Travel product intelligent recommendation method and device, equipment and medium

Abstract

The present application relates to a travel product intelligent recommendation method, device, equipment and medium, the method comprises the following steps: collecting multi-dimensional travel data of a user, performing feature fusion modeling on the multi-dimensional travel data to obtain a user portrait; obtaining multi-dimensional label data of a travel product, performing dimension hierarchical archiving on the multi-dimensional label data to obtain a product resource pool; inputting the user portrait and the product resource pool into a double-engine hybrid recommendation model for bidirectional matching recall to generate a candidate product set; obtaining a dynamic context factor, performing weight correction and multi-objective balanced sorting on the candidate product set based on the factor, and outputting a final recommendation queue. The present application can improve matching accuracy and scene adaptability by combining dynamic context correction and multi-objective balanced sorting through a double-engine hybrid recommendation model, and can balance popular and long-tail product coverage and multi-dimensional operation index collaborative optimization.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent recommendation, and in particular to a method, apparatus, device, and medium for intelligent recommendation of travel products. Background Technology

[0002] With the increasing demand for high-concurrency matching of massive numbers of users and all categories of travel products on online travel platforms, achieving intelligent recommendation that balances personalization accuracy, dynamic response, and multi-objective equilibrium has become a key research focus. Existing solutions typically construct shallow user profiles based on only a few static fields such as age and region, failing to integrate temporal behavior and implicit preference data for multi-dimensional feature modeling. This results in insufficient profile identification and severe homogenization of recommendations. In the recall stage, existing solutions mostly employ a single collaborative filtering algorithm, lacking deep learning representation capabilities. This leads to recall failure when faced with long-tail travel products with sparse interaction samples, resulting in unsold and stockpiled long-tail products. Furthermore, existing solutions lack the ability to perceive dynamic contextual factors such as time, space, weather, and geographical location, causing recommended content to be disconnected from users' actual travel scenarios. In addition, the cold start handling for newly registered users with no prior activity is simplistic, directly downgrading to random blind recommendation, resulting in a poor experience for new customers. In the ranking stage, existing solutions mostly use a single metric for ranking, failing to coordinate and balance multiple operational indicators such as user experience, transaction conversion, and inventory turnover, making it difficult to consider the overall operational efficiency of the platform. Summary of the Invention

[0003] The main objective of this invention is to provide a method, device, equipment, and medium for intelligent recommendation of travel products. By using multi-dimensional data modeling and dual-engine hybrid recall, combined with dynamic context correction and multi-objective balanced ranking, it improves matching accuracy and scenario adaptability, while taking into account the coverage of popular and long-tail products and the collaborative optimization of multi-dimensional business indicators.

[0004] To achieve the above objectives, the present invention provides a method for intelligent recommendation of travel products, comprising: Collect multidimensional travel data from users, perform feature fusion modeling on the multidimensional travel data, and obtain user profiles; Obtain multidimensional tag data of travel products, and archive the multidimensional tag data by dimension layering to obtain a product resource pool; The user profile and the product resource pool are input into a dual-engine hybrid recommendation model for bidirectional matching and recall to generate a candidate product set. Obtain dynamic context factors, perform weight correction and multi-objective balanced ranking on the candidate product set based on the factors, and output the final recommendation queue.

[0005] Furthermore, the process of collecting multidimensional travel data from users and performing feature fusion modeling on the multidimensional travel data to obtain a user profile includes: Collect users' static attribute data and behavioral tracking data, and perform anomaly denoising and missing feature completion on the static attribute data and behavioral tracking data to obtain multi-dimensional travel data; The multidimensional travel data is decomposed into explicit essential needs features and implicit preference features, and a user profile is constructed by weighted fusion of the explicit essential needs features and the implicit preference features. The user profile supports real-time lightweight incremental updates. When new clicks, searches, or add-to-cart actions are collected from users, the feature weights of the user profile are fine-tuned accordingly to keep the profile synchronized with the user's real-time travel decision-making needs.

[0006] Furthermore, the step of acquiring multidimensional tag data for travel products, and then performing dimensional hierarchical archiving of the multidimensional tag data to obtain a product resource pool includes: Obtain the multidimensional tag data of all travel products listed on the platform; The multidimensional tag data is divided into basic attribute tags, rigid tags for audience adaptation, and dynamic real-time operation tags. The basic attribute tags are categorized and archived to generate a product basic attribute index; The rigid tags for audience adaptation are automatically matched and labeled to generate a product audience adaptation index; The dynamic real-time operation tags are synchronized and labeled in real time to generate a product real-time operation status index; The product basic attribute index, the product audience adaptation index, and the product real-time operation status index are jointly encapsulated to construct a multi-dimensional hierarchical tag system, forming the product resource pool.

[0007] Further, the step of inputting the user profile and the product resource pool into a dual-engine hybrid recommendation model for bidirectional matching and recall to generate a candidate product set includes: The user profile and the product resource pool are input into the dual-engine hybrid recommendation model, which includes a collaborative filtering sub-engine and a deep learning representation sub-engine that are launched in parallel. The collaborative filtering sub-engine matches the user profile with popular products in the product resource pool to generate a first candidate product subset. The deep learning representation sub-engine is used to mine latent features and construct vector representations for the user profile and the product resource pool to generate a second candidate product subset. The first subset of candidate products and the second subset of candidate products are weighted and merged to generate a candidate product set.

[0008] Further, the step of matching the user profile with the product resource pool using the collaborative filtering sub-engine to generate a first candidate product subset includes: The interaction co-occurrence calculation layer of the collaborative filtering sub-engine performs co-occurrence analysis on the travel preference features in the user profile and the historical interaction data of other users in the product resource pool to determine the similar user group in the same circle of the user image. The collaborative filtering sub-engine filters interactive travel products from the product resource pool that belong to similar user groups in the same circle, and sorts them by interaction frequency to generate the first subset of candidate products.

[0009] Furthermore, the step of generating a second candidate product subset by mining latent features and constructing vector representations of the user profile and the product resource pool through the deep learning representation sub-engine includes: The user profile and the product resource pool are input into the latent feature mining layer of the deep learning representation sub-engine, and the latent feature mining layer is used to mine the latent travel association features between the user profile and the product resource pool; The implicit travel-related features are input into the vector representation construction layer of the deep learning representation sub-engine to construct vectors, and are represented as user feature vectors and product feature vectors. The similarity between the user feature vector and each product feature vector is calculated through the long-tail product recall layer. Based on the similarity, the target travel products in the product resource pool are recalled to generate the second candidate product subset.

[0010] Further, the step of obtaining dynamic context factors, performing weight correction and multi-objective balanced ranking on the candidate product set based on the factors, and outputting the final recommendation queue includes: Obtain dynamic context factors, which include time-series scene factors, meteorological environment factors, geographical location factors, and passenger flow and business type factors; The initial weights of each travel product in the candidate product set are adjusted based on the time-series scenario factors to obtain the first corrected candidate product set; Based on the meteorological environmental factors and the passenger flow and business type factors, the first correction candidate product set is adjusted for environmental weights to obtain the second correction candidate product set. Based on the geographical location factor, the second corrected candidate product set is adjusted by distance weight to obtain the third corrected candidate product set; Based on preset multi-dimensional business indicators, the third set of candidate products for correction is sorted by target collaborative equilibrium, and the final recommendation queue is output.

[0011] The present invention also provides a smart travel product recommendation device, applied to any one of the smart travel product recommendation methods described above, comprising: A multidimensional data acquisition module is used to collect multidimensional travel data of users, perform feature fusion modeling on the multidimensional travel data, and obtain user profiles. The product tag archiving module is used to acquire multi-dimensional tag data of travel products, perform dimensional hierarchical archiving of the multi-dimensional tag data, and obtain a product resource pool. A hybrid recommendation recall module is used to input the user profile and the product resource pool into a dual-engine hybrid recommendation model for bidirectional matching and recall, generating a candidate product set. The context correction and sorting module is used to obtain dynamic context factors, perform weight correction and multi-objective balanced sorting on the candidate product set based on the factors, and output the final recommendation queue.

[0012] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the intelligent travel product recommendation method as described in any of the preceding claims.

[0013] The present invention also provides a readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent travel product recommendation method as described in any one of claims 1 to 7.

[0014] The present invention provides a method, apparatus, device, and medium for intelligent recommendation of travel products, which has the following beneficial effects: By collecting multi-dimensional travel data and integrating static attributes, temporal behaviors, and implicit preferences to build user profiles through feature modeling, the problem of insufficient shallow profile identification in existing solutions is solved, improving the granularity and matching accuracy of the profiles. By acquiring multi-dimensional tag data of travel products and establishing a product resource pool through dimensional hierarchical archiving, the shortcomings of existing solutions in terms of single product features and inability to perform refined indexing are overcome. A dual-engine hybrid recommendation model is used to perform bidirectional matching and recall between user profiles and the product resource pool, solving the problem of long-tail product recall failure in single collaborative filtering algorithms under sparse data, achieving coverage of both popular and long-tail products. By acquiring dynamic context factors and performing weight correction on the candidate product set, linking temporal scenarios, weather conditions, geographical location, and passenger flow information, the shortcomings of existing solutions in lacking context awareness, leading to a disconnect between recommendations and travel scenarios, are overcome, improving the adaptability of recommended content to real travel scenarios. By performing multi-objective balanced ranking on the corrected candidate product set and coordinating optimization according to multi-dimensional business indicators, the problem of single-indicator ranking being unable to meet the needs of multiple parties is solved, achieving a comprehensive balance of user experience, transaction conversion, and inventory turnover. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention 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.

[0016] Figure 1 This is a schematic diagram of the structure of an intelligent recommendation system for travel products according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall process of a smart recommendation method for travel products in one embodiment of the present invention; Figure 3 This is a schematic diagram of the dimensional hierarchical archiving process in the intelligent recommendation method for travel products in one embodiment of the present invention; Figure 4 This is a schematic diagram of the bidirectional matching and recall process in the intelligent recommendation method for travel products according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the feature mining process in the intelligent recommendation method for travel products according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a smart travel product recommendation device according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention.

[0017] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] 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, not all, of the embodiments of the present invention. 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.

[0019] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. It should also be understood that, as used in this specification and the appended claims, the term "and / or" refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0020] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0022] It should be understood that the sequence number of each step in the following embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0023] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0024] With the increasing demand for high-concurrency matching of massive numbers of users and all categories of travel products on online travel platforms, achieving intelligent recommendation that balances personalization accuracy, dynamic response, and multi-objective equilibrium has become a key research focus. Existing solutions typically construct shallow user profiles based on only a few static fields such as age and region, failing to integrate temporal behavior and implicit preference data for multi-dimensional feature modeling. This results in insufficient profile identification and severe homogenization of recommendations. In the recall stage, existing solutions mostly employ a single collaborative filtering algorithm, lacking deep learning representation capabilities. This leads to recall failure when faced with long-tail travel products with sparse interaction samples, resulting in unsold and stockpiled long-tail products. Furthermore, existing solutions lack the ability to perceive dynamic contextual factors such as time, space, weather, and geographical location, causing recommended content to be disconnected from users' actual travel scenarios. In addition, the cold start handling for newly registered users with no prior activity is simplistic, directly downgrading to random blind recommendation, resulting in a poor experience for new customers. In the ranking stage, existing solutions mostly use a single metric for ranking, failing to coordinate and balance multiple operational indicators such as user experience, transaction conversion, and inventory turnover, making it difficult to consider the overall operational efficiency of the platform.

[0025] To address the aforementioned issues, this application proposes an intelligent recommendation method, apparatus, device, medium, and program product for travel products. By collecting multi-dimensional travel data and integrating static attributes, temporal behaviors, and implicit preferences to construct user profiles through feature modeling, it solves the problem of insufficient shallow profile identification in existing solutions, improving profile refinement and matching accuracy. By acquiring multi-dimensional tag data of travel products and establishing a product resource pool through dimensional hierarchical archiving, it overcomes the shortcomings of existing solutions in terms of single product features and inability to perform refined indexing. Through a dual-engine hybrid recommendation model, it performs bidirectional matching and recall between user profiles and the product resource pool, solving the problem of long-tail product recall failure in single collaborative filtering algorithms under sparse data, achieving comprehensive coverage of both popular and long-tail products. By acquiring dynamic context factors and performing weight correction on the candidate product set, it links temporal scenarios, weather conditions, geographical location, and passenger flow information, overcoming the shortcomings of existing solutions that lack context awareness, leading to a disconnect between recommendations and travel scenarios, and improving the adaptability of recommended content to real travel scenarios. By performing multi-objective equilibrium ranking on the corrected candidate product set and coordinating optimization according to multi-dimensional business indicators, the problem that single-indicator ranking cannot take into account the demands of multiple parties is solved, and a comprehensive balance of user experience, transaction conversion and inventory turnover is achieved.

[0026] The intelligent travel product recommendation method provided in this invention can be applied to, for example... Figure 1 The intelligent travel product recommendation system shown includes a client and an intelligent travel product recommendation device, wherein the terminal device communicates with the server via a network or bus.

[0027] The intelligent travel product recommendation device can be a terminal device, which refers to a device that corresponds to a server and provides local services to customers. This terminal device includes, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices.

[0028] The intelligent travel product recommendation device can also be a server, which can be implemented using a standalone server or a server cluster composed of multiple servers.

[0029] Reference Figure 2 As shown, the present invention provides a method for intelligent recommendation of travel products, including: Step S100: Collect multidimensional travel data of users, perform feature fusion modeling on the multidimensional travel data, and obtain user profiles; Specifically, multi-dimensional travel data of users is collected in compliance with regulations through a unified multi-terminal tracking SDK. The multi-dimensional travel data includes basic compliant static attribute data, full-link time-series dynamic behavior tracking data, and historical travel rating text. The basic compliant static attribute data includes age group stratification, gender tags, cities of residence, consumption level, and travel companion preferences. The full-link time-series dynamic behavior tracking data includes travel channel search keywords, product click browsing trajectory, single product page dwell time, multi-product horizontal comparison behavior, positive feedback such as adding to cart, favorites, and sharing, and negative feedback such as browsing without placing an order and short-term exit.

[0030] The collected basic compliance static attribute data, the full-link time-series dynamic behavior tracking data, and the historical travel rating text undergo standardized preprocessing, including abnormal traffic identification and filtering, duplicate data deduplication, and missing field completion. The preprocessed data is then merged into multidimensional travel data. This multidimensional travel data is input into a multidimensional feature fusion weighted model, which decomposes it into explicit essential needs features and implicit preference features. The explicit essential needs features define users' basic travel demands, while the implicit preference features uncover users' deeper travel preferences. Based on these explicit essential needs features and implicit preference features, a weighted feature fusion is performed to dynamically construct a three-dimensional user profile. This user profile supports real-time lightweight incremental updates. When new clicks, searches, or add-to-cart actions are collected, the weights of each feature in the profile are fine-tuned accordingly to keep the profile synchronized with the user's real-time travel decision-making needs.

[0031] Step S200: Obtain multidimensional tag data of travel products, and archive the multidimensional tag data by dimension layer to obtain a product resource pool; Specifically, multi-dimensional tag data of travel products listed on the entire platform is acquired. These travel products include travel routes, scenic spot packages, hotel accommodations, car rentals, and study tours / wellness products. This multi-dimensional tag data is divided into three dimensions: basic attribute tags, rigid audience adaptation tags, and dynamic real-time operational tags. The basic attribute tags are categorized and archived to generate a product basic attribute index. The rigid audience adaptation tags are automatically matched and labeled with five exclusive adaptation identifiers: family-friendly, age-friendly, romantic for couples, backpacking / outdoor, and business / efficient, generating a product audience adaptation index.

[0032] The dynamic real-time operation tags are annotated in real time with remaining inventory, clearance sales timelines, holiday pricing tiers, real-time sales volume, and supply chain turnover priority, generating a real-time product operation status index. The product basic attribute index, the product target audience index, and the product real-time operation status index are jointly encapsulated to construct a multi-dimensional hierarchical tag system, which is loaded into a high-speed memory cache, forming a product resource pool that supports millisecond-level fast retrieval. All tags are dynamically refreshed in real-time in conjunction with the supply chain backend, ensuring that the tags for each travel product in the product resource pool are authentic, valid, and fulfillable in real time.

[0033] Step S300: Input the user profile and the product resource pool into the dual-engine hybrid recommendation model for bidirectional matching and recall to generate a candidate product set; Specifically, the user profile and the product resource pool are input into a dual-engine hybrid recommendation model, which concurrently launches a collaborative filtering sub-engine and a deep learning representation sub-engine. The collaborative filtering sub-engine, based on travel preference features in the user profile, performs co-occurrence analysis with historical interaction data of other users in the product resource pool to identify a user group with similar travel preferences to the current user. It then filters travel products from the product resource pool whose historical interaction frequency reaches a preset threshold, recalling them in descending order of interaction frequency to generate a first candidate product subset. This first candidate product subset covers highly relevant and popular travel products.

[0034] The deep learning representation sub-engine inputs the user profile and the product resource pool into the latent feature mining layer. Through a deep neural network, it automatically mines the latent travel association features between each latent preference feature in the user profile and the multi-dimensional labels of each travel product in the product resource pool. These latent travel association features are then input into the vector representation construction layer for high-dimensional sparse vector representation, mapping them to user feature vectors and product feature vectors. In the long-tail product recall layer, the similarity between the user feature vector and each of the product feature vectors is calculated. Based on this similarity ranking, long-tail travel products in the product resource pool are recalled, generating a second candidate product subset. The first and second candidate product subsets are then weighted and fused to generate a candidate product set.

[0035] Step S400: Obtain dynamic context factors, perform weight correction and multi-objective balanced ranking on the candidate product set based on the factors, and output the final recommendation queue.

[0036] Specifically, dynamic context factors are obtained, including time-series scenario factors, meteorological environment factors, geographical location factors, and passenger flow and business type factors. Based on the time-series scenario factors, the initial weights of each travel product in the candidate product set generated in step S300 are first corrected. The time-series scenario factors are used to divide the travel rhythm weights according to weekdays, weekends, statutory holidays, winter and summer vacations, and spring and autumn off-peak seasons, and the corresponding time period weight coefficients are matched to each travel product in the candidate product set.

[0037] The candidate product set after the first correction is adjusted based on the meteorological environment factor and the passenger flow industry factor. The meteorological environment factor synchronizes the real-time weather warning information of the destination, and reduces the weight of outdoor travel products affected by severe weather and increases the priority of indoor cultural and tourism products. The passenger flow industry factor identifies the peak passenger flow status during holidays, and reduces the weight of travel products with crowded core scenic spots and increases the priority of travel products with off-peak and low-traffic areas.

[0038] Based on the aforementioned geographic location factor, a third correction is applied to the candidate product set after the second correction. The user's current location (after anonymization) is read, and long-distance intercity travel products are blocked, while nearby local cultural and tourism resources are prioritized. The candidate product set after these three corrections is then sorted according to multi-dimensional business indicators using a multi-objective collaborative equilibrium method, resulting in the final recommendation queue.

[0039] The intelligent recommendation method for travel products provided by this invention has the following beneficial effects: By collecting multi-dimensional travel data and integrating static attributes, temporal behaviors, and implicit preferences to build user profiles through feature modeling, the problem of insufficient shallow profile identification in existing solutions is solved, improving the granularity and matching accuracy of the profiles. By acquiring multi-dimensional tag data of travel products and establishing a product resource pool through dimensional hierarchical archiving, the shortcomings of existing solutions in terms of single product features and inability to perform refined indexing are overcome. A dual-engine hybrid recommendation model is used to perform bidirectional matching and recall between user profiles and the product resource pool, solving the problem of long-tail product recall failure in single collaborative filtering algorithms under sparse data, achieving coverage of both popular and long-tail products. By acquiring dynamic context factors and performing weight correction on the candidate product set, linking temporal scenarios, weather conditions, geographical location, and passenger flow information, the shortcomings of existing solutions in terms of lack of context awareness leading to a disconnect between recommendations and travel scenarios are overcome, improving the adaptability of recommended content to real travel scenarios. By performing multi-objective balanced ranking on the corrected candidate product set and coordinating optimization according to multi-dimensional business indicators, the problem of single-indicator ranking being unable to meet the needs of multiple parties is solved, achieving a comprehensive balance of user experience, transaction conversion, and inventory turnover.

[0040] In one embodiment, step S100: collecting multidimensional travel data of users, performing feature fusion modeling on the multidimensional travel data to obtain a user profile includes: Step S110: Collect the user's static attribute data and behavioral tracking data, and perform anomaly denoising and missing feature completion on the static attribute data and behavioral tracking data to obtain multi-dimensional travel data; Among them, static attribute data, officially known as basic compliant static attribute data, refers to the basic information that users fill in during registration and that does not change in real time. This includes age group segmentation, gender tags, cities of residence and travel, consumption level, and preferences for the number of people traveling with them. This data provides a basic framework for building user profiles and understanding travel needs. Behavioral tracking data, officially known as full-link time-series dynamic behavioral tracking data, refers to the user's operational trajectory information within the platform collected in real time through a unified tracking SDK across multiple terminals. This includes keywords searched in the travel channel, product click and browsing trajectory, duration of stay on single product pages, horizontal comparison behavior of multiple products, positive feedback such as adding to cart, favorites, and sharing, and negative feedback such as browsing without placing an order and short-term exit. This data reflects the user's true travel decision preferences and real-time changes in interest.

[0041] Specifically, a unified event tracking SDK deployed across multiple platforms—APP, mini-programs, and web pages—synchronously collects users' static attribute data and behavioral event tracking data. The collection process employs a compliant, anonymized channel, avoiding sensitive privacy information such as facial recognition, precise location data, and contact lists. The collected static attribute data undergoes field integrity verification, eliminating invalid records with all missing key fields and filling in missing fields with preset default values. The collected behavioral event tracking data undergoes abnormal traffic noise reduction, identifying and filtering invalid data generated by machine-generated traffic boosting and crawler requests—such as high-frequency repeated requests from a single IP address, non-human operation timing characteristics, and abnormal cross-domain redirects—while retaining genuine user operation records.

[0042] For the behavior tracking data after anomaly denoising, field missingness is detected simultaneously. For field gaps caused by network latency or terminal incompatibility, they are filled according to preset default rules to ensure data integrity. The static attribute data and behavior tracking data after anomaly denoising and missing feature completion are associated and merged according to the user's unique identifier to form single-user dimension data records. The merged data is encapsulated in JSON format and written as multi-dimensional travel data into a time-series feature database for subsequent steps. This time-series feature database is deployed in a cloud-based encrypted storage cluster to ensure data security and compliance. The output of this step provides standardized data input for the feature decomposition and profile construction in subsequent step S120.

[0043] Step S120: The multidimensional travel data is decomposed into explicit essential needs features and implicit preference features, and a user profile is constructed by weighted fusion based on the explicit essential needs features and the implicit preference features. Among them, explicit essential needs characteristics refer to the basic travel demands clearly expressed by users directly extracted from multi-dimensional travel data, including essential travel time, affordable budget limit, preferred destination type, fixed mode of transportation and attributes of companions. These characteristics are obtained by directly parsing and analyzing users' search keywords, filtering conditions and historical order information. Implicit preference features refer to the deep-seated travel preferences of users that are mined by analyzing the statistical patterns and correlations of user behavior data. These preferences include a preference for natural scenery or historical sites, a preference for niche or off-peak destinations or quiet travel during off-peak hours, an emphasis on the quality of hotel amenities, aversion to long-distance commutes, and aversion to crowded, high-density environments. These features are derived through correlation analysis of user dwell time, horizontal comparison behavior, and positive feedback behavior.

[0044] Specifically, a layered decomposition operation is performed on the multidimensional travel data: the user's search keywords, filtering conditions, and fields such as travel time, budget amount, destination type, mode of transportation, and number of people traveling with them in historical orders are analyzed and extracted into explicit essential feature vectors. Each feature dimension is encoded and stored in a standardized format.

[0045] We conduct time-series correlation analysis on page dwell time, multi-product horizontal comparison behavior sequences, and positive feedback records of adding to cart, favorites, and sharing in multi-dimensional travel data to uncover users' click preferences, dwell tendencies, and positive feedback patterns in different travel scenarios. The analysis results are mapped into implicit preference feature vectors. Explicit demand feature vectors and implicit preference feature vectors are input into a weighted fusion layer, and a vector weighted summation operation is performed according to preset weight coefficients to generate a comprehensive user profile vector.

[0046] A three-dimensional user profile is dynamically constructed based on the user profile vectors. The user profile supports real-time, lightweight incremental updates: when the front-end collects new user clicks, searches, or add-to-cart actions, the corresponding feature dimensions are locally and incrementally updated, adjusting only the weight values ​​of the affected feature dimensions, without requiring a full profile reconstruction, thus keeping the profile synchronized with the user's real-time travel decision-making needs. The completed user profile is written to the profile feature repository, serving as the matching input for the subsequent steps of the S300 collaborative filtering sub-engine and the deep learning representation sub-engine.

[0047] The user profile supports real-time lightweight incremental updates. When new clicks, searches, or add-to-cart actions are collected from users, the feature weights of the user profile are fine-tuned accordingly to keep the profile synchronized with the user's real-time travel decision-making needs.

[0048] Specifically, when the unified tracking SDK deployed on the front end collects new user clicks, searches, or add-to-cart behaviors, the behavioral data is transmitted in real time to the user profile update engine via a compliant and anonymized channel. The user profile update engine parses the received behavioral data, extracting the behavior type, operation object tag, and operation timestamp. Based on the behavior type and operation object tag, it locates the feature dimensions in the user profile affected by the behavior: search behavior is associated with the destination type and travel time dimension in the explicit essential needs feature; click browsing behavior is associated with the destination tendency and travel scenario preference dimension in the implicit preference feature; and add-to-cart behavior is associated with the product quality preference and consumption level dimension in the implicit preference feature.

[0049] Local incremental weight iterations are performed on the identified feature dimensions: positive behaviors increase the corresponding feature weight value by a preset incremental step size, and negative behaviors decrease the corresponding feature weight value by a preset decrementing step size. The incremental step size adaptively decays according to the frequency of the behavior, with the single adjustment step size decreasing for high-frequency similar behaviors to avoid weight overfitting. Weight fine-tuning only applies to the corresponding components in the user profile feature vector, while other feature dimensions of the profile remain unchanged. After fine-tuning, the user profile is synchronized to the profile feature repository in real time, overwriting the previous version of the profile. The profile feature repository provides the updated user profile to the collaborative filtering sub-engine and deep learning representation sub-engine of subsequent steps S300, ensuring that the recommendation matching input remains synchronized with the user's real-time travel decision requirements.

[0050] The method provided in this embodiment performs anomaly denoising and missing feature completion on user static attribute data and behavioral tracking data, and integrates explicit essential needs features and implicit preference features to construct a multi-dimensional user profile. This overcomes the shortcomings of existing solutions that rely solely on shallow static fields, resulting in insufficient profile identification, and improves the granularity and personalized representation capabilities of user profiles. By hierarchically decomposing and weightedly fusing multi-dimensional travel data, explicit travel demands and implicit behavioral preferences are collaboratively involved in profile modeling, avoiding matching inaccuracies caused by single-dimensional data bias, and providing more comprehensive user feature input for subsequent recommendations. The user profile supports real-time lightweight incremental updates, and when new clicks, searches, or add-to-cart behaviors are collected, the feature weights are locally fine-tuned to maintain dynamic synchronization between the profile and travel decision demands, solving the problem of traditional static profiles lagging behind and unable to reflect changes in user needs in real time.

[0051] like Figure 3 As shown, in one embodiment, step S200: obtaining multidimensional tag data of travel products, and performing dimensional hierarchical archiving of the multidimensional tag data to obtain a product resource pool, includes: Step S210: Obtain the multidimensional tag data of travel products listed on the entire platform; Among them, the travel products listed on the entire platform refer to all travel-related products currently available for sale on the online travel platform that can be searched and ordered by users, covering five major categories: travel routes, scenic spot packages, hotel accommodations, car rentals and self-driving tours, and study tours and health and wellness tours; multi-dimensional tag data refers to the structured attribute information of each travel product pre-labeled from multiple dimensions, including three types: basic attribute tags, rigid tags for audience adaptation, and dynamic real-time operation tags. This data provides a standardized description of product features for subsequent accurate matching.

[0052] Specifically, by calling the platform's standard data interface in the product center, the system batch-fetches raw product information for all travel products currently listed on the platform. The retrieval scope covers five categories: travel routes, scenic spot packages, hotel accommodations, car rentals, and study tours / health resorts. The retrieval requests use a paginated batch query method, with each page request carrying a category identifier, listing status filtering conditions, and a pagination cursor. The raw product information returned by the product center includes a unique product identifier, product name, category, basic description fields, and operational configuration parameters.

[0053] The returned raw product information is parsed, extracting the product's unique identifier as the primary key index value. Basic attributes such as price range, trip days, departure point, destination, and travel theme are extracted from the basic description field. Operational information such as audience suitability identifiers, inventory, delivery time, and price tiers are also extracted from the operational configuration parameters. The parsed and extracted information is then grouped by product unique identifier to generate multidimensional tag raw data records for each product, stored in a key-value pair structure. All multidimensional tag raw data records for all products are grouped and cached by category, serving as the input data source for the subsequent S220 tag dimension segmentation step.

[0054] Step S220: Divide the multidimensional tag data into basic attribute tags, rigid tags for audience adaptation, and dynamic real-time operation tags; Among them, the basic attribute tags refer to the tag dimensions that describe the inherent characteristics of travel products, covering five attributes: price gradient, travel duration, departure city circle, destination business type and travel theme style. These attribute tags are relatively stable and do not change frequently after the product is launched. The rigid tags for target audiences refer to exclusive matching labels that are automatically matched based on the characteristics of the target audience for travel products. These include five categories: family-friendly, senior-friendly, romantic for couples, backpacking and outdoor, and business-efficient. These labels are automatically generated based on the product's itinerary characteristics and service configuration. Dynamic real-time operation tags refer to the operational status tags that change in real time with the supply chain backend. These include five indicators: remaining inventory, near-expiry clearance time, holiday premium levels, real-time hot-selling status, and supply chain turnover priority. These tags need to be updated in real time with the supply chain backend.

[0055] Specifically, the tag fields describing the inherent characteristics of the product are categorized and extracted: the price range field is mapped to the price gradient in the basic attribute tags, the number of travel days is mapped to the travel duration, the departure location configuration field is mapped to the departure location circle, the destination city and scenic spot type field is mapped to the destination business type, and the itinerary theme label field is mapped to the travel theme style, and these are all grouped into the basic attribute tag group.

[0056] The tag fields describing the target audience of the product are categorized and extracted: child-friendly facilities and parent-child interactive projects in the product configuration are mapped to parent-child friendly tags; low-intensity itineraries and barrier-free facilities are mapped to elderly-friendly tags; double-person packages and romantic themes are mapped to romantic couple tags; outdoor sports intensity and hiking routes are mapped to backpacking outdoor tags; and meeting venues and quick itineraries are mapped to business efficiency tags. These are categorized into the rigid tag group for target audience suitability. The tag fields describing the real-time operational status of the product are categorized and extracted: the inventory quantity synchronized by the inventory management system is mapped to the remaining inventory balance; the number of days until the expiration date is mapped to the near-expiration clearance period; the price adjustment coefficient during holidays is mapped to the holiday premium level; the browsing and order statistics of the past seven days are mapped to real-time hot-selling status; and the allocation priority configuration of the supply chain system is mapped to the supply chain turnover priority. These are categorized into the dynamic real-time operation tag group. The results of the three tag group classifications are processed separately in subsequent steps S230 to S250.

[0057] Step S230: Classify and archive the basic attribute tags to generate a product basic attribute index; Specifically, the tag values ​​of each product are extracted from the basic attribute tag group output in step S220. A first-level inverted index is built according to the price gradient dimension, with the index key being the price gradient range code and the index value being the unique identifier set of the corresponding product. All products on the platform are bucketed and archived according to five gradients: low price, low-medium price, medium price, medium-high price, and high price.

[0058] A second-level inverted index is created based on the length of the trip. The index key is the code for the number of days of the trip, and the index value is the set of unique identifiers for the corresponding product. The products are then divided into bins and archived according to short-distance one-day trips, two- to three-day trips to surrounding areas, four- to seven-day trips to medium and long distances, and eight-day or longer trips to long distances.

[0059] A third-level inverted index is established based on the departure city layer, with the index key being the departure city code and the index value being the set of unique identifiers for the corresponding product.

[0060] A fourth-level inverted index is built based on destination type, with the index key being the destination type code, categorizing destinations into five types: natural scenery, cultural history, urban leisure, island resorts, and outdoor adventure. A fifth-level inverted index is built based on the theme / style of the trip, with the index key being the theme tag code and the index value being the set of unique identifiers for the corresponding product. Cross-index relationships are established between the five levels of inverted indexes through unique product identifiers, supporting multi-dimensional combined queries. The completed five-level index structure is loaded into a memory cache in the form of a hash map table, forming a basic product attribute index, with single-dimensional query response latency controlled within milliseconds.

[0061] Step S240: Automatically match and label the rigid tags for the target audience to generate a product target audience matching index; Specifically, the itinerary feature configuration parameters of each product are read from the rigid label group for audience adaptation output in step S220. The parent-child friendly adaptation judgment rule is activated to detect whether the product is equipped with child-friendly facility labels, whether it includes parent-child interactive itinerary items, and whether it supports child prices and family packages. Products that meet the judgment conditions are automatically labeled with parent-child friendly labels.

[0062] The system activates age-friendly adaptation rules, checking if the product's average daily walking distance is below a preset threshold, whether it has accessible pathways and rest areas, and whether the itinerary pace is rated as relaxed. Products meeting these criteria are automatically labeled as age-friendly. The system also activates couple-friendly adaptation rules, checking if the product includes a double-person package, romantic themed accommodations or dining, and whether the itinerary includes private attractions suitable for couples. Products meeting these criteria are automatically labeled as couple-friendly.

[0063] The system activates backpack / outdoor compatibility rules to determine whether a product includes outdoor hiking or mountaineering trips, offers outdoor equipment rental services, and has a moderate or higher activity level rating. Products meeting these criteria are automatically labeled with a backpack / outdoor tag. Similarly, the system activates business / efficiency compatibility rules to determine whether a product includes meeting venues or business transportation services, whether the trip duration is within a preset efficient travel time range, and whether it supports flexible cancellations and changes. Products meeting these criteria are automatically labeled with a business / efficiency tag.

[0064] After the products have been labeled, an inverted index is created for each audience type. Each audience type is used as an independent index key. The index value stores a set of unique identifiers for all products labeled with that type and loads them into the memory cache to form the product audience matching index.

[0065] Step S250: Synchronize and label the dynamic real-time operation tags in real time to generate a product real-time operation status index; Among them, dynamic real-time operation tags refer to operation-related tag categories that change in real time with product inventory status, sales popularity and promotion cycle, including remaining inventory, near-expiry clearance time, holiday premium levels, real-time hot sales and supply chain turnover priority. This tag data changes frequently and needs to be updated in real time or near real time.

[0066] Specifically, by subscribing to message queue topics in the supply chain backend, a real-time data synchronization link is established with the inventory management system, price management system, and timeliness management system to monitor three types of messages: inventory changes, price adjustments, and timeliness updates. When a message about a change in remaining inventory is received, the unique product identifier and the latest inventory value in the message body are parsed, the remaining inventory tag value for that product in memory is updated, and the product sorting position in the inventory turnover priority tag is adjusted synchronously.

[0067] Upon receiving a message about a change in the expiration date for a clearance sale, the system parses the product's unique identifier and the number of days remaining until the expiration date, updates the clearance sale tag value for that product, and automatically marks expired products as unavailable and removes them from the index. Upon receiving a message about a change in holiday premium tiers, the system parses the product's unique identifier and the adjusted price coefficient, and updates the holiday premium tier tag value for that product.

[0068] Real-time best-selling heat tags are calculated based on a sliding window of browsing and order statistics over the past seven days. The latest window statistics are retrieved from the data warehouse at preset time intervals to calculate the browsing-to-order conversion rate for each product and update the heat value. Supply chain turnover priority tags are synchronously refreshed according to the allocation priority configuration of the supply chain management system. Using the product's unique identifier as the primary key, the updated values ​​of the five types of operational tags are written into the product's real-time operational status index. The index is stored using an ordered set data structure, with each operational indicator serving as an independent sorting dimension, supporting fast retrieval by indicator value in ascending or descending order.

[0069] Step S260: Combine and encapsulate the product basic attribute index, the product audience adaptation index, and the product real-time operation status index to construct a multi-dimensional hierarchical tag system and form the product resource pool.

[0070] Specifically, the product basic attribute index, product audience adaptation index, and product real-time operational status index are jointly encapsulated. A cross-index mapping table is constructed with the product's unique identifier as the primary association key. Each record in the mapping table contains the product's unique identifier and the product's bucketing information and sorting position reference in the three indexes, enabling the retrieval of all tag dimension data for a product at once using its unique identifier.

[0071] The three-layer progressive architecture of the multi-dimensional hierarchical tagging system is as follows: The bottom layer is the basic attribute layer, which connects to the product's basic attribute index and carries the product's inherent characteristic information; the middle layer is the audience adaptation layer, which connects to the product's audience adaptation index and carries the adaptation relationship information between the product and the travel audience; the top layer is the dynamic operation layer, which connects to the product's real-time operation status index and carries the product's real-time fulfillment status information.

[0072] Data interaction between the three layers is achieved through a cross-index mapping table. When inventory or timeliness information changes in the dynamic operations layer, the corresponding product's recommendability status in the intermediate layer's adaptation relationship is synchronously updated via the mapping table. The encapsulated two-dimensional hierarchical tag system is serialized into a continuous in-memory data structure and loaded into a high-speed memory cache to form a product resource pool. The product resource pool provides a unified retrieval interface, receiving tag dimension and tag value parameters from the search conditions. It routes the query to the corresponding dimension's index through the cross-index mapping table, and returns the query results after deduplication and merging using the product's unique identifier. The product resource pool and user profiles together serve as the input basis for the dual-engine hybrid recommendation model's matching and recall, supporting subsequent collaborative filtering sub-engines to retrieve similar user-preferred products based on the audience adaptation dimension, and deep learning representation sub-engines to mine implicit association features of long-tail products by combining basic attributes and dynamic operations dimensions.

[0073] The method provided in this embodiment establishes a multi-dimensional hierarchical tagging system by dividing and archiving the multi-dimensional tag data of travel products across the entire platform into three dimensions: basic attributes, audience adaptation, and dynamic real-time operation. This overcomes the shortcomings of existing solutions, such as single product features and the inability to perform refined indexing, providing a structured multi-dimensional retrieval basis for subsequent recommendations. Dynamic real-time operation tags are used to synchronously label remaining inventory, near-expiry clearance sales, and holiday premium levels, enabling the product resource pool to reflect the immediate fulfillment status of products and avoiding recommendation inaccuracies caused by static tags failing to perceive changes in inventory availability. By jointly encapsulating the basic attribute index, audience adaptation index, and real-time operation status index into a unified product resource pool, it supports combined queries and rapid retrieval of different dimension tags, allowing the popular product matching and long-tail product mining of the hybrid recommendation model to share the same standardized product feature input.

[0074] like Figure 4 As shown, in one embodiment, step S300: inputting the user profile and the product resource pool into a dual-engine hybrid recommendation model for bidirectional matching and recall to generate a candidate product set includes: Step S310: Input the user profile and the product resource pool into the dual-engine hybrid recommendation model, which includes a collaborative filtering sub-engine and a deep learning representation sub-engine that are started in parallel. Among them, the dual-engine hybrid recommendation model refers to a hybrid recommendation computing architecture that consists of a collaborative filtering sub-engine and a deep learning representation sub-engine in parallel, and performs weighted fusion processing on the two recall results. The dual-engine hybrid recommendation model is divided into a four-layer closed-loop rigid structure with unidirectional data flow from top to bottom, ensuring no cyclic interference and no feature leakage: The first layer is a unified sparse behavioral data input layer, serving as a shared underlying data source for both engines. It synchronously feeds the two engines with the same user behavior matrix, multi-dimensional product label features, and positive and negative feedback vectors, ensuring that the baseline data of the two recommendation computing power is completely consistent. The second layer is a differentiated feature representation layer, where the collaborative filtering sub-engine enters the group co-occurrence feature statistics branch, and the deep learning representation sub-engine enters the high-dimensional dense vector semantic representation branch, each performing its own function without interfering with each other. The third layer is a dual-engine independent candidate recall computing power inference layer, where the collaborative filtering sub-engine completes global popular association recall, and the deep learning representation sub-engine completes long-tail implicit preference accurate recall, each outputting an independent subset of candidate products. The fourth layer is a dynamic weighted fusion decision output layer, configured with an adaptive fusion weight gating unit, which performs cross-duplicate removal, interest weighting, and priority rearrangement on the two candidate sets, merging them to generate an initial candidate product set.

[0075] Specifically, the system reads the user's 3D profile and the product resource pool. The feature vectors of the user profile are serialized and encoded into a dual-engine common input format, with the encoding containing standardized vector representations of explicit essential feature components and implicit preference feature components. The unique identifiers of each product in the product resource pool and their corresponding multi-dimensional label data are encapsulated item by item into a feature input sequence, categorized by product dimension. This encapsulation includes a structured key-value pair set containing basic attribute label values, user group suitability identifiers, and real-time operational status label values. The dual-engine hybrid recommendation model's runtime environment is initialized, loading the interaction co-occurrence calculation layer and product recall layer components of the collaborative filtering sub-engine, and loading the implicit feature mining layer, vector representation construction layer, and long-tail product recall layer components of the deep learning representation sub-engine.

[0076] The serialized user profile feature vector and the product resource pool feature input sequence are simultaneously sent to the input buffers of the collaborative filtering sub-engine and the deep learning representation sub-engine. Upon receiving the input data, both sub-engines initiate their respective matching and recall processes in parallel: the collaborative filtering sub-engine initiates interaction co-occurrence analysis, and the deep learning representation sub-engine initiates deep neural network forward inference. The two processes execute in parallel without interference. After parallel startup, the dual-engine hybrid recommendation model enters a state where it waits for the recall results from both paths to be returned. Subsequently, steps S320 and S330 respectively complete the recall calculations for each sub-engine.

[0077] Step S320: The collaborative filtering sub-engine matches the user profile with popular products in the product resource pool to generate a first subset of candidate products; Step S330: Using the deep learning representation sub-engine, perform latent feature mining and vector representation construction on the user profile and the product resource pool to generate a second candidate product subset; Step S340: The first candidate product subset and the second candidate product subset are weighted and merged to generate a candidate product set.

[0078] Specifically, the unique identifiers of the product records in the first and second candidate product subsets are compared. Products that appear in both subsets are marked as duplicates. For duplicates, the higher score from the two recall scores is retained and merged into a single record.

[0079] Before performing weighted fusion, the current computing power status and user interaction density are read, and the two recall weights are dynamically and adaptively adjusted: when the platform is in a high-concurrency period during the peak tourist season, the weight of the collaborative filtering sub-engine is automatically increased to ensure the stable operation of the system with low latency; when users are in a low-interaction and sparse scene, the weight of the deep learning representation sub-engine is automatically increased to enhance personalized and accurate matching.

[0080] The product recall score in the first candidate product subset is multiplied by the recall confidence weight coefficient preset by the collaborative filtering engine, and the product recall score in the second candidate product subset is multiplied by the recall confidence weight coefficient preset by the deep learning representation engine. The weight coefficient is set according to the independent accuracy of each engine in historical evaluations.

[0081] The weighted products from the two streams are merged into a single candidate list, and then sorted in descending order according to their weighted composite scores. The composite score reflects the overall matching strength of the product from the perspectives of the two engines. After sorting, a predetermined number of products at the top of the ranking are selected to generate a candidate product set. Each record in the candidate product set contains a unique product identifier, a weighted composite score, and a tag indicating that the product originated from an engine. The candidate product set is written to the candidate set cache as input data for step S400, which involves dynamic context factor weight correction and multi-objective balanced sorting.

[0082] The method provided in this embodiment employs a dual-engine hybrid architecture, with a collaborative filtering sub-engine and a deep learning representation sub-engine running in parallel. This architecture performs popular product matching and long-tail product latent feature mining on user profiles and product resource pools, respectively. This overcomes the limitation of single collaborative filtering algorithms failing to recall long-tail products under sparse interaction data, achieving comprehensive coverage of both popular essential products and long-tail products. The collaborative filtering sub-engine's interaction co-occurrence calculation layer performs similarity calculations for travel preference features within the same user group and filters frequently interacting products, improving the accuracy and user coverage of popular product matching. This avoids the one-way push bias of traditional popular product ranking, which relies solely on sales volume and ignores differentiated user needs. The deep learning representation sub-engine uses high-dimensional sparse vector representation modeling and cosine similarity recall to cross-domain associate user latent preference features with multi-dimensional product tags in a high-dimensional space. This allows niche, regionally specific, near-expiry, high-value, and off-peak long-tail products to emerge from sparse interaction data and be effectively recalled, revitalizing long-term unsold inventory resources in traditional recommendation architectures.

[0083] In one embodiment, step S320: matching the user profile with the product resource pool using the collaborative filtering sub-engine to generate a first candidate product subset includes: Step S321: The interaction co-occurrence calculation layer of the collaborative filtering sub-engine performs co-occurrence analysis on the travel preference features in the user profile and the historical interaction data of other users in the product resource pool to determine the similar user group in the same circle of the user image; The interaction co-occurrence calculation layer refers to the computational component in the collaborative filtering sub-engine responsible for calculating the similarity of travel preferences based on the co-occurrence relationship of user behaviors. This layer receives user profile feature vectors and historical user behavior data, and outputs a set of identifiers of similar user groups that are highly similar to the current user's travel preferences. Travel preference features refer to explicit, essential feature components directly related to travel decisions extracted from the user profile feature vectors, including feature values ​​in five dimensions: travel time preference, affordable budget limit, preferred destination type, fixed mode of transportation, and companion group attributes. The collaborative filtering sub-engine adopts a product-based industrial-grade lightweight collaborative filtering architecture, which contains three rigid serial processing logics: the first step is to construct the interaction sparse co-occurrence matrix between users and travel products in real time; the second step is to batch calculate the similarity association matrix between travel products and generate a product similarity index table; and the third step is to retrieve highly similar popular travel products based on the current user's positive behavior product set.

[0084] Specifically, the interaction co-occurrence computing layer extracts travel preference feature components from the user profile feature vector, combining the feature values ​​of five dimensions—travel time preference, affordable budget limit, preferred destination type, fixed mode of transportation, and companion group attributes—into a travel preference feature vector. The interaction co-occurrence computing layer constructs a sparse co-occurrence matrix of interactions between all users and travel products in real time. The matrix row index is the anonymized user identifier registered across the entire platform, and the matrix column index is the unique resource code for all categories of travel products. Matrix cells are filled with a comprehensive score based on user behavior type: 1 point for a single valid click, 2 points for a valid stay exceeding 15 seconds, 4 points for successfully adding to cart, and 5 points for a complete order. Positions without interaction behavior are left empty with a sparse mask by default, without consuming computing power or storage resources. The layer loads a dataset of interaction behaviors from all historical users belonging to the same departure point stratum as the current user from the historical behavior data warehouse. Each record in the historical interaction behavior dataset contains a unique user identifier, a unique identifier for the corresponding travel product, the type of interaction behavior, and the frequency of interaction.

[0085] The interaction co-occurrence calculation layer uses the unique product identifier as the association key. It performs an intersection operation on the current user's historical interaction product set and the historical interaction product sets of each historical user, and counts the number of products that the two users interacted with together and the interaction frequency weighted value of the products. The interaction frequency weighted value is calculated based on the comprehensive score in the interaction sparse co-occurrence matrix. Products belonging to the same destination business type and play theme style in the co-interaction product set are clustered to extract the co-occurrence patterns of the two users in terms of travel destination and play preferences.

[0086] The cosine similarity of the travel preference feature vectors and the weighted value of interaction co-occurrence are fused according to a preset weight ratio to obtain the comprehensive travel preference similarity score between the current user and each historical user. The comprehensive travel preference similarity scores are sorted in descending order, and historical users whose similarity scores reach a preset similarity threshold are selected to form a similar user group within the same circle. The similar user group within the same circle is output to the product recall layer of the collaborative filtering sub-engine in the form of a list of unique user identifiers, as the user scope input for product screening and recall in step S322.

[0087] Step S322: The product recall layer of the collaborative filtering sub-engine filters interactive travel products of similar user groups in the same circle from the product resource pool, and recalls them in order of interaction frequency to generate the first candidate product subset.

[0088] The product recall layer refers to the computational component in the collaborative filtering sub-engine that is responsible for filtering and sorting candidate travel products from the product resource pool based on the historical interaction behavior of similar user groups. This layer receives a list of user identifiers from similar user groups in the same circle and the product resource pool index, and outputs a collection of popular travel products recalled in order of interaction frequency.

[0089] Specifically, the product recall layer receives the list of unique identifiers of users in the same circle and similar user groups output in step S321. Using each user's unique identifier as the query condition, it batch-reads the historical positive interaction records of each user in the same circle and similar user groups from the historical behavior data warehouse. The positive interaction records only retain the behavior data of the interaction type of click, favorite, and order, and remove negative behaviors such as browsing without ordering and short-term exit, as well as abnormal daily volume boosting noise behavior and invalid samples of one-time erroneous clicks.

[0090] Based on the sparse co-occurrence matrix of interactions constructed in step S321, the product recall layer uses the set of commonly interacting users as the similarity calculation benchmark to batch calculate the cosine similarity association matrix between all travel products, generating a real-time similarity index table between products to characterize the strong association travel pairing relationships between products. Using the product's unique identifier as the aggregation key, cross-user aggregation statistics are performed on the positive interaction records of all users in similar user groups within the same circle, calculating the cumulative number of interacting users, the cumulative total interaction frequency, and the most recent interaction timestamp for each interacted travel product. Based on the time interval between the most recent interaction timestamp and the current time, the cumulative total interaction frequency of each product is weighted by time-sensitivity decay, with the decay coefficient decreasing as the interaction time interval increases. Recent interaction frequencies retain higher weights, while older interaction frequencies have their weights decayed. The interaction frequency weighted by time-sensitivity decay is used as the product's recall ranking score.

[0091] Based on the set of positive behavior products that the current user has viewed, added to cart, and favorited, highly similar popular travel products are retrieved from the product similarity index table. The availability of each product in the product resource pool is verified, and products that are no longer available or have zero inventory are removed from the recall candidates. The remaining products are sorted in descending order of recall ranking score, and a predetermined number of travel products at the top of the ranking are selected to generate a popular fallback candidate subset. For each product, its unique identifier, recall ranking score, and recall source engine identifier are extracted to form the first candidate product subset. The first candidate product subset is written into the fusion buffer of the dual-engine hybrid recommendation model and, together with the second candidate product subset generated in step S330, undergoes weighted fusion processing in step S340.

[0092] The method provided in this embodiment analyzes the co-occurrence of travel preference features in user profiles with historical user interaction data through the interaction co-occurrence calculation layer of the collaborative filtering sub-engine. It integrates the multi-dimensional preference similarity of users with the weighted value of interaction co-occurrence to determine similar user groups within the same circle, overcoming the shortcomings of traditional collaborative filtering that relies solely on the co-occurrence of single product clicks, which leads to biased identification of similar users and improving the accuracy of similar user group positioning. The product recall layer performs positive interaction filtering and time-sensitivity decay weighted sorting on interactive travel products for similar user groups within the same circle, prioritizing the recall of recently high-frequency positive interaction products. This avoids the problem of recommendation timeliness decay caused by outdated historical interaction data, improving the timeliness and relevance of popular product recalls. By verifying the availability of products from the product resource pool and removing products that are no longer available or have zero inventory, the recall results are consistent with the real-time fulfillment status of products, preventing invalid recommendations and traffic waste caused by unsellable products entering the candidate set.

[0093] like Figure 5 As shown, in one embodiment, step S330: the step of performing latent feature mining and vector representation construction on the user profile and the product resource pool through the deep learning representation sub-engine to generate a second candidate product subset includes: Step S331: Input the user profile and the product resource pool into the latent feature mining layer of the deep learning representation sub-engine, and mine the latent travel association features between the user profile and the product resource pool through the latent feature mining layer; The implicit feature mining layer refers to the computational component in the deep learning representation sub-engine responsible for automatically extracting deep association patterns between users and products that cannot be directly captured by shallow data through multi-layer nonlinear transformations of deep neural networks. This layer takes user profile feature vectors and multi-dimensional label data from the product resource pool as input and outputs implicit travel association feature vectors that represent the matching strength between users and various products in the dimension of potential travel preferences. Implicit travel association features refer to the deep association between travel tendencies and product features that users do not explicitly express through search keywords or filtering conditions, but that are reflected by implicit signals such as dwell time, horizontal comparison patterns, and positive feedback distribution in behavioral data. These include users' implicit preferences for specific destination styles, their potential sensitivity to travel rhythm and crowd density, and their implicit requirements for product quality and service configuration. The deep learning representation sub-engine adopts a lightweight deep neural network embedded attention representation architecture, which contains a four-layer progressive operation structure. The first layer is a multimodal feature unified embedding preprocessing layer, which maps all user discrete attribute data and product tag data into low-dimensional dense floating-point vectors to eliminate sparse data blank points. The second layer is a multi-layer fully connected hidden layer automatic cross-high-order feature learning layer. The network is set with three hidden layers to automatically learn the deep implicit travel patterns of users layer by layer, and complete the secondary purification of personalized deep profile.

[0094] Specifically, the latent feature mining layer extracts latent preference feature components from the user profile feature vector. These components include standardized values ​​across four dimensions: destination preference, travel scenario preference, product quality preference, and consumption level preference. The latent feature mining layer first initiates a unified multimodal feature embedding preprocessing step, mapping user discrete age groups, consumption levels, travel time-series behavior, product price tags, destination theme tags, and real-time inventory tags into low-dimensional dense floating-point vectors. This uniformly eliminates sparse data gaps and fills the gaps in low-interaction user features. Multidimensional tag data is read from the product resource pool product by product. For each product, the layer extracts destination business type and travel theme style codes from the basic attribute tags, the audience type adaptation identifier sequence from the audience adaptation tags, and the near-expiration clearance period and real-time hot-selling value from the dynamic real-time operation tags.

[0095] The four-dimensional feature values ​​of the user's implicit preference features are concatenated with the multi-dimensional label encoding values ​​of the single product to form a fixed-length input feature vector. This input feature vector is then fed into the input layer of a pre-trained deep neural network, initiating automatic cross-feature learning across multiple fully connected hidden layers. The deep neural network contains three fully connected hidden layers, each employing a non-linear activation function for layer-by-layer feature transformation. Adjacent hidden layers are fully connected via weight matrices. The input feature vector is transformed by the first hidden layer to obtain a low-order feature interaction representation, by the second hidden layer to obtain a mid-order feature cross-representation, and by the third hidden layer to obtain a high-order semantic feature representation. This process automatically learns the user's deep implicit travel patterns layer by layer, automatically identifying fine-grained preference features that cannot be captured by manual rules, such as "preferring secluded locations with low foot traffic," "implicitly valuing accommodation quality," and "disliking long-distance travel," thus completing a secondary refinement of the personalized deep profile. The output dimensions of each hidden layer are adjusted layer by layer according to preset network structure parameters.

[0096] The mining results are obtained from the output layer of the deep neural network, and the output is a latent travel association feature vector between the user and the product. Each dimension of the latent travel association feature vector corresponds to the matching strength value of different latent preference dimensions. The above mining process is performed on all products in the product resource pool one by one to generate a set of latent travel association feature vectors that correspond one-to-one between the user and each product. The latent travel association feature vector sets are sorted by the unique identifier of the product and output as a whole to the vector representation construction layer as the input of step S332.

[0097] Step S332: Input the implicit travel association features into the vector representation construction layer of the deep learning representation sub-engine to construct vectors and represent them as user feature vectors and product feature vectors; The vector representation construction layer refers to the computational component in the deep learning representation sub-engine responsible for transforming implicit travel-related features into fixed-length feature vectors in a unified space through high-dimensional sparse embedding mapping. This layer completes the conversion from implicit related features to vector representations that can be used for similarity calculation. The user feature vector refers to a fixed-dimensional vector obtained by mapping the implicit preference features of the user side through high-dimensional sparse embedding mapping. The vector is located in a high-dimensional sparse space, and the non-zero components of the dimension correspond to the key feature signals of user preferences. The product feature vector refers to a fixed-dimensional vector obtained by mapping the multi-dimensional label features and implicit related features of the product side through high-dimensional sparse embedding mapping. Its vector space is located in the same high-dimensional sparse coordinate system as the user feature vector, and the vector dimension is completely aligned with the user feature vector. The third layer in the four-layer progressive computation structure is a lightweight attention mechanism dynamic weight allocation layer, which assigns high attention weight to the user's recent travel behavior and reduces the weight to the user's long-term historical behavior, thus aligning with the user's current real-time travel decision-making rhythm.

[0098] Specifically, the vector representation construction layer receives the implicit travel association feature vector group of users and each product output in step S331. For each single-product implicit travel association feature vector in the implicit travel association feature vector group, user-side feature components and product-side feature components are separated. The user-side feature components correspond to the four-dimensional original feature values ​​of user implicit preference features and their higher-order semantic feature values ​​after transformation by three hidden layers of a deep neural network. The product-side feature components correspond to the encoded values ​​of destination business type, travel theme style, audience adaptation identifier, and dynamic operation features in the product's multi-dimensional tags and their higher-order semantic representation after multi-layer transformation by a deep neural network.

[0099] The vector representation construction layer initiates a lightweight attention mechanism with dynamic weight allocation. Higher attention weights are assigned to high-order semantic features corresponding to the user's recent travel behavior within the last 7 days, while lower attention weights are assigned to features corresponding to older historical behaviors with time intervals exceeding six months. This avoids misleading recommendations based on outdated preferences and ensures the recommendation results align with the user's current, real-time travel decision-making rhythm. The attention-weighted user-side feature components are then multiplied with a pre-trained high-dimensional sparse embedding mapping matrix. The number of rows in the embedding mapping matrix is ​​the preset vector dimension, the number of columns is the input feature dimension, and the matrix elements are the embedding weight parameters after training convergence. This maps the user features in the original feature dimension space to a high-dimensional sparse space, generating user feature vectors. The user feature vectors have a higher dimension than the original input feature dimensions, with most dimensional components being zero values ​​and only a few dimensions having non-zero activation values, forming a sparse representation.

[0100] Simultaneously, product-side feature components are multiplied with the same embedding mapping matrix to map to a high-dimensional sparse space identical to the user feature vector, generating product feature vectors. The dimension of the product feature vectors is completely consistent with that of the user feature vectors, ensuring that the two vectors can be directly used to calculate the spatial distance metric. After constructing the product feature vectors for all products, they are cataloged and stored according to their unique product identifiers. The cataloged sets of user feature vectors and product feature vectors for each product are then output to the long-tail product recall layer as input for the similarity calculation in step S333.

[0101] Step S333: Calculate the similarity between the user feature vector and each product feature vector through the long-tail product recall layer, and recall the target travel products in the product resource pool based on the similarity ranking to generate the second candidate product subset.

[0102] The long-tail product recall layer refers to the computational component in the deep learning representation sub-engine responsible for retrieving long-tail travel products from the product resource pool that best match the user's feature vectors through vector space similarity measurement. Target travel products refer to long-tail travel products that deeply match the user's implicit travel preferences, selected through vector similarity calculation. These include niche local specialty products, high-value products nearing their expiration date, and off-peak products for less popular periods. These products typically rank in the middle to lower positions in the interaction frequency ranking of the collaborative filtering engine but have a high degree of implicit preference matching. The second candidate product subset refers to the set of candidate products generated by the deep learning representation sub-engine based on vector representation and similarity retrieval. This subset uses vector space similarity as the primary ranking criterion, focusing on covering high-quality long-tail products that are difficult to effectively recall using traditional collaborative filtering algorithms. The fourth layer in the four-layer progressive computation structure is the global vector space nearest neighbor similarity retrieval layer, which calculates the spatial distance between the user's deep representation vector and the product vectors across the entire platform to accurately recall long-tail travel products with highly matched vectors.

[0103] Specifically, the long-tail product recall layer receives the user feature vector output from step S332 and the catalog set of product feature vectors for each product. It initiates a global vector space nearest neighbor similarity search, using the user feature vector as the query vector, and iterates through each product feature vector in the catalog set, calculating a cosine similarity score for each product. The cosine similarity calculation formula is the inner product of the user feature vector and the product feature vector divided by the product of the magnitudes of the two vectors; the calculated similarity score ranges from negative one to positive one. After initial sorting of the similarity scores for all products, an initial ranking list is formed.

[0104] The long-tail product recall layer reads the list of unique product identifiers of the first candidate product subset from the fusion buffer, marks popular products in the initial ranking result list that have the same unique product identifiers as the first candidate product subset as duplicate items, and determines whether to retain the duplicate items according to the preset deduplication judgment rules: if the similarity score of the product in the initial ranking result list exceeds the preset retention threshold, the product is retained and marked as jointly recommended by the dual engines; if the score is lower than the retention threshold, it is removed from the initial ranking result list.

[0105] The initial ranking list after deduplication is reordered in descending order based on similarity score. A predetermined number of products are extracted from the top-ranked list to precisely recall niche and unique routes with highly matched vectors, high-value packages nearing their expiration dates, and newly launched local cultural and tourism long-tail products. These products constitute the long-tail product recall results.

[0106] Each record in the long-tail product recall results is labeled with the recall source engine identifier. The product unique identifier, cosine similarity score, and recall source engine identifier are extracted for each product to generate a second candidate product subset. The second candidate product subset is written into the fusion buffer of the dual-engine hybrid recommendation model and, together with the first candidate product subset generated in step S322, serves as the input for the weighted fusion processing in step S340.

[0107] The method provided in this embodiment performs multi-layer nonlinear transformations on user implicit preference features and product multi-dimensional labels through the implicit feature mining layer of the deep learning representation sub-engine. This automatically extracts implicit travel-related features that cannot be directly captured by shallow data, overcoming the deficiency of traditional recommendations that rely solely on explicit interaction data, resulting in the ineffective association between long-tail products and user latent preferences. Through a vector representation construction layer, the implicit association features of the user side and the product side are mapped to the same vector space via high-dimensional sparse embedding. This transforms the matching relationship between user preferences and product features from discrete label collisions to a distance metric in a continuous vector space, improving the semantic accuracy and generalization ability of long-tail product matching. The long-tail product recall layer calculates the cosine similarity between user feature vectors and product feature vectors. After filtering popular products already covered by the collaborative filtering engine from the recall results, long-tail products are sorted by similarity and extracted. This allows niche, localized, high-value-for-money products nearing their expiration date, as well as off-peak products during off-peak periods, to effectively emerge from sparse interaction data, revitalizing inventory resources that have long been unsold due to insufficient interaction samples in traditional recommendation architectures.

[0108] In one embodiment, step S400: obtaining dynamic context factors, performing weight correction and multi-objective balanced ranking on the candidate product set based on the factors, and outputting the final recommendation queue includes: Step S410: Obtain dynamic context factors, which include time-series scene factors, meteorological environment factors, geographical location factors, and passenger flow and business type factors; Among them, the time-series scenario factor refers to the travel rhythm weight parameter divided according to the time period type of the current date. The time period types include five categories: weekdays, weekends, statutory long holidays, winter and summer vacations, and spring and autumn off-peak seasons; the meteorological environment factor refers to the real-time meteorological warning information synchronized from the destination meteorological service interface, including four types: sunny and rainy status, high temperature warning, strong wind warning, and cold wave warning; the geographical location factor refers to the city-level geocode obtained after parsing the de-identified current location authorized by the user terminal, which is used to determine whether the user is in a travel status in another location; the passenger flow industry factor refers to the real-time passenger flow density status information obtained from the passenger flow heat map data interface of the scenic spot or the local area, which is used to identify whether the core scenic spot is in the peak passenger flow state during holidays.

[0109] Specifically, by calling the platform's unified time and date service interface, the timestamp and date type identifier of the current date are obtained. The date type identifier is automatically determined and returned by the time and date service according to the preset holiday calendar and solar term division rules, including five enumerated values: weekday identifier, weekend identifier, statutory long holiday identifier, winter and summer vacation identifier, and spring and autumn off-peak season identifier.

[0110] The acquired date type identifier and its corresponding preset travel rhythm weight coefficient are encapsulated into time-series scenario factors. By calling the destination weather service interface and passing in the destination city codes of each product in the candidate product set, weather warning information for the current and future preset time periods of the destination is obtained in batches. The weather warning information includes the warning type and warning level, and the weather warning type and level of each destination city are encapsulated into meteorological environmental factors.

[0111] By calling the location service authorization interface of the user's terminal device, the city-level geocode of the user's current anonymized location is obtained. The anonymization process retains only the city-level location information without obtaining the precise latitude and longitude, and the city geocode is encapsulated as a geographic location factor. By calling the scenic area visitor flow data service interface, real-time visitor flow density data of the core scenic area and the judgment indicator of whether the preset peak threshold has been reached are obtained, and the visitor flow peak judgment indicator is encapsulated as a visitor flow business type factor.

[0112] The above four factors are merged into a dynamic context factor data group and written into the dynamic context cache, so that steps S420 to S440 can read and call them in the corrected order.

[0113] Step S420: Adjust the initial weights of each travel product in the candidate product set based on the time-series scenario factors to obtain the first corrected candidate product set; Specifically, the date type identifier and corresponding travel rhythm weight coefficient are read from the time-series scenario factor in the dynamic context cache. The candidate product set generated in step S340 is read from the candidate set cache. Each record in the candidate product set contains a unique product identifier, a weighted comprehensive score, and travel duration and travel theme style information from the product's basic attribute tags.

[0114] The travel rhythm characteristics of the current time period are determined based on the date type identifier of the time sequence scenario factor: weekdays correspond to short-distance efficient travel rhythm, weekends correspond to local leisure travel rhythm, statutory long holidays correspond to long-distance inter-provincial travel rhythm, winter and summer vacations correspond to parent-child study tour travel rhythm, and spring and autumn off-peak seasons correspond to off-peak long-distance travel rhythm.

[0115] The matching degree of each product's travel duration tag in the candidate product set is calculated with the current travel rhythm: a high matching degree is set when the product's travel duration matches the typical duration range of the current rhythm, a medium matching degree is set when there is a partial match, and a low matching degree is set when there is no match. The matching degree value is multiplied by the travel rhythm weight coefficient corresponding to the date type to obtain the time series adjustment coefficient.

[0116] The weighted composite score of each product is multiplied by a time-series adjustment factor to obtain the time-series corrected product weight score. The candidate product set is then reordered according to the corrected product weight scores to generate the first corrected candidate product set. The first corrected candidate product set is written into the correction intermediate buffer as input for step S430, environmental weight adjustment.

[0117] Step S430: Adjust the environmental weights of the first correction candidate product set based on the meteorological environmental factors and the passenger flow and business type factors respectively to obtain the second correction candidate product set; Specifically, the meteorological warning type and warning level of each destination city are read from the meteorological environmental factors in the dynamic context cache, and the peak passenger flow determination indicators of each core scenic spot are read from the passenger flow business type factors. The first correction candidate product set is read from the correction intermediate cache.

[0118] Meteorological matching is performed on each product in the first set of candidate products for correction: the site type information in the business type tag of the product destination is extracted to identify whether the site attribute of the product is outdoor or indoor. When the meteorological warning type of the product destination city is a rainstorm, strong wind, high temperature or cold wave warning and the warning level reaches the preset severe weather threshold, it is determined that the destination is currently in a severe weather state.

[0119] For products whose destination is in severe weather and whose venue is outdoor, their weighted score is reduced by multiplying it by a preset outdoor severe weather attenuation coefficient. For products whose destination is in severe weather and whose venue is indoor, their weighted score is increased by multiplying it by a preset indoor severe weather gain coefficient. For products with no weather warning or whose warning level is below the severe weather threshold, their weighted score remains unchanged.

[0120] Simultaneously, passenger flow matching and determination are performed: Peak passenger flow determination markers for the destination scenic area are read. When the peak passenger flow determination marker indicates that the peak threshold has been reached, the product's site attributes and tour theme are identified. Products whose destination is a peak passenger flow core scenic area and whose itinerary type is a regular sightseeing tour are marked as high-passenger-flow-restricted products, and their weight score is reduced by multiplying it by a preset high-passenger-flow attenuation coefficient. For products whose destination is a peak passenger flow core scenic area but whose itinerary type is off-peak time or niche route, their weight score is increased by multiplying it by a preset off-peak gain coefficient.

[0121] For products whose peak passenger flow determination is below the peak threshold, their weight scores remain unchanged. After dual adjustments based on weather matching and passenger flow matching, the weight scores of each product are weighted and synthesized according to a preset ratio of weather weight and passenger flow weight to obtain the environmentally corrected product weight score. The first correction candidate product set is reordered according to the environmentally corrected product weight scores to generate the second correction candidate product set. The second correction candidate product set is written into the correction intermediate buffer, overwriting the first correction candidate product set, and serves as the input for distance weight adjustment in step S440.

[0122] Step S440: Adjust the distance weight of the second correction candidate product set based on the geographic location factor to obtain the third correction candidate product set; Specifically, the city-level geocode of the user's current de-identified location is read from the dynamic context cache. The second correction candidate product set generated in step S430 is read from the correction intermediate cache. The origin layer code and destination business type code are extracted from the product basic attribute tags of each product in the second correction candidate product set.

[0123] The system compares the user's current city code with the departure city code of each product: if the user's current city code matches the product's departure city code, the user is determined to be in their place of residence. If the user's current city code does not match the product's departure city code, the user is determined to be in a travel scenario in another location.

[0124] For users in their place of residence, products whose departure city code matches the user's current city code from the second set of corrected candidate products will have their weight scores unchanged. For cross-city departure products whose departure city code does not match the user's current city code, they will be judged as products that do not conform to the user's travel habits in their place of residence, and their weight scores will be reduced by multiplying them by a preset cross-city attenuation coefficient. The attenuation range is positively correlated with the distance range between the departure city and the user's place of residence.

[0125] For users traveling in different locations, localized products whose destination city code matches the user's current city code are selected from the second set of candidate products for correction. These products are then identified as nearby local cultural and tourism resources. Their weighted scores are multiplied by a preset nearby local gain coefficient to increase the weight. The gain is set according to the degree of matching between the product destination and the user's current location.

[0126] For long-distance intercity products whose destination city code does not match the user's current city code, these are deemed invalid products as they exceed the user's daily round-trip or short-distance travel range, and are directly blocked and removed from the candidate product set. After geolocation matching and distance weight adjustment, the remaining products are re-ranked according to the adjusted weight scores to generate a third corrected candidate product set.

[0127] When the number of products in the third correction candidate product set is lower than the preset minimum recommended number threshold, some products are retrieved from the blocked and removed long-distance cross-city products in descending order of their original weight scores to make up the number. The third correction candidate product set is written into the correction intermediate buffer as input for step S450 multi-objective collaborative equilibrium sorting.

[0128] Step S450: Based on the preset multi-dimensional business indicators, perform target collaborative equilibrium sorting on the third correction candidate product set, and output the final recommendation queue.

[0129] Among them, the multi-dimensional business indicators refer to the quantitative indicator system of three core business objectives that need to be optimized in the platform operation, including user experience objectives, transaction conversion objectives and inventory turnover objectives; the objective coordination and balance ranking refers to the weighted comprehensive score of each product under the three business indicator dimensions, so that no single indicator overly dominates the ranking result, and the ranking method achieves a balance between the demands of the three parties. Specifically, the third calibration candidate product set generated in step S440 is read from the calibration intermediate cache. Preset multi-dimensional operational indicator weight parameters are loaded from the platform operational indicator configuration library. These weight parameters include user experience target weight, transaction conversion target weight, and inventory turnover target weight, with the sum of these three weights being a preset normalized total.

[0130] For each product in the third correction candidate product set, calculate the scores of the three categories of indicators. Calculation of user experience indicator score: Extract the weighted comprehensive score of the product in step S340 from the product records of the candidate product set as the personalized matching base value. Multiply this base value by the product weight score ratio coefficient after correction in the current steps S420 to S440 to obtain the user experience indicator score.

[0131] Transaction conversion metric score calculation: Extract the high reputation and high fulfillment rate labels and commission tier information of products from the product dynamic real-time operation tags. Assign a basic transaction conversion score to products marked as having a high fulfillment rate, and then make a linear adjustment according to the ratio of its commission tier to the platform's average commission tier to obtain the transaction conversion metric score.

[0132] Inventory turnover index score calculation: Extract the number of days for near-expiry clearance and the remaining inventory from the product dynamic real-time operation tags. For products with near-expiry clearance days lower than the preset near-expiry threshold, assign an inventory turnover urgency score according to the gradient. For products with remaining inventory higher than the preset backlog threshold, assign an inventory turnover backlog score according to the gradient. The urgency score and the backlog score are weighted and combined to form the inventory turnover index score.

[0133] The scores of each product across the three categories of indicators are multiplied by their respective platform operation indicator weights, and then weighted and summed to obtain a multi-objective collaborative comprehensive score. The third set of candidate products for correction is then sorted in descending order based on this multi-objective collaborative comprehensive score. After sorting, a predetermined number of products are selected from the top-ranked products to generate the final recommendation queue. Each product record in the final recommendation queue contains a unique product identifier, the multi-objective collaborative comprehensive score, and summary information required for product display. The final recommendation queue is output from the correction intermediate cache to the recommendation interface layer for display and distribution across multiple platforms.

[0134] The method provided in this embodiment acquires four types of dynamic context factors—time-series scenarios, meteorological environment, geographical location, and passenger flow and business types—and performs step-by-step weight correction on the candidate product set. This overcomes the shortcomings of existing recommendation architectures, which lack full-dimensional context awareness capabilities, leading to a disconnect between recommended content and users' actual travel scenarios. The method dynamically aligns recommendation results with travel rhythms, real-time weather, and passenger flow conditions. By using meteorological environment factors and passenger flow and business type factors to differentiate between restricted and priority scenario products, it avoids ineffective pushes of outdoor products during severe weather and products crowding scenic spots during peak hours, improving the accuracy of matching recommended content with travel scenarios and enhancing users' willingness to fulfill their commitments. Furthermore, by using preset multi-dimensional operational indicators to perform a synergistic and balanced ranking of the corrected candidate product set based on user experience, transaction conversion, and inventory turnover, it solves the problem that single-indicator ranking cannot simultaneously address the diverse operational needs of different parties. This ensures that the final recommendation queue satisfies personalized matching while also considering the platform's overall operational efficiency.

[0135] In one embodiment, the output final recommendation queue includes: The final recommendation queue is allocated concurrently in real time, and then accurately divided and distributed to multiple terminal ports for display. The multiple terminal ports include homepage recommendation slots, travel pop-up cards, private domain message pushes, and personalized special pages.

[0136] Among them, concurrent real-time computing power allocation refers to the dynamic allocation of computing resources based on the request concurrency of each port in high-load scenarios such as peak travel seasons when millions of users access the site simultaneously, ensuring that the recommendation queues of each port complete the distribution within millisecond-level response latency. Specifically, the dedicated computing power scheduling server receives the final recommendation queue output from step S450 from the recommendation interface layer. The final recommendation queue contains a product list sorted in descending order of multi-objective collaborative comprehensive score and summary information for each product. The load monitoring module of the dedicated computing power scheduling server collects the current request concurrency of each port, the CPU and GPU utilization of each computing node, and the memory usage in real time.

[0137] When the concurrent request volume on a certain port exceeds the preset concurrency threshold, the load monitoring module triggers an elastic computing power allocation strategy, allocating additional GPU inference acceleration units from the idle computing node pool to the push channel of that port to ensure that the distributed computing tasks on that port are completed within the preset latency limit. When the peak concurrency falls below the normal threshold, the allocated computing resources are returned to the idle node pool to achieve peak shaving and bottom-line protection. The dedicated computing power scheduling server reads the user profile features of the currently requesting user from the profile feature repository, extracting user stratification identifiers and destination type preferences and consumption levels from travel preference features.

[0138] Based on user segmentation identifiers, the priority configuration for reaching a user across multiple platforms is determined: highly active users are reached first through private domain message push channels, moderately active users are reached through homepage recommendations and travel pop-up cards, and low-active users are reached through personalized special pages.

[0139] Based on the display requirements of each platform, the final recommendation queue is segmented and adapted to different product groups: The homepage recommendation slots require the display of the preset homepage display limit, with product summaries including the main image, title, price range, and travel duration. A corresponding number of products are extracted from the final recommendation queue in sorted order and used to populate the display fields. Travel pop-up cards require the display of the preset pop-up display limit, with product summaries including the main image, title, and limited-time offer label. Products with "near-expiry clearance" or "holiday premium" tags are prioritized for pop-up cards. Private domain message pushes require the display of the preset private domain display limit, with product summaries including the title, key selling points, and a one-click redirect link. Personalized special pages require the display of the preset special page display limit, with the final recommendation queue divided into multiple themed product groups based on destination business type or travel theme.

[0140] The dedicated computing power scheduling server serializes the adapted product display list and display field data for each port into the corresponding port's response format, and distributes it in parallel to the front end for rendering and display through the independent push channels of each port. After distribution is complete, the dedicated computing power scheduling server writes the distribution logs of each port to the recommendation distribution log storage, recording the distribution timestamp, port identifier, user anonymized session identifier, list of unique identifiers of distributed products, and computing power resource usage statistics, for subsequent review of recommendation performance and optimization of computing power planning.

[0141] The method provided in this embodiment allocates high-concurrency real-time computing power to the final recommendation queue through a dedicated computing power scheduling server. It dynamically adjusts GPU inference acceleration units and implements elastic peak shaving based on the concurrency of requests from each port, overcoming the drawback of increased response latency due to computing power bottlenecks in recommendation services during peak travel seasons with high concurrency. This ensures millisecond-level real-time response and stable availability of recommendation distribution. Through precise segmentation and distribution based on user profile hierarchical identification and travel preference characteristics, the same recommendation queue is adapted to different ports such as homepage recommendation slots, travel pop-up cards, private domain message pushes, and personalized special pages according to user activity and preference differences. This solves the problem of low reach efficiency caused by traditional unified push notifications across all platforms neglecting port format differences, and improves the acceptance and interaction willingness of users at different activity levels across various reach channels.

[0142] In one embodiment, after the final recommendation queue is output, the system further includes: The system collects de-identified interaction signals from users in real time for the final recommendation queue. These de-identified interaction signals include clicks and pauses, quick swipes, adding to favorites and placing orders, and refunds / modifications / negative reviews. Among them, the desensitized interaction signal refers to the interactive feedback data formed by collecting users' operational behavior on the front-end recommendation interface and removing sensitive information such as personal identifiers and location through a compliant desensitization channel. This data is used for online model learning without infringing on user privacy.

[0143] Specifically, after the final recommendation queue is distributed to each terminal for display, a unified event tracking SDK deployed on the front end is used to start listening for recommendation interaction events. The types of events to be listened to include product card exposure, clicks, dwell time statistics, quick swipe skip, favorites, order placement, unsubscription requests, rescheduling requests, and review submissions. For each interaction event listened to, the front end SDK extracts the event type, the corresponding unique product identifier, the event trigger timestamp, and behavioral parameters such as interaction duration or swiping rate. The extracted data is then processed locally on the terminal through an anonymization channel to remove the user's unique identifier, terminal device identifier, and precise location information, retaining only the anonymized session identifier, behavior type code, product unique identifier, and behavior parameter values, which are then encapsulated into an anonymized interaction signal.

[0144] The anonymized interaction signals are reported to the platform's data receiving gateway in real time via an encrypted transmission channel. The platform's data receiving gateway verifies the legitimacy of the anonymized interaction signals, checking the integrity of fields and the validity of behavior type codes. For verified anonymized interaction signals, a server-side receiving timestamp is appended before writing them into the online learning message queue. The online learning message queue caches anonymized interaction signals in a first-in, first-out order. Each message contains an anonymized session identifier, behavior type code, product unique identifier, and behavior parameters. The online learning message queue provides continuous signal input to subsequent incremental parameter update modules, marking messages as processed after consumption.

[0145] The desensitized interaction signal is fed back to the dual-engine hybrid recommendation model to incrementally update the model's internal weights, user profile feature ratios, context correction coefficients, and multi-objective ranking weighting parameters, thereby enabling the model to continuously learn incrementally online.

[0146] Among them, reverse feedback refers to the closed-loop learning mechanism that uses front-end user interaction signals as feedback data to input back into the parameter update link of the recommendation model, so that the model can adjust its internal parameter configuration according to the actual user behavior results; the internal weight of the model refers to the similarity weight coefficient of the interaction co-occurrence calculation layer of the collaborative filtering sub-engine and the connection weight matrix of each hidden layer of the deep neural network of the deep learning representation sub-engine in the dual-engine hybrid recommendation model.

[0147] Specifically, the de-identified interaction signals in the online learning message queue are consumed one by one by the incremental parameter update module in the order of arrival time. The incremental parameter update module parses the behavior type code, product unique identifier, and behavior parameters of each de-identified interaction signal, and determines the positive or negative polarity of the signal based on the behavior type code: clicks and stays, and adding to favorites and placing orders are marked as positive signals, while quick swipes away and cancellations / reviews are marked as negative signals. Using the product unique identifier as the association key, the intermediate calculation parameters of the product at each stage of the recommendation chain are traced back from the recommendation log storage, including the recall score and similar user group identifier of the collaborative filtering sub-engine, the implicit travel association feature vector and cosine similarity score of the deep learning representation sub-engine, the correction coefficient values ​​of each correction stage, and the original scores of each indicator in the multi-objective ranking stage.

[0148] Incremental updates to the model's internal weights: When a positive signal is triggered, the preference weights of similar user groups within the same circle corresponding to the product in the collaborative filtering sub-engine are slightly increased, with the increase step size calculated as the product of the preset learning rate and the signal confidence. At the same time, the category weights corresponding to the product in the output layer of the deep learning representation sub-engine are increased by the same step size. When a negative signal is triggered, the weights are decreased by the same step size.

[0149] Incremental updates to user profile feature ratios: Analyze the audience matching type and game theme style in the product multi-dimensional tags corresponding to positive signals, slightly increase the weight ratio of the implicit preference feature components in the user profile corresponding to the matching type and theme style, and correspondingly decrease the weight ratio of explicit essential feature ratios in proportion, so as to keep the sum of feature ratios normalized.

[0150] Incremental updates to context correction coefficients: The distribution of positive and negative signals under various correction scenarios is statistically analyzed. When the proportion of positive signals in a specific weather or passenger flow scenario consistently exceeds a preset self-learning threshold, the gain coefficient for that scenario is slightly increased by a preset adjustment step size; conversely, it is decreased by the same step size. Incremental updates to multi-objective ranking weighting parameters: The distribution characteristics of products in the positive signals across various operational indicator dimensions are statistically analyzed. When the proportion of positive signals for products with high inventory turnover scores consistently exceeds a preset threshold, the weight of the inventory turnover target is slightly increased, while the weights of other targets are proportionally decreased, achieving dynamic adaptive balance of the weights of each target.

[0151] After all incremental updates are completed, the updated model internal weights, user profile feature ratios, context correction coefficients, and multi-objective ranking weighting parameters are written back to their respective model parameter storage areas, overwriting the previous parameter versions. The updated parameters automatically take effect when the next round of recommendation requests occurs, affecting the overall recommendation calculation process in subsequent steps, thus completing the closed-loop iteration of continuous incremental learning of the model online.

[0152] The method provided in this embodiment collects anonymized user interaction signals with the final recommendation queue in real time and feeds them back to the dual-engine hybrid recommendation model. This allows for incremental iterative updates to the model's internal weights, user profile feature ratios, context correction coefficients, and multi-objective ranking weighting parameters. This overcomes the shortcomings of existing recommendation models, which are statically locked during the day and unable to capture real-time user behavior anomalies, enabling unattended online adaptive optimization around the clock. By mapping positive and negative interaction signals to incremental adjustments of corresponding parameters, the method ensures that the adjustment direction of model parameters is causally consistent with real user feedback, avoiding the recommendation effect degradation caused by the training cycle lag in traditional full offline retraining. Furthermore, by synchronously introducing incremental iterative updates at the context correction coefficient and multi-objective ranking weighting parameter levels, the scenario correction strategy and operational indicator weight ratios continuously and adaptively converge with user behavior feedback, reducing the frequency of manual adjustments and the computational cost of nighttime full retraining.

[0153] like Figure 6 As shown, the present invention also provides a smart travel product recommendation device, applied to any of the above-described smart travel product recommendation methods, comprising: A multidimensional data acquisition module is used to collect multidimensional travel data of users, perform feature fusion modeling on the multidimensional travel data, and obtain user profiles. The product tag archiving module is used to acquire multi-dimensional tag data of travel products, perform dimensional hierarchical archiving of the multi-dimensional tag data, and obtain a product resource pool. A hybrid recommendation recall module is used to input the user profile and the product resource pool into a dual-engine hybrid recommendation model for bidirectional matching and recall, generating a candidate product set. The context correction and sorting module is used to obtain dynamic context factors, perform weight correction and multi-objective balanced sorting on the candidate product set based on the factors, and output the final recommendation queue.

[0154] This application also provides a computer device, such as... Figure 7 As shown, the computer device includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the processor executes the computer program, it implements the steps in any of the above method embodiments, or when the processor executes the computer program, it implements the functions of each module / unit in the above device embodiments.

[0155] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete this application. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the computer device.

[0156] Those skilled in the art will understand that Figure 7 The computer device described is merely an example and does not constitute a limitation on the computer device. It may include more or fewer components than shown, or combine certain components, or different components. For example, the computer device may also include input / output devices, network access devices, buses, etc.

[0157] The aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs). Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0158] The memory can be an internal storage unit of the computer device, such as a hard drive or RAM. The memory can also be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units of the computer device.

[0159] This application also provides a readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0160] This application provides a computer program product that, when run on an electronic device, enables the electronic device to perform the steps described in the various method embodiments above.

[0161] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0162] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0163] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0164] In the embodiments provided in this application, it should be understood that the disclosed apparatus / devices and methods can be implemented in other ways. For example, the apparatus / device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0165] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0166] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for intelligent recommendation of travel products, characterized in that, include: Collect multidimensional travel data from users, perform feature fusion modeling on the multidimensional travel data, and obtain user profiles; Obtain multidimensional tag data of travel products, and archive the multidimensional tag data by dimension layering to obtain a product resource pool; The user profile and the product resource pool are input into a dual-engine hybrid recommendation model for bidirectional matching and recall to generate a candidate product set. Obtain dynamic context factors, perform weight correction and multi-objective balanced ranking on the candidate product set based on the factors, and output the final recommendation queue.

2. The intelligent recommendation method for travel products according to claim 1, characterized in that, The process involves collecting multidimensional travel data from users, performing feature fusion modeling on the multidimensional travel data, and obtaining user profiles, including: Collect users' static attribute data and behavioral tracking data, and perform anomaly denoising and missing feature completion on the static attribute data and behavioral tracking data to obtain multi-dimensional travel data; The multidimensional travel data is decomposed into explicit essential needs features and implicit preference features, and a user profile is constructed by weighted fusion of the explicit essential needs features and the implicit preference features. The user profile supports real-time lightweight incremental updates. When new clicks, searches, or add-to-cart actions are collected from users, the feature weights of the user profile are fine-tuned accordingly to keep the profile synchronized with the user's real-time travel decision-making needs.

3. The intelligent recommendation method for travel products according to claim 1, characterized in that, The process involves acquiring multidimensional tag data for travel products, performing dimensional hierarchical archiving of the multidimensional tag data, and obtaining a product resource pool, including: Obtain the multidimensional tag data of all travel products listed on the platform; The multidimensional tag data is divided into basic attribute tags, rigid tags for audience adaptation, and dynamic real-time operation tags. The basic attribute tags are categorized and archived to generate a product basic attribute index; The rigid tags for audience adaptation are automatically matched and labeled to generate a product audience adaptation index; The dynamic real-time operation tags are synchronized and labeled in real time to generate a product real-time operation status index; The product basic attribute index, the product audience adaptation index, and the product real-time operation status index are jointly encapsulated to construct a multi-dimensional hierarchical tag system, forming the product resource pool.

4. The intelligent recommendation method for travel products according to claim 1, characterized in that, The step of inputting the user profile and the product resource pool into a dual-engine hybrid recommendation model for bidirectional matching and recall to generate a candidate product set includes: The user profile and the product resource pool are input into the dual-engine hybrid recommendation model, which includes a collaborative filtering sub-engine and a deep learning representation sub-engine that are launched in parallel. The collaborative filtering sub-engine matches the user profile with popular products in the product resource pool to generate a first candidate product subset. The deep learning representation sub-engine is used to mine latent features and construct vector representations for the user profile and the product resource pool to generate a second candidate product subset. The first subset of candidate products and the second subset of candidate products are weighted and merged to generate a candidate product set.

5. The intelligent recommendation method for travel products according to claim 4, characterized in that, The step of matching the user profile with the product resource pool using the collaborative filtering sub-engine to generate a first candidate product subset includes: The interaction co-occurrence calculation layer of the collaborative filtering sub-engine performs co-occurrence analysis on the travel preference features in the user profile and the historical interaction data of other users in the product resource pool to determine the similar user group in the same circle of the user image. The collaborative filtering sub-engine filters interactive travel products from the product resource pool that belong to similar user groups in the same circle, and sorts them by interaction frequency to generate the first subset of candidate products.

6. The intelligent travel product recommendation method according to claim 4, characterized in that, The step of generating a second candidate product subset by mining latent features and constructing vector representations of the user profile and the product resource pool through the deep learning representation sub-engine includes: The user profile and the product resource pool are input into the latent feature mining layer of the deep learning representation sub-engine, and the latent feature mining layer is used to mine the latent travel association features between the user profile and the product resource pool; The implicit travel-related features are input into the vector representation construction layer of the deep learning representation sub-engine to construct vectors, and are represented as user feature vectors and product feature vectors. The similarity between the user feature vector and each product feature vector is calculated through the long-tail product recall layer. Based on the similarity, the target travel products in the product resource pool are recalled to generate the second candidate product subset.

7. The intelligent recommendation method for travel products according to claim 1, characterized in that, The process of obtaining dynamic context factors, performing weight correction and multi-objective balanced ranking on the candidate product set based on the factors, and outputting the final recommendation queue includes: Obtain dynamic context factors, which include time-series scene factors, meteorological environment factors, geographical location factors, and passenger flow and business type factors; The initial weights of each travel product in the candidate product set are adjusted based on the time-series scenario factors to obtain the first corrected candidate product set; Based on the meteorological environmental factors and the passenger flow and business type factors, the first correction candidate product set is adjusted for environmental weights to obtain the second correction candidate product set. Based on the geographical location factor, the second corrected candidate product set is adjusted by distance weight to obtain the third corrected candidate product set; Based on preset multi-dimensional business indicators, the third set of candidate products for correction is sorted by target collaborative equilibrium, and the final recommendation queue is output.

8. A smart travel product recommendation device, characterized in that, The intelligent travel product recommendation method applied to any one of claims 1-7 includes: A multidimensional data acquisition module is used to collect multidimensional travel data of users, perform feature fusion modeling on the multidimensional travel data, and obtain user profiles. The product tag archiving module is used to acquire multi-dimensional tag data of travel products, perform dimensional hierarchical archiving of the multi-dimensional tag data, and obtain a product resource pool. A hybrid recommendation recall module is used to input the user profile and the product resource pool into a dual-engine hybrid recommendation model for bidirectional matching and recall, generating a candidate product set. The context correction and sorting module is used to obtain dynamic context factors, perform weight correction and multi-objective balanced sorting on the candidate product set based on the factors, and output the final recommendation queue.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent recommendation method for travel products as described in any one of claims 1 to 7.

10. A readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the intelligent recommendation method for travel products as described in any one of claims 1 to 7.

Images