Local characteristic delicious food intelligent recommendation and consumption map navigation system
By leveraging a distributed data processing engine, a dynamic recommendation engine, and a multimodal interaction interface, combined with real-time traffic data, the system addresses the issues of homogenized recommendations and fragmented navigation in existing systems. This enables personalized, real-time food recommendations and navigation in a closed loop, enhancing the user experience.
Patent Information
- Application Number
- CN202511218754.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-28
- Publication Date
- 2025-12-16
AI Technical Summary
Existing local specialty food recommendation systems suffer from problems such as homogenized recommendations, disconnect between navigation and services, and lagging information updates. They are unable to accurately match users' deep-seated needs for local specialty foods and lack real-time response capabilities and multimodal data utilization.
Employing a distributed data processing engine, a dynamic recommendation engine, a real-time navigation service module, and a multimodal interaction interface, it achieves a personalized recommendation and navigation closed loop by combining collaborative filtering and deep learning to create dynamic user profiles, along with real-time traffic data and multimodal interaction.
It enables personalized and accurate food recommendations, reduces homogenization, provides real-time route planning, improves user experience, and enhances the system's real-time performance and multimodal data utilization.
Smart Images
Figure CN121146860A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of navigation recommendation, and specifically relates to an intelligent recommendation and consumption map navigation system for local specialty foods. Background Technology
[0002] Currently, the technological background of intelligent recommendation and consumer map navigation systems for local specialty foods is rooted in the integration of widespread mobile internet access, improved location service accuracy, enhanced big data analytics capabilities, and advancements in artificial intelligence. Consumers have a strong demand for exploring localized, personalized, and culturally rich dining experiences, but existing technologies still face significant bottlenecks in meeting this demand. Existing systems largely rely on user historical ratings and basic merchant information (such as category and average price), severely lacking in deep semantic understanding and multi-dimensional feature extraction of the core attributes of food. This includes the inability to accurately analyze the strict geographical origin of ingredients, the inheritance of unique traditional cooking techniques, seasonal supply limitations, the historical and cultural allusions contained in the dishes, and subtle and authentic descriptions of flavor layers (such as the balance of sour and spicy, crispness, and the complexity of oil and aroma). As a result, recommended content tends to be superficial and homogenized, failing to accurately match users' deeper expectations for "local specialties" and unable to effectively distinguish between truly authentic flavors with unique cultural value and complex craftsmanship from commercially mass-produced products.
[0003] Existing data acquisition and update mechanisms suffer from severe lag. Key dynamic information such as merchant operating status (e.g., sudden closure, seasonal shutdown, temporary relocation), real-time menu changes (especially for seasonal dishes reliant on daily ingredients), estimated queue times, and customer traffic saturation rely heavily on manual reporting or infrequent platform review, lacking automated perception and verification capabilities at the minute or even second level. This results in a significant gap between the static POI database and real-world dynamic consumption scenarios, frequently leading to users arriving at locations only to find them closed or out of stock, severely damaging system credibility. Furthermore, the recommendation engine, navigation module, and consumption decision support functions (e.g., reservations, discounts, sharing) are often disconnected within the system architecture. Recommendation results lack deep integration with real-time traffic conditions, user location, personalized travel preferences (e.g., walking tolerance, public transport connections), and multi-target point (e.g., attractions, accommodation) collaborative route planning, forcing users to manually switch between multiple applications to piece together information, making it difficult to generate a one-stop closed-loop "recommendation-navigation-consumption" solution based on comprehensive contexts (e.g., time budget, companions, dietary restrictions).
[0004] Furthermore, the current human-computer interaction methods are rigid and monotonous, making it difficult to effectively handle complex, ambiguous, or emotionally charged natural language queries from users. They lack the ability for multi-turn conversational interactions to clarify intent and guide exploration, and fail to effectively integrate multimodal content such as images and short videos for intuitive decision-making assistance. Personalized recommendation models generally face the cold-start problem for new users or users with sparse behavior. Meanwhile, collaborative filtering algorithms that over-rely on historical behavioral similarity can easily lead to "information cocoons," limiting users' exploration of unknown but potentially interest-related culinary specialties. Some systems, in pursuit of commercial conversion, tend to have recommendation results excessively influenced by paid promotions, damaging fairness and user trust. In summary, existing technologies have significant shortcomings in the depth and breadth of food understanding, the dynamism and accuracy of information, the integration of service chains, the natural intelligence of interaction, and the fairness and exploratory nature of recommendations. Technological innovation is urgently needed to build a next-generation system that truly understands the essence of local cuisine, responds to changes in real time, seamlessly connects the entire process, intelligently understands intent, and stimulates the desire for exploration. Summary of the Invention
[0005] This invention proposes an intelligent recommendation and consumption map navigation system for local specialty foods, which solves the problems of homogenized recommendations, disconnect between navigation and services, and lagging information updates in existing food recommendation systems, and realizes personalized and accurate recommendations and dynamic navigation closed loops for local specialty foods.
[0006] The technical solution of the present invention is implemented as follows: a local specialty food intelligent recommendation and consumption map navigation system, characterized in that it includes a distributed data processing engine, a dynamic recommendation engine, a real-time navigation service module, and a multimodal interaction interface;
[0007] The distributed data processing engine processes user behavior data in real time and generates a dynamic, multi-dimensional user profile tag set.
[0008] The dynamic recommendation engine includes a collaborative filtering sub-model and a deep learning sub-model. The collaborative filtering sub-model calculates the correlation between users and food based on the user profile tag set, the deep learning sub-model mines the non-linear features of behavioral data, and the model fusion unit dynamically weights the output of the two sub-models to generate a recommendation list. The model parameters are optimized online through a real-time feedback mechanism.
[0009] The real-time navigation service module integrates a geographic information system, generates dynamic path planning schemes based on user location and recommended targets, performs path replanning in conjunction with real-time traffic data to avoid traffic obstacles, and provides real-scene navigation through an augmented reality interface.
[0010] The multimodal interaction interface includes: an image recognition unit and a natural language processing unit driven by a convolutional neural network, wherein the image recognition unit parses the visual information of the dishes and maps it to the image feature space; the natural language processing unit extracts the sentiment and flavor entities of the comment text and records them; the image recognition unit and the natural language processing unit output a feedback channel to the dynamic recommendation engine to correct the user profile weights.
[0011] By coordinating the data flow between the front end and the back end through a caching acceleration layer, closed-loop processing of behavioral data collection, profile updating, recommendation calculation, path planning and navigation is achieved.
[0012] Existing technologies suffer from four main shortcomings: First, recommendation systems rely on single historical behavioral data or static labels, lacking the ability to dynamically capture deep user preferences (such as flavor sensitivity and seasonal ingredient preferences), resulting in highly homogenized recommendation results and an inability to distinguish between industrialized fast food and truly regionally complex specialty foods. Second, navigation functions and recommendation services are completely disconnected, requiring users to manually switch between application-planned routes. Furthermore, traditional route planning does not integrate real-time traffic conditions (such as sudden construction or congestion) and personalized user travel constraints (such as walking tolerance and multi-point connection), leading to a disconnect between navigation results and actual scenarios. Third, the utilization rate of multimodal data (images and text) is low. Existing systems can only perform basic classification of food images or extract comment keywords, failing to establish semantic connections between visual features and flavor descriptions, and even less able to correct user profiles through cross-modal verification, resulting in biased recommendation weights. Fourth, the system response latency is significant, with minute-level delays in behavioral data updates, model optimization, and route replanning, failing to support users' real-time decision-making needs.
[0013] The technical challenges are: how to construct dynamically evolving user profiles to accurately quantify subjective dietary preferences; how to achieve deep coupling between recommendation models and navigation engines to generate a one-stop service loop; how to design a joint analysis framework for cross-modal data to solve semantic conflict problems; and how to ensure the real-time performance of the system with millisecond-level feedback optimization under a data scale of hundreds of millions.
[0014] As a preferred embodiment, the behavioral data processing flow of the distributed data processing engine specifically includes:
[0015] The behavior sequence parsing unit encodes user clicks, favorites, and order behaviors into spatiotemporally correlated explicit behavior feature vectors;
[0016] The implicit preference inference unit extracts implicit preference features derived from the frequency and duration of behavioral intervals through temporal pattern analysis;
[0017] The dynamic profile building unit orthogonally fuses explicit behavioral feature vectors with implicit preference features to generate a multi-dimensional tag set that includes flavor sensitivity, price tolerance range, and regional flavor preference, and dynamically evolves the tag weights through a time decay factor.
[0018] In a preferred embodiment, the model fusion unit of the dynamic recommendation engine includes: a weight allocator that dynamically allocates the output weights of the collaborative filtering sub-model and the deep learning sub-model based on the freshness of user behavior data and the scene context; a feature concatenation layer that concatenates the feature vectors output by the two sub-models after dimensional alignment, and generates a recommendation score through a fully connected network; and a real-time feedback mechanism that includes a parameter incremental update module that monitors real-time user interaction behavior. When a recommendation feedback deviation is detected to exceed a threshold, the module triggers fine-tuning of the model parameters and locks high-confidence weight nodes in the fully connected network to prevent overfitting.
[0019] As a preferred implementation, the real-time navigation service module generates an initial navigation path based on the user-defined maximum walking distance, public transportation connection signs, and recommended target priority sequence; the dynamic obstacle avoider continuously receives traffic status event streams, and when it detects construction or congestion events overlapping with the current path, it initiates a local topology update and retains valid nodes of the original path;
[0020] In a preferred embodiment, the feedback channel of the multimodal interaction interface maps the visual feature vector of the dish and the entity vector of the flavor description to the shared embedding space through the feature association module; the weight regulator calculates the credibility of the flavor attribute based on the sentiment polarity value, and when the similarity between the visual feature and the text description in the shared space exceeds a preset value, it triggers the weight enhancement of the corresponding flavor preference in the user profile; when the difference processing module detects a contradiction between the visual feature and the text sentiment, it prioritizes using high-frequency user behavior data to cover the low credibility modal input.
[0021] After adopting the above technical solution, the beneficial effects of the present invention are as follows: Based on the collaborative filtering and deep learning hybrid recommendation mechanism of dynamic multi-dimensional user profile tag set, it breaks through the linear correlation limitation of traditional models, mines non-linear features in behavioral data (such as the interaction between short-term impulsive attempts and long-term taste preferences) through deep learning sub-models, and optimizes parameters online in combination with real-time feedback mechanism, so that the recommendation list accurately matches the user's personalized needs for regional characteristic ingredients, traditional cooking techniques and niche flavor combinations, and reduces the homogenization rate of recommendations;
[0022] The recommendation and navigation functions are deeply integrated. Based on the geographic information system, the route planning scheme is dynamically generated. Real-time traffic data is used to detect traffic obstacles (such as road closures and scenic area flow restrictions) and trigger millisecond-level route replanning. Combined with the augmented reality interface, real-scene navigation is realized. The recommended target POI and traffic information are superimposed on the camera screen, reducing the average time for users to reach the target merchant.
[0023] An innovative multimodal feedback channel is designed, which uses a convolutional neural network to analyze the visual features of dishes (such as the gloss of oil and the texture of cut ingredients) and map them to the feature space. The natural language processing unit extracts the sentiment polarity of comments and flavor entities. The two modal outputs are validated for consistency in a shared semantic space and the profile weights are corrected. This solves the problem of insufficient credibility of a single data source and improves the accuracy of flavor preference recognition.
[0024] By constructing a low-latency closed-loop system through a distributed data processing engine and a caching acceleration layer, the system achieves millisecond-level response times across the entire process of behavioral data collection, profile updates, recommendation calculations, and path planning, significantly improving the user experience in high-concurrency scenarios. The overall technical solution breaks down the inherent boundaries between recommendation, navigation, and interaction, forming a self-evolving intelligent service ecosystem. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 This is a system flowchart of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] Example:
[0029] like Figure 1 As shown, the intelligent recommendation and consumption map navigation system for local specialty foods is characterized by including a distributed data processing engine, a dynamic recommendation engine, a real-time navigation service module, and a multimodal interaction interface.
[0030] The distributed data processing engine processes user behavior data in real time and generates a dynamic, multi-dimensional user profile tag set.
[0031] The dynamic recommendation engine includes a collaborative filtering sub-model and a deep learning sub-model. The collaborative filtering sub-model calculates the correlation between users and food based on the user profile tag set, the deep learning sub-model mines the non-linear features of behavioral data, and the model fusion unit dynamically weights the output of the two sub-models to generate a recommendation list. The model parameters are optimized online through a real-time feedback mechanism.
[0032] The real-time navigation service module integrates a geographic information system, generates dynamic path planning schemes based on user location and recommended targets, performs path replanning in conjunction with real-time traffic data to avoid traffic obstacles, and provides real-scene navigation through an augmented reality interface.
[0033] The multimodal interaction interface includes: an image recognition unit and a natural language processing unit driven by a convolutional neural network, wherein the image recognition unit parses the visual information of the dishes and maps it to the image feature space; the natural language processing unit extracts the sentiment and flavor entities of the comment text and records them; the image recognition unit and the natural language processing unit output a feedback channel to the dynamic recommendation engine to correct the user profile weights.
[0034] By coordinating the data flow between the front end and the back end through a caching acceleration layer, closed-loop processing of behavioral data collection, profile updating, recommendation calculation, path planning and navigation is achieved.
[0035] Before tourists enter the target city's food street area, the system captures their behavioral data stream in real time through a distributed data processing engine. This includes interactive behaviors such as zooming in and out of the map interface, dwell time while browsing the merchant list, and click sequences on preview images of featured dishes. The behavior sequence parsing unit encodes these spatiotemporally continuous operations into explicit feature vectors (such as regional flavor attention coefficients and price-sensitive range identifiers). Simultaneously, the implicit preference inference unit extracts latent preference features (such as a preference for soups in the morning and a preference for fried foods at night) by analyzing the millisecond intervals between adjacent operations and the daily time distribution patterns. The dynamic profile construction unit integrates explicit and implicit features to generate a multi-dimensional tag set, initially relying on collaborative filtering. The model matches the historical preferences of nearby users based on label similarity and outputs a basic relevance score. At the same time, the deep learning sub-model analyzes the non-linear patterns in the behavioral sequence through a convolutional neural network (such as skipping Sichuan restaurants three times in a row and then choosing Cantonese dim sum) to generate a deep feature vector. The weight allocator of the model fusion unit dynamically increases the weight of the deep learning sub-model based on the freshness of the behavioral data (such as the latest operation occurring within 30 seconds) and the scene context (GPS positioning shows that it is located at the entrance of the snack street). The feature splicing layer reduces the dimensionality of the two feature vectors and aligns them before inputting them into the fully connected network to generate a list of merchants sorted in descending order of recommendation score, with the old-brand soup dumpling restaurant being the top recommendation, accompanied by the special tag "crab roe xiaolongbao".
[0036] The real-time navigation service module calls the geographic information system interface, taking the tourist's current location as the origin and the recommended destination, Tangbao Restaurant, as the destination. The multi-constraint path generator generates an initial walking path based on a preset walking tolerance threshold (800 meters) and public transportation connection signs (subway prohibited). The dynamic obstacle avoidance device simultaneously monitors traffic event flow. After detecting a municipal construction event 200 meters ahead, it initiates a local topology update algorithm to retain the valid nodes of the original path to the nearest intersection and replans an alternative route through the alley. The augmented reality interface fuses visual SLAM data through the smartphone's motion sensor and overlays dynamic directional arrows on the camera image. When the device detects that the user is turning into an alley, the POI dynamic fusion unit renders a virtual label for the Tangbao Restaurant sign in the real scene and adjusts the label's transparency through an occlusion perception algorithm to avoid obscuring the actual road conditions.
[0037] After arriving at the soup dumpling restaurant, tourists take photos of the "crab roe xiaolongbao" dish and upload them. The multimodal interaction interface initiates collaborative processing: the image recognition unit driven by the convolutional neural network extracts visual features such as the bamboo texture on the surface of the steamer and the light transmittance of the soup, and maps them to the image feature space to generate feature vectors; the natural language processing unit simultaneously scans the merchant's recent reviews, extracting flavor entities such as "fresh and sweet without being greasy" and "plump crab meat" as well as positive sentiment tendencies; the feature association module projects the visual vector and the text flavor vector to the shared embedding space, calculates the cosine similarity to 0.87, triggers the weight regulator, and enhances the weight of the "seafood acceptance" tag in the user profile; when the tourist subsequently gives a negative review to the "fresh meat xiaolongbao" ("the filling is greasy"), the difference processing module detects the contradiction between the visual features (high oil reflectivity) and the text sentiment (negative), and initiates a high-frequency behavior verification mechanism. Based on the behavior data of the tourist who frequently chooses steamed dishes in their historical orders, it corrects the interference of the current modal input on the "meat preference" tag.
[0038] The system's closed-loop processing achieves millisecond-level response through a caching acceleration layer: When the front-end Vue3 component loads the AR navigation module on demand, Redis caches and prefetches intermediate path planning data; when the real-time feedback mechanism detects negative reviews, the parameter incremental update module locks the high-confidence "cooking method preference" weight node in the fully connected network, only fine-tuning the "ingredient category" related parameters, and the model optimization results are broadcast to all service nodes via a distributed message queue; the updated recommendation list responds immediately, prioritizing the display of non-fried dishes such as "steamed ribbonfish" that match the new profile in the visitor's next planning step, while the path planning scheme is automatically associated with nearby recommendation targets, forming a continuous optimization closed loop of "behavior collection - profile update - recommendation calculation - path planning - real-world guidance". The entire process, under a cloud-native microservice architecture, achieves elastic scaling of each module through a service mesh, ensuring end-to-end latency of hundreds of milliseconds in high-concurrency scenarios.
[0039] The behavioral data processing flow of the distributed data processing engine specifically includes: a behavioral sequence parsing unit encodes user clicks, favorites, and order behaviors into spatiotemporally correlated explicit behavioral feature vectors; an implicit preference inference unit extracts implicit preference features derived from the frequency and duration of behavioral intervals through temporal pattern analysis; and a dynamic profile construction unit orthogonally fuses the explicit behavioral feature vectors and implicit preference features to generate a multi-dimensional tag set containing flavor sensitivity, price tolerance range, and regional flavor preference, and dynamically evolves the tag weights through a time decay factor.
[0040] When tourists enter a traditional food street in the city, the system captures their map zoom focus area, merchant list browsing sequence, and menu preview click behavior in real time. The behavior sequence analysis unit encodes the timestamps and spatial coordinates of user operations into spatiotemporally correlated explicit feature vectors, and generates structured behavioral representations by combining merchant categories and price ranges. The implicit preference inference unit analyzes the time interval distribution and behavior continuity patterns between consecutive operations, extracts potential preferences for specific cooking methods from high-frequency, short-interval operations, and infers flavor preference patterns by combining time period characteristics. The dynamic profile construction unit uses an orthogonal fusion algorithm to eliminate the linear correlation between explicit behavioral features and implicit preference features, generating a multi-dimensional label set that includes flavor sensitivity dimensions, price acceptance ranges, and regional flavor preferences. It also dynamically reduces the weight contribution of historical behaviors through a time decay mechanism to ensure that the profile continuously evolves with the user's real-time behavior.
[0041] The model fusion unit of the dynamic recommendation engine includes: a weight allocator that dynamically allocates the output weights of the collaborative filtering sub-model and the deep learning sub-model based on the freshness of user behavior data and scene context; a feature concatenation layer that concatenates the feature vectors output by the two sub-models after dimensional alignment, and generates a recommendation score through a fully connected network; and a real-time feedback mechanism that includes a parameter incremental update module that monitors real-time user interaction behavior. When a recommendation feedback deviation is detected to exceed a threshold, the module triggers fine-tuning of the model parameters and locks high-confidence weight nodes in the fully connected network to prevent overfitting.
[0042] When tourists are located in high-density food areas, the weight allocator dynamically increases the decision weights of the deep learning sub-model based on scene features such as the user's last operation time and current location to capture non-linear preference mutations. The feature concatenation layer performs dimensionality reduction and alignment on the user-food association vector output by the collaborative filtering sub-model and the behavioral pattern feature vector extracted by the deep learning sub-model, forming a hybrid feature representation input to the fully connected network. The fully connected network outputs a list of merchants sorted by recommendation score. When the negative feedback intensity of the user's recommendation results exceeds a preset threshold, the parameter incremental update module locks the high-confidence weight nodes in the model and performs gradient fine-tuning only for specific preference dimensions, completing model iteration and outputting the updated recommendation list in a very short time.
[0043] The real-time navigation service module generates an initial navigation path based on the user-defined maximum walking distance, public transportation connection signs, and recommended target priority sequence. The dynamic obstacle avoidance device continuously receives traffic status event streams. When it detects construction or congestion events that overlap with the current path, it initiates a local topology update and retains valid nodes of the original path.
[0044] After tourists select recommended target merchants, a multi-constraint path generator generates an initial optimal walking path based on preset walking distance limits, public transportation usage signs, and a priority sequence of multiple recommended targets. A dynamic obstacle avoidance system continuously receives urban traffic event stream data. When construction closures or congestion events are detected that spatially overlap with the current navigation path, a local path update algorithm is activated to preserve the connectivity of verified path nodes and replan the detour route based on the nearest valid node. The augmented reality navigation interface integrates motion sensor data and visual positioning information, dynamically overlaying path indicator arrows and virtual merchant signs onto the mobile device's camera view. A spatial occlusion handling algorithm ensures visual consistency between the navigation information and the real-world scene.
[0045] The feedback channel of the multimodal interaction interface maps the visual feature vector of the dish and the entity vector of the flavor description to the shared embedding space through the feature association module; the weight regulator calculates the credibility of the flavor attribute based on the sentiment polarity value. When the similarity between the visual feature and the text description in the shared space exceeds a preset value, the weight enhancement of the corresponding flavor preference in the user profile is triggered; when the difference processing module detects a contradiction between the visual feature and the text sentiment, it prioritizes using high-frequency user behavior data to cover the low credibility modal input.
[0046] When a visitor uploads a dish image and triggers comment analysis, the feature association module projects the visual feature vector extracted by the image recognition unit and the flavor description entity vector parsed by the natural language processing unit into a shared semantic space. The weight adjuster, based on a composite evaluation of cross-modal feature similarity and sentiment polarity, increases the weight coefficient of the corresponding flavor attribute in the user profile when the consistency between visual features and text descriptions in the semantic space exceeds a baseline value. If a semantic conflict is detected between the dish's visual feature recognition results (e.g., high oiliness and shine) and the text sentiment (e.g., negative reviews), the difference processing module prioritizes using high-frequency historical user behavior data to cover the influence of low-confidence modal inputs, maintaining the stability of the core preferences in the user profile.
[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A local specialty food intelligent recommendation and consumption map navigation system, characterized in that: It includes a distributed data processing engine, a dynamic recommendation engine, a real-time navigation service module, and a multimodal interaction interface; The distributed data processing engine processes user behavior data in real time and generates a dynamic, multi-dimensional user profile tag set. The dynamic recommendation engine includes a collaborative filtering sub-model and a deep learning sub-model. The collaborative filtering sub-model calculates the correlation between users and food based on the user profile tag set, the deep learning sub-model mines the non-linear features of behavioral data, and the model fusion unit dynamically weights the output of the two sub-models to generate a recommendation list. The model parameters are optimized online through a real-time feedback mechanism. The real-time navigation service module integrates a geographic information system, generates dynamic path planning schemes based on user location and recommended targets, performs path replanning in conjunction with real-time traffic data to avoid traffic obstacles, and provides real-scene navigation through an augmented reality interface. The multimodal interaction interface includes: an image recognition unit and a natural language processing unit driven by a convolutional neural network, wherein the image recognition unit parses the visual information of the dishes and maps it to the image feature space; the natural language processing unit extracts the sentiment and flavor entities of the comment text and records them; the image recognition unit and the natural language processing unit output a feedback channel to the dynamic recommendation engine to correct the user profile weights. By coordinating the data flow between the front end and the back end through a caching acceleration layer, closed-loop processing of behavioral data collection, profile updating, recommendation calculation, path planning and navigation is achieved.
2. The intelligent recommendation and consumption map navigation system for local specialty foods as described in claim 1, characterized in that: The behavioral data processing flow of the distributed data processing engine specifically includes: The behavior sequence parsing unit encodes user clicks, favorites, and order behaviors into spatiotemporally correlated explicit behavior feature vectors; The implicit preference inference unit extracts implicit preference features derived from the frequency and duration of behavioral intervals through temporal pattern analysis; The dynamic profile building unit orthogonally fuses explicit behavioral feature vectors with implicit preference features to generate a multi-dimensional tag set that includes flavor sensitivity, price tolerance range, and regional flavor preference, and dynamically evolves the tag weights through a time decay factor.
3. The intelligent recommendation and consumption map navigation system for local specialty foods as described in claim 1, characterized in that: The model fusion unit of the dynamic recommendation engine includes: a weight allocator that dynamically allocates the output weights of the collaborative filtering sub-model and the deep learning sub-model based on the freshness of user behavior data and scene context; a feature concatenation layer that concatenates the feature vectors output by the two sub-models after dimensional alignment, and generates a recommendation score through a fully connected network; and a real-time feedback mechanism that includes a parameter incremental update module that monitors real-time user interaction behavior. When a recommendation feedback deviation is detected to exceed a threshold, the module triggers fine-tuning of the model parameters and locks high-confidence weight nodes in the fully connected network to prevent overfitting.
4. The intelligent recommendation and consumption map navigation system for local specialty foods as described in claim 1, characterized in that: The real-time navigation service module generates an initial navigation path based on the user-defined maximum walking distance, public transportation connection signs, and recommended target priority sequence. The dynamic obstacle avoidance device continuously receives traffic status event streams, and when it detects construction or congestion events that overlap with the current path, it initiates a local topology update and retains valid nodes of the original path.
5. The intelligent recommendation and consumption map navigation system for local specialty foods as described in claim 1, characterized in that: The feedback channel of the multimodal interaction interface maps the visual feature vector of the dish and the entity vector of the flavor description to the shared embedding space through the feature association module; the weight regulator calculates the credibility of the flavor attribute based on the sentiment polarity value. When the similarity between the visual feature and the text description in the shared space exceeds a preset value, the weight enhancement of the corresponding flavor preference in the user profile is triggered; when the difference processing module detects a contradiction between the visual feature and the text sentiment, it prioritizes using high-frequency user behavior data to cover the low credibility modal input.