ancient City". At the same time, the complete "experience feedback vector" corresponding to this record is read, which represents the final evaluation and feelings of the tourist on the entire trip. Then, the pair of information "subsequent scenic spot identification" and "experience feedback vector" is classified under the "behavior feature cluster" to which the record belongs. The association between "behavior state", "subsequent action" and "final experience" is established, aiming to answer the key question: when in a certain specific fatigue state (such as high fatigue cluster A), choosing different subsequent activities (going to Hongcun or Huizhou Ancient City), does the final experience result differ?

[0118] Detailed statistical analysis is performed within each "behavior feature cluster". It first groups all records within a cluster according to their selected "next attraction identifier", for example in "high fatigue cluster A", it will form multiple "candidate option subsets" such as "group that selected Hongcun" and "group that selected Huizhou Ancient Town". Then, it calculates the average of all "experience feedback vectors" within each subset, which quantitatively reflects the overall experience level of the group that selected that option. Then, it performs statistical significance test (such as t-test) between these subsets, to judge whether the difference in experience average caused by different options is large enough to be unlikely caused by random chance. For example, it might find that in "high fatigue cluster A", the experience feedback average of "group that selected Huizhou Ancient Town" is significantly, rather than accidentally, higher than that of "group that selected Hongcun". Finally, it extracts the attraction identifier (such as "Huizhou Ancient Town") that leads to statistically significantly higher experience feedback as the "experience optimization decision set" of that behavior feature cluster. This means that an empirical rule is verified through historical data: "when a tourist is in high fatigue state A, turning to Huizhou Ancient Town will statistically lead to better overall experience than going to Hongcun as originally planned".

[0119] Finally, the mapping relationship between each "behavior feature cluster" and corresponding "experience optimization decision set" is established for each "specific attraction" (such as "Huangshan -> high fatigue cluster A -> recommend {Huizhou Ancient Town}"), and all such mapping relationships of all specific attractions are integrated, which constitutes the "association rule base" that drives the entire intelligent planning. Through distillation of massive, fuzzy historical experience into clear, verifiable, executable decision knowledge, the purpose is not simply to record where the tourists went, but to discover and solidify the best practices of "what action should be taken in what state to achieve better results". This rule base is the "brain" and "knowledge base" of intelligent decision-making. It enables the initial itinerary planning or dynamic adjustment to go beyond simple geographical and popularity logic, and instead make recommendations based on empirically proven, state-matched experience optimization strategies. For example, when a tourist is predicted or detected to be in "high fatigue state A", it will unhesitatingly retrieve "Huizhou Ancient Town" as the optimization decision from the rule base, rather than recommending a popular attraction that might exacerbate fatigue.

[0120] In another preferred embodiment of the present application, the specific process of initial itinerary planning in S3 is:

[0121] Parse the current user request to obtain a list of target attraction identifiers, generate a set of attractions to be planned; match the set of attractions to be planned with the set of specific attractions, and mark the identifiers that belong to the set of specific attractions as to-be-optimized nodes;

[0122] For each to-be-optimized node, query the mapped experience optimization decision set from the association rule base, and extract the scenic spot identifiers contained in the set; add the extracted scenic spot identifiers to the to-be-planned scenic spot set to form an expanded scenic spot set;

[0123] Based on the expanded scenic spot set and the space-time constraint data between scenic spots, generate an initial scenic spot access sequence; traverse the initial scenic spot access sequence to identify to-be-optimized nodes arranged as continuous access nodes in the sequence;

[0124] When continuous to-be-optimized nodes are identified, obtain the experience optimization decision set mapped by the first to-be-optimized node from the association rule base;

[0125] Determine whether the second to-be-optimized node belongs to the scenic spot identifiers in the experience optimization decision set; if not, select a scenic spot identifier from the experience optimization decision set to replace the second to-be-optimized node in the sequence or insert it before the second to-be-optimized node;

[0126] Combine the adjusted scenic spot access sequence with traffic and time consumption data to generate the initial itinerary plan.

[0127] When a user inputs a target scenic spot list that he or she expects to visit, such as "Huangshan, Hongcun, and Xidi", through an application interface, first analyze the request and convert it into a machine-processable "to-be-planned scenic spot set". By standardizing the user's free-form language expression into a structured data object, an explicit input range is established for all subsequent automated decisions. Next, compare the user's intention set with the "specific scenic spot set" obtained through historical data analysis. For example, if "Huangshan" is found in the specific scenic spot set, it is immediately marked as a "to-be-optimized node" in the to-be-planned set. It can be understood that based on historical knowledge, it is known in advance that "Huangshan" is a scenic spot that needs special careful arrangement and is prone to fatigue accumulation, so it is marked at the beginning of planning to give it special attention.

[0128] After identifying all to-be-optimized nodes, it queries for each to-be-optimized node, such as "Huangshan", which subsequent choices have been verified to bring better experience when the tourists are in different fatigue states triggered by this scenic spot. Extract the corresponding "experience optimization decision set" from the rule base, such as the suggestion set "{Huizhou Ancient City}". Then, add these high-quality alternative scenic spot identifiers (such as "Huizhou Ancient City") refined from historical experience to the user's original intention set to form an "expanded scenic spot set". The principle is that the initial plan should not be limited to the limited options proposed by the user, but should actively introduce historically verified effective buffer or alternative resources to expand the solution space for generating a better solution.

[0129] Next, based on the "extended attraction set", combined with the actual distance between each attraction, the conventional tour length, the opening time and other "spatial and temporal constraint data", the path planning algorithm is run to generate a preliminary, spatio-temporally feasible "initial attraction visiting sequence". A possible generation result is "Day 1: Hongcun, Day 2: Huangshan, Day 3: Xidi". After generation, a key "risk flow simulation" check will be performed on this sequence: it traverses the entire sequence and specifically examines whether the marked "to-be-optimized nodes" are arranged too densely, that is, whether two or more such nodes are arranged for continuous access. For example, it will find that "Huangshan" and "Xidi" are next to each other in this sequence. It can be understood that continuous access to multiple high-fatigue-risk points is a typical trap that leads to experience collapse, and such dangerous patterns must be actively checked.

[0130] Once such consecutive risk nodes are identified, automatic optimization logic is triggered. It first focuses on the first to-be-optimized node (Huangshan) and obtains its corresponding "experience optimization decision set" from the associated rule base. Then, it determines whether the second to-be-optimized node (Xidi) that follows is within this optimization decision set. In this example, if "Xidi" is not within the set "{Hui-style ancient city}", it means that, according to historical experience, it is not the optimal choice to visit "Xidi" immediately after the fatiguing "Huangshan". Therefore, a node (such as Hui-style ancient city) will be selected from the optimization decision set to perform a surgical adjustment on the original sequence: either replace "Xidi" directly with "Hui-style ancient city", or insert "Hui-style ancient city" between "Huangshan" and "Xidi". By using data-driven rules, the identified high-risk structure is actively intervened, and a historically verified high-quality choice is used to buffer or replace the arrangement that may trigger experience decline.

[0131] Finally, the adjusted new attraction visiting sequence is filled with detailed transportation connection methods, time required for transfer, recommended stay length at attractions, etc., thereby generating an "initial itinerary plan" that can be directly delivered to the user, with a timetable and action guide. By changing the post-event, passive experience management into a pre-event, active risk avoidance. Traditional planning methods only string resources based on the user's explicit preferences and are blind to potential experience risks. The process described in the present invention simulates the "fatigue load" of the itinerary in the planning stage and pre-judges and automatically recombines the risk points based on historical group wisdom (associated rule base). The purpose is to inject "resilience" into the itinerary before it starts, so that the generated plan itself has built-in optimization strategies to combat common experience traps.

[0132] In another preferred embodiment of the present invention, the specific calculation process of the behavior feature vector in S4 is:

[0133] obtaining a real-time positioning sequence and a stay duration sequence uploaded by a current user mobile terminal; extracting a leaving time point of a last visited scenic spot from the real-time positioning sequence, and taking the leaving time point as a starting point to obtain a subsequent real-time positioning sequence in a fixed time window;

[0134] based on the subsequent real-time positioning sequence in the fixed time window, calculating a moving speed sequence in the fixed time window, and calculating a change rate of the moving speed sequence compared with a moving speed sequence in a control time window before the leaving time point;

[0135] based on the stay duration sequence, counting a frequency of a stay event in the fixed time window, and calculating a change rate of the frequency compared with a stay event frequency in the control time window;

[0136] combining the calculated change rate of the moving speed sequence and the change rate of the stay event frequency to generate a current behavior feature vector of the current user.

[0137] When a user moves in his journey, the application on the smart phone or other mobile terminal he carries will continuously and anonymously record two core data streams in the background: one is the latitude and longitude coordinates collected at fixed time intervals, forming a "real-time positioning sequence"; the other is the length of time the user is stationary at each coordinate point, forming a "stay duration sequence". The principle of collecting these data is that a person's movement pattern in a large-scale space (such as between scenic spots) and his micro-downtime habits are the most direct and objective external reflection of his physical condition, interest level and action willingness, and can be obtained completely by the built-in sensors of the device without the need for the user to actively report.

[0138] Then, a key reference time point needs to be determined, for example, the "leaving time point of the last visited scenic spot". The real-time positioning sequence will be compared with the known scenic spot geographic fence range. For example, when it is detected that the user's continuous coordinate points move from the preset geographic fence of "the south gate of Huangshan" to outside the fence, and do not return afterwards, it can be determined that the user has left the scenic spot "Huangshan", and the timestamp of the last coordinate point located in the fence is recorded as the "leaving time point". The principle of doing so is that taking the time when the user leaves a clear activity node (scenic spot) as the starting point of evaluating the subsequent state change can ensure that the time anchor point of the analysis has clear physical meaning and consistency.

[0139] After the departure time point is determined, all positioning data within a pre-defined length of "fixed time window" (e.g. the following 6 hours) immediately backward (future direction) from the time point is intercepted as "subsequent real-time positioning sequence". Meanwhile, in order to establish a comparable individualized baseline, positioning data within a "control time window" (e.g. the last 6 hours during the Huangshan tour) of the same length forward (past direction) from the same departure time point is intercepted. The principle of selecting windows of the same length and adjacent to each other for comparison is to maximize the fairness of comparison: the difference between the behavior of the same user in the late stage of touring the same core scenic spot (Huangshan) and the early stage after finishing the tour, so as to more purely isolate the immediate impact of the event of "finishing the tour" on the behavior pattern, and reduce the confusion factors caused by different time periods in a day (such as usually more energetic in the morning than in the afternoon) or different activity types.

[0140] Subsequently, quantitative calculation is carried out based on the positioning sequences of the two time windows. For the sequence in the "control time window", the average moving speed of the user in the late stage of touring the scenic spot (such as the last stage of climbing and descending the mountain) is calculated, which can be called "baseline speed". For the sequence in the "fixed time window" (i.e. the subsequent observation window), the system calculates the average moving speed of the user after leaving the scenic spot (such as on the way back to the hotel and evening activities), which is called "observation speed". The "observation speed" is divided by the "baseline speed" and then subtracted by 1 to obtain a ratio value, which is the "change rate of moving speed". If the value is negative, it indicates that the user's movement has slowed down. The principle is that muscle fatigue, mental relaxation or reduced interest will directly lead to a decrease in the displacement efficiency of the human body, and the speed change rate is a continuous index sensitive to the overall physiological and psychological energy level.

[0141] At the same time, the "stay duration sequence" is analyzed, and a "stay event" determination standard (e.g. continuous stationary for more than 90 seconds within a 50-meter diameter) is defined within the above-mentioned same two time windows. The total number of stay events occurring in the "control window" is counted and divided by the window duration to obtain the "baseline stay frequency"; the total number of stay events in the "subsequent observation window" is counted and divided by the window duration to obtain the "observation stay frequency". Similarly, the change rate of the "observation stay frequency" relative to the "baseline stay frequency" is calculated. The principle is that in the process of non-target-oriented movement (such as the return trip after the tour ends), an increase in the number of purposeless and unnecessary stays is often a signal of the body seeking rest, distraction of attention or decision-making delay (such as not knowing where to go next), which is closely related to fatigue accumulation.

[0142] Finally, the two scalar values of "moving speed change rate" and "stay frequency change rate" are combined into a two-dimensional "behavior feature vector". For example, a possible vector is [-0.25, +0.40], indicating that the speed has decreased by 25% and the stay frequency has increased by 40%. The principle of doing this is that a single indicator can be disturbed by a specific situation (such as slow speed because of taking pictures), but two indicators reflecting fatigue from different sides (action efficiency and rest demand) combined together can confirm each other, forming a more stable and comprehensive digital characterization of the user's current comprehensive behavior state. This vector is no longer a string of original and ambiguous coordinate points, but a standardized mathematical representation that condenses key behavior change information.

[0143] It can be understood that the traditional method cannot know whether the user is "tired" after visiting a scenic spot, and the method can convert the abstract "tired" into specific and comparable digital vectors through the ingenious design of the comparison window and the calculation of the double indicators. Real-time indirect but measurable diagnosis of the user's fatigue state is achieved. It does not depend on the user's subjective expression (which may not be accurate or willing to disclose), nor does it depend on special medical equipment (such as a heart rate band), but only uses the ubiquitous smart phone data to achieve it. This "behavior feature vector" is a crucial "sensing signal". It is a bridge connecting "user real-world behavior" and internal "intelligent decision-making model". Subsequently, the classification model identifies the fatigue category to which the vector belongs, and accordingly triggers the corresponding trip adjustment.

[0144] In another preferred embodiment of the present application, in the S4, the specific construction process of the preset behavior deviation classification model is:

[0145] From the association rule base, extract the clustering center feature vector of each behavior feature cluster corresponding to the specific scenic spot; take each clustering center feature vector as a model training sample, and take the identification of the behavior feature cluster from which the clustering center feature vector is derived as the class label of the training sample;

[0146] Collect all the training samples and class labels corresponding to all the specific scenic spots to construct a labeled training sample set; use the labeled training sample set to supervise the training of a multi-class machine learning model with a preset basic structure;

[0147] Save the multi-class machine learning model that has completed training and reached the preset precision index as the behavior deviation classification model.

[0148] In another preferred embodiment of the present application, in the S5, the specific process of dynamically adjusting the remaining trip plan of the current user is:

[0149] obtaining a deviation category to which the real-time behavior of the current user belongs, output by the behavior deviation classification model; and obtaining a corresponding experience optimization decision set from the association rule base according to the deviation category;

[0150] obtaining real-time geographic location information of the current user and a sequence of scenic spots to be visited in the remaining trip plan; generating a subsequent trip segment, taking the real-time geographic location as a starting point, taking the experience optimization decision set as a candidate destination constraint, and combining real-time traffic condition data and scenic spot opening time data;

[0151] fusing and comparing the generated subsequent trip segment with an original scenic spot sequence in the remaining trip plan that has not been executed, and replacing a scenic spot identifier in the original scenic spot sequence that does not belong to the experience optimization decision set with a scenic spot identifier in the subsequent trip segment that belongs to the experience optimization decision set;

[0152] combining the new trip plan after the fusion and replacement processing with real-time updated traffic modes and predicted time consumption data to obtain an adjusted dynamic trip plan.

[0153] When the user is executing the initial trip, for example, just after ending the tour of “Huangshan” and entering the hotel, the mobile phone application continuously collects and uploads anonymized positioning and stay data in the background. The pre-trained “behavior deviation classification model” is used to analyze these fresh data in real time. The principle of the model is that it has learned and stored different fatigue behavior patterns (such as “high fatigue cluster A” and “medium fatigue cluster B”) exhibited by tourists after visiting various “specific scenic spots” in history as classification standards. At this time, the model compares the behavior data features (such as slow movement and frequent short stays) generated by the current user with these standards, and finally outputs a judgment result, for example, “current behavior attribution: high fatigue cluster A”. The essence of this step is to complete a transformation from raw sensor data to high-order physiological state cognition, and to “understand” the fatigue level in which the user may currently be.

[0154] Once the clear "offset category" diagnosis is obtained, it is immediately sent as a precise query instruction into the "association rule base" for retrieval. The principle of the rule base is to store the best mapping of "state-action-result" that has been historically validated. According to the key "high fatigue cluster A", the corresponding "experience optimization decision set" is quickly found, such as a set containing the labels "Hui Zhou Ancient City", "Hot Spring Experience", etc. This means that historical big data shows that tourists in a similar state of fatigue will obtain better subsequent experiences by choosing these activities. Next, two key pieces of information are obtained simultaneously: one is the user's "real-time geographic location" (such as a hotel at the foot of Huangshan) obtained through the mobile phone GPS, and the other is the "sequence of remaining attractions to be visited" in the user's itinerary plan (such as planned to visit "Hongcun" tomorrow). The principle of obtaining these two is to start the re-planning under the constraints of the real physical world: knowing "where to start from" and "where to go originally".

[0155] Taking the user's "real-time geographic location" as the starting point of the planning, all the attraction labels in the "experience optimization decision set" are taken as strong constraint candidate destinations - that is, the target of the planning must be selected from these verified excellent options. At the same time, real-time traffic data (such as the current road congestion from the hotel to each candidate location, available public transportation schedules) and real-time opening time data of each attraction are accessed. Based on all these dynamic inputs, the engine quickly calculates and generates one or more optimal "subsequent itinerary segments". For example, it may generate a segment: "tomorrow morning 9:30, take a direct bus from the current hotel to Hui Zhou Ancient City, and plan to visit for 3 hours". The principle of this step is to solve the most feasible and efficient execution scheme at the current time and location under the premise of ensuring the experience optimization goal (going to the attractions in the decision set).

[0156] Next, the intelligent decision fusion stage is entered. It compares the newly generated "subsequent itinerary segment" with the user's established, yet unexecuted "original sequence of attractions". The core logic is to check whether the original plan conflicts with the optimization suggestion. In this example, it is found that the next attraction in the original sequence, "Hongcun", is not in the "experience optimization decision set" {Hui Zhou Ancient City, Hot Spring Experience}. According to the predetermined optimization rules, the replacement logic is immediately started: it uses the attraction label in the newly generated itinerary segment that belongs to the decision set (i.e. "Hui Zhou Ancient City") to replace the one in the original sequence that does not belong to the decision set (i.e. "Hongcun"). The principle is that when the historically validated optimal choice is inconsistent with the original plan, the choice that data proves can bring better experience results is preferred to achieve active gain for the user's journey.

[0157] Finally, the new sequence of scenic spots obtained after the replacement is combined with real-time updated specific traffic mode, departure time, estimated time consumption and other detailed information, encapsulated into a complete and instant "adjusted dynamic travel plan", and immediately pushed to the mobile terminal of the user, realizing the leap from "static planning" to "companion intelligent agent". The traditional travel planning is basically completed after the plan is delivered, while the dynamic adjustment mechanism described in the present application enables continuous perception of the user state during the journey and active intervention at critical moments (such as when significant fatigue is detected), providing an evidence-based and personalized optimization scheme. The benefit of this is to greatly improve the adaptability and resilience of the journey, enabling it to flexibly respond to unexpected changes in the user's state.

[0158] The above describes one embodiment of the present application in detail, but the content described is only a preferred embodiment of the present application and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made within the scope of the present application should still be attributed to the patent coverage of the present application.

Claims

1. A tourism itinerary planning method based on artificial intelligence, characterized in that, Includes the following steps: S1. Obtain the historical travel sequences and corresponding travel evaluation data of anonymous user groups in the target area, and construct a travel-evaluation mapping library; S2, perform group behavior analysis on the itinerary-evaluation mapping library, identify common behavior shift feature values ​​after visiting a specific attraction, and cluster the attraction visit records in the historical itinerary sequence that are located after the specific attraction based on the common behavior shift feature values ​​to form an association rule library between behavior shift feature values ​​and subsequent attraction selection clusters. The specific process for obtaining information about a particular attraction is as follows: Extract all attraction visit sequences from the trip-evaluation mapping database and aggregate the attraction identifiers to form an initial attraction set; for each attraction identifier in the initial attraction set, retrieve all trip-evaluation mapping records containing that identifier from the trip-evaluation mapping database to form an associated record set; For each record in the associated record set, locate the visit time point of the attraction identifier in its attraction visit sequence, divide the sequence into a preceding subsequence and a subsequent subsequence based on the time point; obtain the mean local experience feedback corresponding to the preceding subsequence and the subsequent subsequence respectively, and calculate the difference between the two as the individual experience feedback attenuation degree of the record. Statistical analysis is performed on the individual experience feedback decay rate of all records in the associated record set to obtain the group experience feedback decay rate and its statistical distribution parameters of the scenic spot identifier; after traversing the initial scenic spot set, the group experience feedback decay rate and its statistical distribution parameters of each scenic spot identifier are filtered through statistical significance test. Scenic spots with negative group experience feedback decay and statistical significance exceeding a preset threshold are marked as specific scenic spots and form a specific scenic spot set; S3, responding to the current user's planning request, and generating an initial itinerary plan containing a sequence of attraction visits based on the association rule base; S4, during the process of the current user executing the initial itinerary planning, the user's location sequence and dwell time sequence are obtained in real time and the behavior feature vector is calculated. The behavior feature vector is input into the preset behavior offset classification model and the offset category to which the current user's real-time behavior belongs is output. S5. Obtain the corresponding target attraction selection cluster from the association rule base according to the offset category, dynamically adjust the remaining itinerary plan of the current user based on the target attraction selection cluster, and synchronize it to the current user's terminal device.

2. The artificial intelligence-based tourism itinerary planning method according to claim 1, characterized in that, In S1, the specific construction process of the trip-evaluation mapping library is as follows: Obtain the historical itinerary records and historical evaluation records of anonymous user groups; parse out the attraction visit sequence indexed by itinerary identifiers from the historical itinerary records, the attraction visit sequence containing attraction identifiers arranged by timestamps; The evaluation data corresponding to each itinerary identifier is parsed from the historical evaluation record set. The evaluation data includes numerical overall score data and text-based comment data. Sentiment analysis is performed on the text-based comment data to extract the sentiment tags and intensity values ​​that represent experience tendencies and generate sentiment vectors. The numerical overall rating data is concatenated with the emotion vector to form an experience feedback vector corresponding to the itinerary identifier; the itinerary identifier is used as the association key to bind the attraction visit sequence with the experience feedback vector corresponding to the itinerary identifier, generating an itinerary-evaluation mapping record; all itinerary-evaluation mapping records corresponding to all itinerary identifiers are collected to generate an itinerary-evaluation mapping library.

3. The artificial intelligence-based tourism itinerary planning method according to claim 2, characterized in that, In S2, the specific process of the common behavior offset feature value is as follows: For each specific attraction in the specific attraction set, extract its corresponding associated record set from the itinerary-evaluation mapping library; select records from the associated record set whose individual experience feedback decay is negative and whose absolute value exceeds a first preset threshold to form a high decay record subset; For each record in the high-attenuation record subset, the trajectory data within a fixed time window starting from the visit time of the specific scenic spot in its scenic spot visit sequence is extracted; based on the trajectory data, the rate of change of movement speed and the rate of change of dwell event frequency of the record within the fixed time window compared to a control time window before the visit time point are calculated; The rate of change of movement speed and the rate of change of dwell event frequency of all records in the high-attenuation record subset are collected, and their arithmetic mean and standard deviation are calculated respectively. The combination of the arithmetic mean and standard deviation of the rate of change of movement speed and the arithmetic mean and standard deviation of the rate of change of dwell event frequency is used as the common behavior offset feature value corresponding to the specific scenic spot.

4. The artificial intelligence-based tourism itinerary planning method according to claim 1, characterized in that, In S2, the specific construction process of the association rule base is as follows: For each specific attraction in the set of specific attractions, obtain its high-attenuation record subset; perform cluster analysis based on the common behavioral offset feature values ​​of each record in the high-attenuation record subset to form multiple behavioral feature clusters; For each record in the high-decay record subset, extract the identifier of the first subsequent attraction following that specific attraction in its attraction visit sequence and its corresponding experience feedback vector; The subsequent attraction identifiers and the experience feedback vectors are assigned to the corresponding behavioral feature clusters; for each behavioral feature cluster, the records within it are grouped according to the subsequent attraction identifiers to form multiple candidate option subsets; Calculate the mean of the experience feedback vectors corresponding to all records in each subset of candidate options; Perform a statistical significance test on the mean of the experience feedback vectors of different candidate option subsets within the same behavioral feature cluster; The subsequent attraction identifiers corresponding to the subset of candidate options whose mean experience feedback vector is significantly higher than that of other options in the test results are extracted as the experience optimization decision set of the behavioral feature cluster; Establish a mapping relationship between the behavioral feature clusters of the specific scenic spot and the experience optimization decision set; integrate all mapping relationships of all specific scenic spots to construct the association rule base.

5. The artificial intelligence-based tourism itinerary planning method according to claim 1, characterized in that, In S3, the specific process of initial trip planning is as follows: Parse the current user request to obtain a list of target attraction identifiers, and generate a set of attractions to be planned; match the set of attractions to be planned with the specific attraction set, and mark the identifiers belonging to the specific attraction set as nodes to be optimized; For each node to be optimized, query its mapped experience optimization decision set from the association rule base, and extract the attraction identifiers contained in the set; add the extracted attraction identifiers to the set of attractions to be planned to form an expanded attraction set; Based on the expanded set of attractions and the spatiotemporal constraint data between attractions, an initial attraction visit sequence is generated; the initial attraction visit sequence is traversed to identify the nodes to be optimized that are arranged as consecutive visit nodes; When consecutive nodes to be optimized are identified, the experience optimization decision set mapped to the first node to be optimized is obtained from the association rule base; Determine whether the second node to be optimized belongs to the scenic spot identifier in the experience optimization decision set; if it does not belong, select a scenic spot identifier from the experience optimization decision set and replace the second node to be optimized in the sequence or insert it before it; The adjusted attraction visit sequence is combined with traffic and time data to generate the initial itinerary plan.

6. The artificial intelligence-based tourism itinerary planning method according to claim 1, characterized in that, In S4, the specific calculation process of the behavior feature vector is as follows: Obtain the real-time location sequence and dwell time sequence uploaded by the current user's mobile terminal; extract the departure time of the most recently visited attraction from the real-time location sequence, and extract the subsequent real-time location sequence with a fixed duration window starting from that time point; Based on the subsequent real-time positioning sequence within the fixed time window, calculate the movement speed sequence within the fixed time window, and calculate the rate of change of the movement speed sequence compared to the movement speed sequence within the control time window before the departure time point; Based on the dwell time sequence, the frequency of dwell events within the fixed duration window is statistically analyzed, and the rate of change of this frequency relative to the frequency of dwell events within the control duration window is calculated. The calculated rate of change of the movement speed sequence is combined with the rate of change of the dwell event frequency to generate the current user's current behavior feature vector.

7. The artificial intelligence-based tourism itinerary planning method according to claim 1, characterized in that, In S4, the specific construction process of the preset behavior offset classification model is as follows: From the association rule base, extract the cluster center feature vector of each behavioral feature cluster corresponding to all specific attractions; use each cluster center feature vector as a model training sample, and use the identifier of the behavioral feature cluster from which the cluster center feature vector originates as the category label of the training sample; Collect all training samples and category labels corresponding to all the specific attractions to construct a labeled training sample set; use the labeled training sample set to perform supervised training on a multi-class machine learning model with a preset basic structure; The multi-class machine learning model that has been trained and has reached the preset accuracy index is saved as the behavior offset classification model.

8. The artificial intelligence-based tourism itinerary planning method according to claim 1, characterized in that, In step S5, the specific process of dynamically adjusting the remaining travel plans of the current user is as follows: Obtain the offset category to which the current user's real-time behavior belongs, as output by the behavior offset classification model; and query the corresponding experience optimization decision set from the association rule base according to the offset category. Obtain the current user's real-time geographic location information and the sequence of attractions to be visited in the remaining itinerary plan; using the real-time geographic location as the starting point, the experience optimization decision set as the candidate destination constraint, and combining real-time traffic data and attraction opening time data, generate subsequent itinerary segments; The generated subsequent itinerary segments are merged and compared with the original attraction sequences that have not been executed in the remaining itinerary plan. If the attraction identifier in the original attraction sequence does not belong to the experience optimization decision set, then the attraction identifier in the subsequent itinerary segments that belongs to the experience optimization decision set is used to replace the attraction identifier that does not belong. The new itinerary sequence, after fusion and replacement processing, is combined with real-time updated traffic movement data and estimated travel time to obtain the adjusted dynamic itinerary plan.

Images