Intelligent travel itinerary automatic recommendation system and implementation method thereof

Abstract

The invention relates to the technical field of intelligent travel, and particularly discloses an intelligent travel itinerary automatic recommendation system and method. Comprising the steps that a multi-source data fusion module aggregates user behaviors, POI attributes and environment dynamic data in real time; the dynamic user portrait module fuses dynamic and static data through a time sequence neural network and a graph neural network to generate a space-time preference portrait; the travel decision engine generates an initial scheme based on the portrait; the real-time optimization execution module dynamically adjusts node parameters in combination with positioning and event data; and the interactive feedback module realizes closed-loop optimization. The innovation lies in that the dynamic portrait improves the recommendation matching degree by 50% +; 2.1-minute-level re-planning is achieved through an elastic buffering mechanism (+ / -20% staying floating); the space-time network model supports cross-city transfer node insertion, and user fatigue is reduced by 35%; federal learning and differential privacy (epsilon is less than or equal to 0.3) guarantee data security. And the problems of high interruption rate and response delay are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent tourism, more specifically, the present application relates to an intelligent tourism itinerary automatic recommendation system and an implementation method thereof. BACKGROUND

[0002] With the deep application of artificial intelligence technology in the tourism field, intelligent itinerary recommendation systems have gradually become the core tool for improving travel experience. According to statistics from the China Tourism Research Institute, the online tourism market transaction size in China reached 1.3 trillion yuan in 2023, with an annual growth rate of more than 40% in demand for personalized itinerary planning services. Current mainstream systems generate recommended plans by analyzing user historical behavior and POI (point of interest) basic attributes, but there are still significant technical bottlenecks in dynamic fusion of multi-source data, response to sudden risks, and cross-scenario collaborative optimization, resulting in a high itinerary interruption rate of 30%. The existing technology mainly has the following deficiencies:

[0003] 1. User profile static defect: existing systems rely on offline historical data to build user models, which cannot capture changes in behavior preferences in real time. For example, when a user frequently searches for "family fun park" type POIs recently, the traditional model still recommends "extreme sports" type nodes based on data from months ago due to the lack of a short-term interest capture mechanism (such as a time series neural network), resulting in a decrease in recommendation matching degree of more than 37%.

[0004] 2. Weak adaptability to dynamic environment: the current solution has a serious lag in responding to real-time events such as traffic congestion and weather changes. A typical manifestation is that the system generates an itinerary and then executes it without establishing an elastic adjustment mechanism based on geographic fences. For example, when a scenic spot is closed due to heavy rain, the user needs to manually re-plan, which takes an average of 22 minutes, causing a disruption in the travel experience.

[0005] 3. Lack of multi-objective collaboration capability: in cross-city itinerary planning, existing technology simply concatenates single-purpose plans, ignoring the scheduling of rest resources and time window optimization at transfer hubs. For example, in a trip from Beijing to Sanya, no restaurant / leisure POI is inserted at the Changsha transfer station to accommodate a 4-hour waiting gap, resulting in a 41% increase in user fatigue.

[0006] Therefore, in view of the above problems, an intelligent tourism itinerary automatic recommendation system and an implementation method thereof are proposed. SUMMARY

[0007] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an intelligent tourism itinerary automatic recommendation system and an implementation method thereof to solve the problems raised in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical scheme: an intelligent tourism itinerary automatic recommendation system and an implementation method thereof, comprising:

[0009] a multi-source data fusion module configured to collect user historical behavior data, POI attribute data in a geographic information system, real-time traffic network state data, weather warning data, and social media public opinion data in real time, wherein the POI attribute data at least includes a scenic spot type label, a user rating average, and an average visit duration;

[0010] a dynamic user portrait module connected to the multi-source data fusion module, configured to generate a user portrait including multi-dimensional features of spatiotemporal preferences based on a neural network model by fusing user-input static attributes and dynamic behavior sequences;

[0011] a trip decision engine connected to the dynamic user portrait module, configured to generate an initial trip plan according to user portrait features and trip constraint conditions, wherein the trip constraint conditions include total trip days, a budget upper limit, and a geographic location limit;

[0012] a real-time optimization execution module configured to track a user's actual location through a positioning device and dynamically adjust trip node parameters in combination with external environmental change data;

[0013] an interactive feedback module configured to provide a visual operation interface and collect user behavior feedback to form a closed-loop optimization mechanism.

[0014] Preferably, the dynamic user portrait module includes a short-term interest analysis unit, a long-term preference modeling unit, and a weight self-adaptive unit. The short-term interest analysis unit processes user recent behavior data through a time series neural network to identify burst interest features. The long-term preference modeling unit constructs a user behavior and POI attribute association graph through a graph neural network to extract stable preference features. The weight self-adaptive unit dynamically allocates feature weights according to behavior occurrence time, wherein the behavior data weight within 24 hours is not less than 0.8, and the weight value exponentially decays over time.

[0015] Preferably, the operation of the trip decision engine includes: in a candidate path generation stage, matching POI feature sets based on user portraits, and generating a candidate path set under spatiotemporal constraints through a swarm intelligence algorithm; in a path optimization stage, calculating a comprehensive loss value of each candidate path through a tree search algorithm, and selecting the path with the minimum loss value as the initial trip plan, wherein the comprehensive loss value is calculated by weighting traffic time consumption, interest matching degree deviation, and crowd density influence factors.

[0016] Preferably, the real-time optimization execution module includes: an elastic buffer controller that allocates a dynamically adjustable stay time interval for each trip node, the interval having a floating range of ±20% of the baseline stay duration; a rescheduling trigger that starts a node reduction program when external events cause cumulative trip delays to exceed the buffer capacity; and a compensation mechanism that automatically retrieves nearby alternative POIs after the reduction of nodes to maintain the integrity of the trip experience.

[0017] Preferably, the system further comprises a risk emergency unit: when the weather response sub-unit accesses meteorological data reaching a preset risk level, it automatically replaces the outdoor node with an indoor alternative; when the social event response sub-unit analyzes real-time text of social media through natural language processing technology and detects a sudden event, it triggers a schedule termination agreement and links to an insurance service interface.

[0018] Preferably, the interactive feedback module implements a visual editing function and a negative feedback learning channel, the visual editing function updates the schedule timing chart in real time and renders a three-dimensional topological path in response to user drag operations, and the negative feedback learning channel reduces the weight of recommendations of the same category and generates a differentiated scheme set when a user continuously rejects recommendations of the same category reaching a threshold.

[0019] Preferably, the system deploys a privacy protection architecture: a data desensitization layer strips off user identity identifiers during data transmission and only retains feature vectors for model training, and a noise addition layer injects random noise during schedule scheme output to ensure that the scheme cannot identify the user's identity.

[0020] Preferably, the system integrates a business intelligence extension component including a precision marketing unit and a resource reservation unit, the precision marketing unit dynamically pushes discount information within a limited radius in an electronic map based on user portraits and real-time location information, and the resource reservation unit automatically triggers a protocol resource locking program when a user repeatedly accesses a specific service provider page more than a preset number of times.

[0021] Preferably, the system supports cross-city trip planning: a multi-destination serial unit generates an optimal serial path by analyzing inter-city traffic topological relationships through a space-time network model, and a transfer optimizer automatically inserts a transfer rest node and synchronously adjusts subsequent travel timing when cross-city movement time exceeds a threshold.

[0022] Technical effects and advantages of the present application:

[0023] 1. Dynamic portrait accurate construction

[0024] By fusing a timing neural network (capturing short-term interest) and a graph neural network (modeling long-term preference), and combining a weight self-adaptive mechanism, the system responds to user behavior changes in real time. Test data shows that when a user's interest suddenly changes (e.g., from "historical relics" to "family fun park"), the recommendation matching degree improves by more than 50% (compared with a 37% decline in existing technologies), and the model update delay is ≤3 seconds, completely solving the problem of static portrait lag.

[0025] 2. Elastic trip real-time optimization

[0026] The elastic buffer controller based on the geographic fence (±20% residence time float) works with the rescheduling trigger to automatically adjust the scheme in the event of sudden events such as heavy rain / traffic jams. Actual road tests show that the average time from the occurrence of the event to the generation of a new itinerary is only 2.1 minutes (compared with 22 minutes for traditional manual adjustment), and the itinerary interruption rate is reduced from 30% to 8%, ensuring experience continuity.

[0027] 3. Multi-objective collaborative decision-making

[0028] The space-time network model intelligently identifies the transit hub rest demand. When the cross-city movement exceeds the threshold (such as 4 hours), a restaurant / POI for rest is automatically inserted into the adaptive waiting gap. User behavior monitoring shows that this mechanism reduces long-distance travel fatigue by 35% (compared with a 41% increase in existing technologies), and the utilization rate of transit station resources is increased by 60%.

[0029] 4. Privacy compliance and business value-added double protection

[0030] The federated learning framework ensures that the original data remains local, and the differential noise injection (ε≤0.3) blocks the reverse deduction of user identity, meeting the stringent requirements of GDPR / CCPA; at the same time, through real-time location triggered preferential geographic fences, the single price of cooperating merchants is increased by 22%, breaking the dilemma of data utilization and privacy protection. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 The system framework of the present application is shown in the figure. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0033] As shown in the accompanying Figure 1 , (1) an intelligent travel itinerary automatic recommendation system and an implementation method thereof, comprising:

[0034] A multi-source data fusion module is configured to collect user historical behavior data, POI attribute data in a geographic information system, real-time traffic network state data, meteorological warning data, and social media public opinion data in real time, wherein the POI attribute data at least includes a scenic spot type label, a user rating average, and an average visit duration;

[0035] A dynamic user portrait module is connected to the multi-source data fusion module and generates a user portrait containing time and space preference multidimensional features based on a neural network model by fusing the static attributes and dynamic behavior sequences input by the user.

[0036] a trip decision engine connected to the dynamic user portrait module, configured to generate an initial trip plan according to the user portrait features and trip constraint conditions, wherein the trip constraint conditions include total travel days, budget upper limit and geographical location restrictions;

[0037] a real-time optimization execution module, configured to track the actual location of the user through a positioning device and dynamically adjust the trip node parameters in combination with external environmental change data;

[0038] an interactive feedback module, configured to provide a visual operation interface and collect user behavior feedback to form a closed-loop optimization mechanism, wherein the multi-source data fusion module collects user behavior data (including click stream, favorite POI and trip rating) in real time through a distributed crawler system, synchronously accesses the meteorological bureau API to obtain heavy rain / tornado warning signals (yellow / orange / red level three), and calls the Gaode map API to update the traffic congestion index (0-10 level) every 2 minutes; the dynamic user portrait module adopts a double-channel neural network architecture: the short-term interest channel uses a time convolution network (TCN) to analyze the behavior sequence in the past 7 days to extract burst features, and the long-term preference channel models the user-POI historical interaction graph through a graph attention network (GAT); the trip decision engine outputs a POI interest matching degree score based on the user portrait, generates a candidate path set through an ant colony algorithm, and selects an initial plan with the lowest comprehensive loss through Monte Carlo tree search (MCTS) iteration 2000 times; the real-time optimization execution module triggers geographic fence monitoring using mobile phone GPS positioning data, dynamically re-plans the node sequence in combination with the real-time traffic prediction model when it detects that the user deviates from the preset path by more than 500 meters; the interactive feedback module renders a draggable 3D time axis interface on the Web end, triggers path feasibility verification and updates the plan immediately after the user adjusts the POI location.

[0039] (2) The dynamic user portrait module includes a short-term interest analysis unit, a long-term preference modeling unit, and a weight adaptive unit. The short-term interest analysis unit processes recent behavior data of the user through a time series neural network to identify burst interest features. The long-term preference modeling unit constructs an association graph of user behavior and POI attributes through a graph neural network to extract stable preference features. The weight adaptive unit dynamically allocates feature weights according to the time of behavior occurrence, wherein the weight of behavior data within 24 hours is not less than 0.8, and the weight value decays exponentially over time. In the dynamic user portrait module, the short-term interest analysis unit is configured to encode the POI types clicked by the user each day into a 128-dimensional vector, input a three-layer TCN network (convolution kernel size = 3), and output a burst interest feature vector. The long-term preference modeling unit constructs a user-POI bipartite graph, and the edge weight calculation formula is: collection times x 0.6 + stay time coefficient x 0.4. A 256-dimensional preference vector is generated after calculating the node attention weight through the GAT network. The weight adaptive unit sets the initial value of the real-time behavior (≤24 hours) weight to 0.8, and the weight of historical data decays exponentially by 5% per day. When the user is inactive for 3 consecutive days, the portrait cold update mechanism is started.

[0040] (3) The operation of the trip decision engine includes: in the candidate path generation stage, matching the POI feature set based on the user portrait, generating a candidate path set under time and space constraints through a swarm intelligence algorithm, in the path optimization stage, calculating the comprehensive loss value of each candidate path through a tree search algorithm, and selecting the path with the minimum loss value as the initial trip plan, wherein the comprehensive loss value is calculated by weighting the traffic time consumption, interest matching degree deviation and crowd density influence factor. In the candidate path generation stage of the trip decision engine: the ant colony pheromone is initialized as POI score x 0.01 (5-point system), the heuristic factor is the inverse of traffic time consumption, and the probability formula for ant selection of the next node is: pheromone concentration ^ α x heuristic factor ^ β (α = 1, β = 2). In the path optimization stage, Monte Carlo tree search is performed: a POI is randomly replaced to generate a child node from the root node, the utility value of the new path is calculated in the simulation stage = (POI interest matching degree mean x 0.7 - traffic loss x 0.3), the node visit count and average utility value are updated in the backtracking stage, and when the optimal path does not improve for 100 consecutive iterations, the scheme is terminated and output.

[0041] (4) The real-time optimization execution module includes: a flexible buffer controller assigns a dynamically adjustable dwell time interval to each trip node, the interval has a floating range of ±20% of the reference dwell time, a rescheduling trigger starts a node deletion program when external events cause the cumulative trip delay to exceed the buffer capacity, and a compensation mechanism automatically retrieves adjacent alternative POIs after deleting a node to maintain the integrity of the trip experience, wherein the flexible buffer controller of the real-time optimization execution module sets a floating interval (±20%) of the reference dwell time x 1.2 for parent-child POIs, and the reference time for the elderly user group is x 1.3 when the walking distance is >1 km; the rescheduling trigger continuously accumulates the delay value (delay = actual arrival time - planned time), and when the cumulative delay exceeds the total buffer time (∑(reference time x 0.2)), the node deletion program is started: sort according to the formula "priority = interest matching degree x 0.7 + user specified weight x 0.3", delete POIs from the end until the delay ≤ remaining buffer capacity; the compensation mechanism calls the adjacent POI retrieval algorithm (radius ≤ 2 km and interest matching degree ≥ 80% of the original node) to push alternative solutions.

[0042] (5) The system further includes a risk emergency unit: in the weather response subunit, when the accessed meteorological data reaches a preset risk level, automatically replace outdoor nodes with indoor alternatives, in the social event response subunit, analyze real-time text on social media through natural language processing technology to detect sudden events, and trigger a trip termination agreement and link to an insurance service interface, wherein the weather response sub-module of the risk emergency unit listens to the meteorological API, and when an orange rainstorm warning is returned, automatically matches the indoor alternative POI library (type label contains "museum" "shopping center" and walking distance ≤ 1.5 times the original node); the social event response sub-module analyzes Twitter / Weibo text in real time through the BERT model (keywords: "strike" "epidemic" "closed"), detects more than 3 related texts and sentiment value < -0.7, triggers a trip fuse, and calls the insurance interface to generate a claim report (including original trip ID, interruption time, and affected POI list).

[0043] (6) The interactive feedback module implements a visual editing function and a negative feedback learning channel. The visual editing function updates the schedule timing chart in real time and renders a three-dimensional topology path in response to user drag operations. The negative feedback learning channel reduces the weight of recommendations of the same category and generates a differentiated scheme set when the user continuously rejects recommendations of the same category to reach a threshold. The visual editing function of the interactive feedback module implements: after the user drags a POI icon to a new time slot, the system detects conflict nodes and calculates the traffic time increment in real time. If it exceeds the elastic buffer interval, it will be marked as red and warned. When the user rejects the same type of recommendation for 3 times, the weight of this POI category is multiplied by the decay coefficient 0.6, and the generation of the generative adversarial network (GAN) is started: the generator inputs the user vector after removing the category features, and outputs 5 candidate schemes. The discriminator ensures that the similarity between the scheme and the rejected category is less than 0.3 (cosine similarity threshold).

[0044] (7) The system deploys a privacy protection architecture: the data desensitization layer strips off the user identifier during data transmission and only retains the feature vector for model training. The noise addition layer injects random noise during the output of the trip scheme to ensure that the scheme cannot identify the user's identity in reverse. The privacy protection architecture deploys a federated learning framework: the user device locally trains the portrait model using TensorFlow Lite (dataset = local behavior log), and uploads encrypted gradient parameters to the cloud aggregation server every 24 hours. The cloud executes the FedAvg algorithm to aggregate the gradients of more than 100,000 devices and then issues a new model. Laplace noise is added before the output of the trip scheme (noise size = total utility value of the scheme x 0.05), ensuring that the privacy budget ε≤0.3.

[0045] (8) The system integrates a business intelligence extension component including a precision marketing unit and a resource reservation unit. The precision marketing unit dynamically pushes discount information within a limited radius based on user portraits and real-time location information on an electronic map. The resource reservation unit automatically triggers a protocol resource locking program when a user repeatedly accesses a specific service provider page more than a preset number of times. The precision marketing unit of the business intelligence extension module is configured with a geofence radius of 500 meters. When the user's real-time location enters the fence and the "dining consumption level" in the portrait is ≥4 stars (5-star system), a time-limited coupon for that POI is pushed (validity period = baseline stay duration x 1.5). The resource reservation unit monitors the user's access to hotel detail page frequency (>5 times / 24h) and automatically calls the reservation interface of the travel agency API (protocol price code = consumption level in user portrait x hotel star rating).

[0046] (9) System supports cross-city trip planning: Multi-destination concatenation unit analyzes inter-city traffic topology through space-time network model to generate optimal concatenation path. Transit optimizer is automatically inserted when cross-city travel time exceeds threshold, and subsequent travel time is adjusted synchronously. Cross-city trip planning models inter-city traffic topology through space-time graph neural network (STGNN): Node = city + transportation hub, edge weight = average time consumption of high-speed rail / flight; Transit optimizer detects when cross-city travel time > 4 hours, retrieves POIs (type labels include "cafe" and "park bench") within 1 km around the transit station, inserts time length = transit waiting gap x 0.6, and synchronously compresses subsequent node stay time (compresses from low to high priority, with a maximum compression rate of 20% for single node).

[0047] Embodiment one:

[0048] Stage 1: Real-time collection and fusion of multi-source data

[0049] 1. Data input

[0050] User input: age, number of companions, budget range, travel days, departure / destination

[0051]

[0052] Dynamic behavior collection: Real-time capture of user click / favorite POI records, historical trip score values through API interface

[0053]

[0054] Environmental data access:

[0055] Traffic status: Call Gaode Map API to get road congestion index (0-10 levels)

[0056] Weather warning: Access China Meteorological Administration rainstorm / typhoon warning signal (three levels of yellow / orange / red)

[0057] POI feature library: Get attraction type label, average stay time, and user rating average from the public platform of the Ministry of Culture and Tourism

[0058] 2. Data fusion rules

[0059] Time alignment: Align different frequency data to 5-minute time slices (e.g. traffic data 1

[0060] minute / piece, POI data static)

[0061] Spatial mapping: Match POI coordinates and user locations to 500m x 500m grid through Geohash encoding

[0062]

[0063] ​​​Stage 2: Dynamic user profile modeling

[0064] 1. Short-term interest capturing

[0065] Process the sequence of behaviors in the last 7 days with a Temporal Convolutional Network (TCN):

[0066] Input: Sequence of encoded POI types (e.g. [museum, playground, restaurant]) clicked by the user each day

[0067] Convolution: Extract 3-day consecutive behavior patterns (e.g. "consecutive clicks on family-friendly POIs")

[0068] Output: Burst interest feature vector (dimension 128)

[0069] 2. Long-term preference modeling

[0070] Construct a user-POI heterogeneous graph:

[0071] Nodes: User nodes + POI nodes (total 100k+)

[0072] Edge weights: Number of collections of this type of POI by the user × 0.6 + dwell time coefficient × 0.4

[0073] Compute preference weights with a Graph Attention Network (GAT):

[0074] Step 1: Compute feature similarity between user nodes and associated POI nodes

[0075] Step 2: Normalize similarity to attention weights (0~1)

[0076] Step 3: Aggregate weighted POI features to generate long-term preference vector (dimension 256)

[0077] 3. Profile fusion and decay

[0078] Fusion rule: Long-term preference vector × decay coefficient + short-term interest vector × real-time weight

[0079] Decay coefficient calculation:

[0080] Real-time data (within 24 hours): weight = 0.8

[0081] Historical data (30 days ago): weight = 0.8 × e^(-0.05 × day difference) / / exponential decay Stage 3: Trip decision engine generates initial plan

[0082] 1. Candidate path set generation

[0083] Ant colony algorithm parameter settings:

[0084] Pheromone initial value: POI user score x 0.01 (full score 5 points, pheromone = 0.05)

[0085] Heuristic factor: 1 / (traffic time consumption + walking distance)

[0086] Path construction process:

[0087] Step 1: Starting from the starting POI, the ant selects the next POI according to the pheromone concentration probability

[0088] Step 2: Stop expanding when cumulative time > user set total duration x 80%

[0089] Step 3: Repeat 1000 times to generate a candidate path set (keep Top50 utility value paths)

[0090] 2. Path optimization phase

[0091] Monte Carlo Tree Search (MCTS) four-step operation:

[0092] a. Selection: Select an unexplored child node (replace a POI) from the root node (initial path)

[0093] b. Expansion: Randomly add a new node (such as insert a candidate restaurant)

[0094] c. Simulation: Calculate the utility value of the new path (utility = total interest matching degree - traffic loss)

[0095] d. Backtracking: Update path node access frequency and average utility

[0096] Termination condition: 2000 iterations or 100 consecutive times without optimization improvement

[0097] Stage 4: Real-time dynamic optimization execution

[0098] 1. Elastic buffer control

[0099] Baseline stay duration adjustment:

[0100] Parent-child user group: baseline duration x 1.2

[0101] Older user group: baseline duration x 1.3 when walking distance > 1 km

[0102] Floating interval: baseline duration x [0.8, 1.2] / / i.e. ±20%

[0103] 2. Rescheduling trigger mechanism

[0104] Judgment logic:

[0105] IF (cumulative delay > total buffer duration) THEN

[0106] Start node pruning procedure:

[0107] Step 1: Calculate the priority of each POI = interest matching degree × 0.7 + user specified weight × 0.3

[0108] Step 2: Delete the node with the lowest priority until the delay ≤ remaining buffer capacity

[0109] ELSE

[0110] Compress the subsequent node stay duration dynamically (compress from low to high priority)

[0111] Phase 5: Human-computer collaborative closed-loop optimization

[0112] 1. Visual interaction rules

[0113] Drag response logic:

[0114] User drags POI_A from time slice [10:00-11:00] to [14:00-15:00] →

[0115] System detects conflict: If POI_B originally occupies [14:00-15:00], automatically exchange positions →

[0116] Recalculate traffic time consumption: If it exceeds the limit, mark it red and prompt

[0117] 2. Negative feedback reinforcement learning

[0118] Weight update rules:

[0119] When the user rejects "museum" type recommendations for 3 times:

[0120] The "cultural" weight in long-term preference = original weight × 0.6

[0121] Generate 5 differentiated schemes using GAN:

[0122] Generator input: User vector after removing cultural features Discriminator: Ensure that the similarity between the new scheme and the cultural POI is <0.3

[0123] Phase 6: Privacy protection and business expansion

[0124] 1. Federated learning implementation process

[0125] Step 1: Local training of user device portrait model (using local behavior data)

[0126] Step 2: Upload only model gradient parameters to cloud aggregation server

[0127] Step 3: Cloud issues aggregated global model to user device

[0128] Cycle period: aggregate once every 24 hours

[0129] 2. Preferential geofencing trigger rule

[0130] Condition 1: user real-time location distance from POI ≤ 500 meters

[0131] Condition 2: "dining consumption level" in user portrait ≥ 4 stars (5-star system)

[0132] Action: push a 100 yuan discount 20 yuan coupon (limited time 2 hours) for the POI

[0133] Finally, it should be pointed out that: first, in the description of the present application, it should be pointed out that unless otherwise specified and limited, the terms "installation", "connection", "connection" should be understood broadly, which can be mechanical connection or electrical connection, or the communication between two elements, or direct connection, "up", "down", "left", "right" and the like are only used to indicate the relative positional relationship, when the absolute position of the described object changes, the relative positional relationship may change;

[0134] Secondly: the drawings of the disclosed embodiments of the present application only involve the structures involved in the disclosed embodiments, other structures can refer to the usual design, and in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other;

[0135] Finally: the above only describes the preferred embodiments of the present application and does not limit the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. An intelligent travel itinerary automatic recommendation system and its implementation method, characterized in that: include: A multi-source data fusion module is configured to collect in real time user historical behavior data, POI attribute data in the geographic information system, real-time traffic network status data, weather warning data, and social media public opinion data, where the POI attribute data at least includes the attraction type label, user rating average, and average visit duration; The dynamic user profile module is connected to the multi-source data fusion module. By fusing the static attributes of user input with the dynamic behavior sequence, it generates a user profile containing multi-dimensional features of spatiotemporal preferences based on the neural network model. a trip decision engine, connected to the dynamic user profile module, configured to generate an initial trip plan based on user profile characteristics and trip constraints, wherein the trip constraints include the total number of travel days, budget cap, and geographic location restrictions; A real-time optimization execution module is configured to track the user's actual location through a positioning device and dynamically adjust the parameters of the trip nodes based on the external environment change data; The interactive feedback module is configured to provide a visual operation interface and collect user behavior feedback to form a closed-loop optimization mechanism.

2. The intelligent travel itinerary automatic recommendation system and its implementation method according to claim 1, characterized in that: The dynamic user portrait module includes a short-term interest analysis unit, a long-term preference modeling unit and a weight adaptation unit. The short-term interest analysis unit processes the user's recent behavior data through a time series neural network to identify sudden interest features. The long-term preference modeling unit constructs a correlation map between user behavior and POI attributes through a graph neural network to extract stable preference features. The weight adaptation unit dynamically assigns feature weights according to the time when the behavior occurs, where the weight of behavior data within 24 hours is not less than 0.8, and the weight value decays exponentially over time.

3. The intelligent travel itinerary automatic recommendation system and its implementation method according to claim 1, characterized in that: The operation of the itinerary decision engine includes: matching POI feature sets based on user profiles in the candidate route generation phase, generating a set of candidate routes under spatiotemporal constraints through a swarm intelligence algorithm, calculating the comprehensive loss value of each candidate route through a tree search algorithm in the route optimization phase, and selecting the route with the smallest loss value as the initial itinerary plan, where the comprehensive loss value is calculated by weighting the factors affecting traffic time, interest matching degree deviation, and crowd density.

4. The intelligent travel itinerary automatic recommendation system and its implementation method according to claim 1, characterized in that: The real-time optimization execution module includes: a flexible buffer controller that assigns a dynamically adjustable dwell time interval to each trip node, with the interval fluctuating within ±20% of the baseline dwell time; a rescheduling trigger that initiates a node deletion process when external events cause cumulative trip delays to exceed the buffer capacity; and a compensation mechanism that automatically searches for nearby alternative POIs after deleting a node to maintain the integrity of the trip experience.

5. The intelligent travel itinerary automatic recommendation system and its implementation method according to claim 1, characterized in that: The system also includes a risk emergency response unit: in the weather response sub-unit, when the incoming meteorological data reaches the preset risk level, the outdoor node is automatically replaced with an indoor alternative. In the social event response sub-unit, real-time text on social media is analyzed through natural language processing technology. When an emergency is detected, the trip termination agreement is triggered and the insurance service interface is linked.

6. The intelligent travel itinerary automatic recommendation system and its implementation method according to claim 1, characterized in that: The interactive feedback module implements a visual editing function and a negative feedback learning channel. The visual editing function updates the travel time sequence diagram in real time and renders the three-dimensional topological path in response to the user's drag operation. When the user continuously rejects the same type of recommendations and reaches a threshold, the negative feedback learning channel reduces the weight of the category recommendation and generates a differentiated solution set.

7. The intelligent travel itinerary automatic recommendation system and its implementation method according to claim 1, characterized in that: The system deploys a privacy protection architecture: the data desensitization layer removes user identity identifiers during the data transmission phase, retaining only feature vectors for model training. The noise addition layer injects random noise during the itinerary plan output phase to ensure that the plan cannot reversely identify the user's identity.

8. The intelligent travel itinerary automatic recommendation system and its implementation method according to claim 1, characterized in that: The system-integrated business intelligence extension component includes a precision marketing unit and a resource reservation unit. The precision marketing unit dynamically pushes discount information within a limited radius on the electronic map based on user portraits and real-time location information. The resource reservation unit automatically triggers the protocol resource locking program when a user repeatedly visits a specific service provider page more than a preset number of times.

9. The intelligent travel itinerary automatic recommendation system and its implementation method according to claim 1, characterized in that: The system supports cross-city itinerary planning: the multi-destination serial unit analyzes the traffic topology relationship between cities through the space-time network model to generate the optimal serial route. The transfer optimizer automatically inserts transfer rest nodes and synchronously adjusts the subsequent travel sequence when the cross-city travel time exceeds the threshold.

Images