Merchant recommendation method and system based on multi-dimensional label system and mixed matching strategy, and storage medium

Abstract

The application discloses a merchant recommendation method and system based on a multi-dimensional label system and a mixed matching strategy and a storage medium, belongs to the technical field of big data processing and intelligent recommendation, and comprises the steps of data collection, label construction, matching calculation, comprehensive scoring and list generation. In the scheme, firstly, full original business data on the two sides of users and merchants is collected, structured user and merchant portraits are generated based on a three-layer multi-dimensional label system, a differentiated similarity calculation method is used for labels of different data types to obtain sub-item matching scores, and finally, a recommendation list is output after sorting by combining with a scenario-based dynamic weight and distance factor calculation final recommendation score. Through the establishment of the multi-layer label system, the application realizes fine description of user and merchant characteristics, simultaneously considers recommendation accuracy and response efficiency, adapts the matching logic to multi-scene differentiated needs, and greatly improves the matching degree of the recommendation result and user satisfaction.

Description

Technical Field

[0001] This invention relates to the field of big data processing and intelligent recommendation technology, and in particular to a merchant recommendation method, system and storage medium based on a multi-dimensional tagging system and a hybrid matching strategy. Background Technology

[0002] In the fields of internet big data processing and artificial intelligence recommendation technology, merchant recommendation systems serve as an important bridge connecting users and merchants. Their performance and accuracy directly affect the user experience and commercial value of the platform. With the booming development of local life service platforms, how to efficiently and accurately recommend merchants that meet users' personalized needs has become a hot topic in industry research.

[0003] Traditional recommendation systems rely heavily on collaborative filtering algorithms or content-based recommendation strategies. These methods achieve initial matching between users and merchants to a certain extent, but their limitations gradually become apparent when dealing with complex and ever-changing user needs and merchant characteristics.

[0004] Firstly, at the tag system design level, most solutions adopt a flat tag structure, which provides a coarse-grained characterization of user needs and fails to cover differentiated needs across multiple dimensions such as spending power, scenario attributes, and time characteristics. Furthermore, the tag update mechanism does not differentiate features with different timeliness, making it difficult to adapt to short-term changes in user preferences and dynamic adjustments in merchant operational status. Secondly, at the matching logic level, most solutions use a uniform similarity calculation rule for tags of different data types, which easily leads to insufficient matching accuracy for discrete features and poor adaptability between aggregate and continuous features. This results in the core needs being diluted by non-core dimensions. In addition, existing solutions often directly perform matching calculations on all merchants without setting up a pre-recall filtering stage, which can easily lead to excessively high response latency in high-concurrency scenarios, making it difficult to balance recommendation accuracy and operational efficiency. Summary of the Invention

[0005] To address the shortcomings mentioned in the background technology, we propose a merchant recommendation method, system, and storage medium based on a multi-dimensional tagging system and a hybrid matching strategy.

[0006] The technical solution mainly consists of a merchant recommendation method based on a multi-dimensional tagging system and a hybrid matching strategy, characterized by the following steps:

[0007] Data collection: Collect registration information, historical order data, and click logs of target users, as well as basic attributes, menu data, and operational data of candidate merchants, as the raw data source for end-to-end calculation;

[0008] Tag Construction: A three-tier tag system is built, comprising primary business dimensions, secondary subdivision dimensions, and tertiary calculation indicators. User profiles and merchant profiles are generated based on this system. The primary business dimensions include four categories: spending power, taste preferences, scenario attributes, and time attributes. The taste preference dimension is further subdivided into cuisine subdivisions, and the spending power dimension is further subdivided into price sensitivity and discount sensitivity subdivisions. The secondary dimensions are the business breakdown items of the primary dimensions, and the tertiary indicators are the specific quantitative values ​​of the tags.

[0009] Matching calculation: For similar tags in user and merchant profiles, differentiated similarity calculation rules are applied according to four data types: discrete, set, continuous numerical, and time period, to obtain the sub-matching scores for each tag dimension.

[0010] Comprehensive Score: Obtain contextual dynamic weights corresponding to dining time, user location, and consumption scenario. Sum the sub-matching scores of each tag dimension according to their corresponding weights to calculate the content matching degree. Convert real-time geographical distance into distance matching score according to the rule that the closer the distance, the higher the score. Sum the distance matching score with the content matching degree according to a preset ratio to obtain the final recommendation score. The dynamic weights are generated by a pre-trained multilayer perceptron model, and the fusion ratio satisfies that the content matching degree accounts for a higher percentage than the distance score.

[0011] List generation: Sort candidate merchants from highest to lowest based on their final recommendation score, and output the recommendation list.

[0012] Preferably, the specific method for constructing a user profile in the tag construction step includes:

[0013] Using a preset fixed decay coefficient as the base, and the difference between the historical order placement time and the current time as the exponent, the time decay factor of the corresponding order is calculated. Based on the user's historical order data, the frequency of occurrence of each flavor keyword is counted. The flavor preference intensity is calculated in combination with the time decay factor. Keywords with intensity exceeding a preset threshold are selected to generate flavor tags. The preset threshold is a pre-set intensity threshold.

[0014] Based on the user's historical order amount, calculate the mean and standard deviation of the amount, and generate price sensitivity labels and consumption stratification labels;

[0015] Based on the distribution of users' order placement times, clustering algorithms are used to generate tags for users' active time periods.

[0016] Preferably, the specific method for constructing merchant profiles in the tag construction step includes:

[0017] Natural language processing is used to extract keywords related to ingredients, flavors, and cooking methods from merchants' menu names and product descriptions.

[0018] By combining merchants' historical order review data, sentiment analysis and weighting are performed on the extracted keywords to generate merchants' taste characteristic tags and signature dish tags.

[0019] Preferably, in the matching calculation step, the differential similarity calculation algorithm for different data types includes:

[0020] For discrete tags, determine whether the user tag value and the merchant tag value are completely consistent. If they are consistent, obtain the full score of the discrete tag matching item; if they are not consistent, no score is awarded.

[0021] For set-type tags, the ratio of the number of elements in the intersection of the user tag set and the merchant tag set to the number of elements in the union is calculated and used as the Jaccard similarity coefficient score.

[0022] For continuous numerical tags, the absolute value of the difference between the user preference value and the merchant attribute value is calculated, and normalization is performed based on a preset maximum tolerance threshold to obtain the corresponding matching score. The preset maximum tolerance threshold is a pre-set numerical threshold.

[0023] For time-based tags, the overlap between the user's preferred time period and the merchant's business hours is calculated to obtain the corresponding matching score.

[0024] Preferably, in the comprehensive scoring step, the tag dimension weight allocation strategy corresponding to the content matching degree is as follows:

[0025] The combined weight of the cuisine preference and flavor preference tags is the highest among all dimensions;

[0026] The weighting of the price sensitivity tag is lower than that of the cuisine and taste preference dimensions;

[0027] The weighting of the time-period preference tag is lower than that of the price preference dimension;

[0028] The weighting of the discount sensitivity tag is the lowest among all dimensions;

[0029] The sum of the weights of each dimension is one. Each weight can be adjusted according to the dining time, user's geographical location, and consumption scenario attributes corresponding to the recommendation request. The proportion of each dimension can be adjusted through the pre-trained multilayer perceptron model. The priority level of the weights remains unchanged. The priority level refers to the fixed sorting rule of the weights of each dimension.

[0030] Preferably, the method further includes a recall filtering step, specifically:

[0031] Before calculating the final recommendation score, the inverted index mechanism of the search engine is used to initially filter out a set of candidate merchants from the full merchant database with high-weight tags in the user profile as query conditions. These high-weight tags are those whose weight ratio exceeds a preset classification threshold, which is a pre-set weight ratio judgment threshold. The basic matching degree is the matching rate between the candidate merchants and the user's high-weight tags, and the matching rate is the proportion of the number of matched tags to the total number of the user's high-weight tags. The preset recall threshold is a pre-set matching rate threshold.

[0032] Preferably, the method further includes a step of dynamically updating the tags, specifically:

[0033] For users' short-term behavior tags, a streaming computing framework is used to update them in real time. The update frequency meets the timeliness requirements of short-term behavior. The short-term behavior tags refer to tags related to the user's behavior in the past seven days.

[0034] For users' long-term preference tags and merchant tags, a batch processing calculation framework is used to update them periodically. The update frequency is adapted to the stability requirements of long-term features. The long-term preference tags refer to tags that are related to the user's behavior over the past three months or more.

[0035] This invention also provides a merchant recommendation system based on a multi-dimensional tag system and a hybrid matching strategy, including a business data storage module, a profile feature calculation module, a recommendation request response module, and a tag data management module;

[0036] The business data storage module is configured to connect to the front-end business system data port, collect and store raw business data of user behavior, raw business data of merchant operation, and metadata of the tag system, and provide a standardized basic data source for end-to-end computing;

[0037] The profile feature calculation module is configured to carry a distributed batch processing calculation engine and a distributed stream processing calculation engine to complete the extraction of tag keywords, the calculation of dimension weights, and the dynamic update calculation of user and merchant profiles, and output standardized profile feature data.

[0038] The recommendation request response module is configured to connect to the user's recommendation request port. After receiving the recommendation query command, it executes a multi-dimensional tag matching algorithm, a comprehensive matching score calculation, and a candidate merchant ranking operation to generate a final recommendation result list and return it to the user.

[0039] The tag data management module is configured to integrate a relational database and a full-text search engine to achieve persistent storage and fast query response for user profile data, merchant profile data, and tag inverted index data.

[0040] The present invention also provides an electronic 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 method described in the above technical solutions.

[0041] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in the above technical solution.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] In this invention, a multi-layered tagging system encompassing business dimensions is constructed to achieve a refined profile of user and merchant characteristics. This system not only covers core dimensions such as spending power, taste preferences, scenario attributes, and time attributes, but also further subdivides them into specific indicators such as cuisine preferences and price sensitivity, ensuring the comprehensiveness and depth of the profile. This refined tagging approach enables the system to accurately capture the diverse needs of users and the unique attributes of merchants, laying a solid foundation for subsequent precise matching. Through a dynamic update mechanism, the system can reflect changes in user preferences and adjustments to merchant operational status in real time, ensuring the timeliness and accuracy of the recommendation results.

[0044] In this invention, the hybrid matching strategy significantly improves the accuracy and personalization of recommendations. For different types of data tags, the system designs differentiated similarity calculation algorithms, including discrete, set-based, continuous numerical, and time-based types, ensuring that all features are matched scientifically and reasonably. Simultaneously, by introducing a scenario-based dynamic weight adjustment mechanism, the system can flexibly adjust the weight allocation of each tag dimension based on real-time factors such as dining time, user location, and consumption scenario, making the recommendation results more closely aligned with the user's actual needs. This intelligent matching strategy effectively solves the problems of simplistic matching and lack of personalization in traditional recommendation systems, greatly improving user experience and satisfaction.

[0045] In this invention, by integrating recall filtering and comprehensive scoring mechanisms, the system achieves dual optimization of recommendation efficiency and quality. In the recall stage, the system uses inverted index technology to quickly filter out a set of candidate merchants that highly match the user profile, effectively narrowing the calculation scope of the subsequent fine-ranking stage and improving the recommendation response speed. In the fine-ranking stage, the system uses a comprehensive scoring model to fully evaluate the candidate merchants and generates the final recommendation list by combining content matching degree and distance matching score. This two-stage design approach ensures both the accuracy of the recommendation and the efficiency of the system's operation, enabling the system to maintain stable recommendation performance even in high-concurrency scenarios. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the overall operation of the present invention.

[0047] Figure 2 This is a flowchart illustrating the user and merchant profile construction process in this invention.

[0048] Figure 3 This is a flowchart of the matching calculation and comprehensive scoring in this invention. Detailed Implementation

[0049] The present invention will now be described in further detail with reference to the accompanying drawings and preferred embodiments.

[0050] Example 1, refer to Figure 1-3 As shown, the merchant recommendation method based on a multi-dimensional tag system and a hybrid matching strategy includes five core steps: data collection, tag construction, matching calculation, comprehensive scoring, and list generation.

[0051] In this embodiment, firstly, all raw business data from the user side and the merchant side are collected as the basic data source for subsequent full-process calculations. Then, based on a preset three-layer tag system, the raw data is feature extracted and quantified to generate structured user profiles and merchant profiles respectively. Next, for the same type of tags in the user and merchant profiles, the corresponding differentiated matching logic is executed according to the data type of the tags to obtain the sub-item matching scores of each dimension. Then, the sub-item scores are weighted and summed in combination with the scenario-based dimension weights to obtain the content matching degree. Then, a distance factor is introduced for secondary weighting to obtain the final recommendation score. Finally, the candidate merchants are sorted according to the final recommendation score and the recommendation list is output.

[0052] Furthermore, in this embodiment, a recall filtering step is set before the comprehensive scoring stage. The inverted index mechanism is used to quickly filter out the candidate merchant set with basic matching from the full merchant database, and then the candidate set is subjected to refined matching and scoring. At the same time, a dynamic tag update mechanism is set up to select the corresponding calculation framework to complete the tag update according to the timeliness requirements of the tag, so as to ensure the accuracy of the profile features.

[0053] This embodiment constructs a full-process recommendation logic covering feature extraction, accurate matching, and dynamic adjustment through the sequential operation of five core steps. It directly integrates users' multi-dimensional needs into each stage of the matching calculation, achieving deep alignment between user needs and merchant attributes. It can complete accurate matching without relying on historical interaction data, effectively solving the problem of insufficient recommendation accuracy in cold start scenarios.

[0054] In this embodiment, the hierarchical matching logic balances the accuracy and response speed of the recommendation. The three-layer tag system makes the recommendation results interpretable and adapts to the differentiated recommendation needs in multiple scenarios, greatly improving the matching degree of the recommendation results and user satisfaction.

[0055] Example 2, refer to Figure 1-3 As shown, the steps for building and dynamically updating the tag system are used to extract structured features from raw business data and generate timely user and merchant profiles.

[0056] Specifically, user profiles are first constructed based on a pre-defined three-tier tag system. The first-level business dimension corresponds to four categories: spending power, taste preferences, scenario attributes, and time attributes. The second-level subdivision dimension is a further subdivision of the first-level dimension. The third-level calculation indicator is the specific quantitative value of the tag.

[0057] For flavor preference tags, the time decay factor of the corresponding order is calculated by combining the timestamps of historical orders with the current time. The frequency of occurrence of each flavor keyword is also counted. The flavor preference intensity is calculated by combining the time decay factor. Keywords with intensity exceeding a preset threshold are selected to generate flavor tags. The specific calculation formula is as follows:

[0058] ;

[0059] In the formula, Score flavor The value represents the preference intensity for a single flavor keyword, where n is the total number of historical orders completed by the user within a preset effective time window (usually the last 3 months). i The frequency of this keyword in the user's historical orders is calculated from the user's historical order data. The menu name and description of each order are broken down into keywords, and the cumulative number of times a single flavor or ingredient keyword appears in all of the user's historical orders is counted. Each occurrence in each order increments the count by 1. λ is a preset fixed decay coefficient, and Day_Ago is the number of days from the current time to the corresponding order, calculated by the difference between the timestamps. The difference between the order's timestamp and the currently calculated timestamp is converted to a duration in days. Each historical order corresponds to an independent value.

[0060] For example: First, map the dish "Mapo Tofu" in the order to {spicy, tofu, soft and sticky}, and map "Kung Pao Chicken" to {sweet, spicy, chicken}. If we count that "spicy" appears 15 times and "sweet" appears 5 times, and calculate the score of "spicy" after normalization based on the time decay factor, if it exceeds 0.6, then we will assign the corresponding flavor tag to the user.

[0061] For price-related tags, the mean and standard deviation of the amount are calculated based on the user's historical order amount to generate price sensitivity tags and consumption stratification tags; for time-related tags, clustering algorithms are used to generate user active time period tags based on the distribution of user order times.

[0062] Then, merchant profiles are constructed: natural language processing is performed on the merchant's menu names and product descriptions to extract keywords related to ingredients, flavors, and cooking methods. Combined with the merchant's historical order evaluation data, sentiment analysis and weighting are performed on the extracted keywords to generate the merchant's flavor feature tags and signature dish tags.

[0063] After the initial tag calculation is completed, the corresponding update mechanism is selected according to the timeliness requirements of the tags: for short-term user behavior tags, a streaming computing framework is used for real-time updates, and the update frequency meets the timeliness requirements of short-term behavior; for long-term user preference tags and merchant tags, a batch processing computing framework is used for periodic updates, and the update frequency adapts to the stability requirements of long-term features.

[0064] This embodiment achieves a structured mapping between user needs and merchant attributes through the setting of a three-layer tagging system. The introduction of a time decay factor ensures that the taste preference tags can accurately reflect the user's recent preferences, and the batch-stream combined update mechanism takes into account both the timeliness of the tags and the reasonable allocation of computing resources.

[0065] In this embodiment, a standardized tagging system enables unified management and flexible expansion of features. The time decay calculation logic ensures the accuracy of user profiles. The batch-stream update mechanism ensures the timeliness of tags while significantly reducing the consumption of computing resources, effectively supporting the accuracy requirements of subsequent matching processes.

[0066] Example 3, referring to Figure 1-3 As shown, after the construction of each tag is completed, the matching calculation step is entered. The matching calculation step is used to complete the multi-dimensional alignment of user profile and merchant profile, and calculate the sub-item matching score of each dimension.

[0067] Specifically, the corresponding similarity calculation algorithm is executed for tags of different data types. At the same time, the sub-matching scores of all tag dimensions are normalized to the [0,1] interval through the corresponding matching rules to ensure that tag features of different dimensions can be directly weighted and calculated.

[0068] For discrete tags, it is determined whether the user tag value and the merchant tag value are completely consistent. If they are consistent, the full score for that item is obtained; if they are not consistent, no score is awarded for that item. The specific calculation formula is as follows:

[0069] ;

[0070] In the formula, Score cat For discrete label matching scores, U category The values ​​for the corresponding tags on the user side are directly read from the discrete tag fields corresponding to the user profile, and are generated by statistical analysis of user historical behavior data. categoryThe value of the corresponding tag on the merchant side is directly read from the discrete tag field corresponding to the merchant profile. It is generated by the merchant's basic attributes and operational data statistics. == is the equality comparison operator, which is used to determine whether the values ​​on both sides of the operator are equal.

[0071] For set-type tags, the ratio of the number of elements in the intersection of the user tag set and the merchant tag set to the number of elements in their union is calculated and used as the Jaccard similarity coefficient score. The specific calculation formula is as follows:

[0072] ;

[0073] In the formula, Score set For set-type tag matching scores, Set U This is a set of user-side tags, read from the multi-value tag fields corresponding to the user profile, and a set of tags obtained by splitting and statistically analyzing user historical behavior (such as a set of taste preferences). S This is a set of merchant-side tags, read from the multi-value tag fields corresponding to the merchant profile, and a set of tags obtained by splitting and statistically analyzing merchant menus and evaluation data (such as a set of merchant features).

[0074] │Set U ∩Set S │ is a set operation that calculates the number of elements in the intersection of two sets. │Set U ∪Set S | represents the number of elements in the union;

[0075] For example: if the user preference set is {spicy, salty} and the merchant feature set is {spicy, sweet, crispy}, then the intersection is {spicy} (1), the union is {spicy, salty, sweet, crispy} (4), and the tag matching score is 0.25;

[0076] Furthermore, for continuous numerical tags, the absolute value of the difference between the user preference value and the merchant attribute value is first calculated, and then normalized based on a preset maximum tolerance threshold to obtain the corresponding matching score. The specific calculation formula is as follows:

[0077] ;

[0078] In the formula, Score cont V is the continuous numerical label matching score. u The numerical value representing user preferences is read from the corresponding numerical label field in the user profile and calculated statistically from historical user behavior data (such as user preference for average order value). V sThe merchant-side attribute values ​​are read from the corresponding numerical tag fields in the merchant profile and calculated from the merchant's historical operational data (such as the average spending per merchant). T is the preset maximum tolerance threshold. Taking the average order value tag as an example, if the user prefers an average order value of 35 yuan, T can be set to 15 yuan, that is, the average spending per merchant can get a matching score if it is between 20 and 50 yuan.

[0079] Furthermore, for time-based tags, the matching score is obtained by calculating the proportion of overlap between the user's preferred time period and the merchant's business hours to the user's preferred time period. The specific calculation formula is as follows:

[0080] ;

[0081] In the formula, Score time For time-based label matching scores, Window U The time window representing user preferences is read from the corresponding time period labels in the user profile and is obtained by clustering the distribution of users' historical order times. S The time window for merchant operations, obtained directly from the merchant's basic attribute data, is called Overlap (Window). U Window S ) is a time overlap calculation function used to calculate the overlap duration of two time windows. Length(Window) U This is a duration calculation function that takes a time window as input and outputs the total duration of the user's preferred time window.

[0082] For example: the user's preferred time period is 11:30-13:30, the merchant's business hours are 11:00-14:00, the overlap is 120 minutes, the user's preferred time period is 120 minutes, and the time period tag matching score is 1;

[0083] This embodiment designs differentiated matching algorithms for different data type tags, constructs multi-dimensional quantitative matching logic, transforms fuzzy demand matching into calculable quantitative scores, and achieves accurate alignment of users' multi-dimensional needs.

[0084] In this embodiment, differentiated matching logic is designed for four different data types of tags: discrete, set, continuous numerical, and time period. A multi-dimensional quantitative matching system is constructed, which not only adapts to the matching rules of different types of features and greatly improves the accuracy of matching in each dimension, but also converts fuzzy demand matching into calculable quantitative scores, making the recommendation results interpretable and reproducible, effectively reducing the probability of matching deviation, and providing a reliable computational basis for subsequent comprehensive scoring and ranking.

[0085] Example 4, refer to Figure 1-3 As shown, after calculating the matching score of each tag, the process proceeds to the recall filtering and comprehensive scoring steps to quickly narrow down the candidate range and calculate the final recommendation score of the merchant, thus completing the sorting of the candidate set. Specifically, before calculating the final recommendation score, the inverted index mechanism of the search engine is used to select candidate merchants with a basic matching degree greater than the preset recall threshold from the full merchant database, using high-weight tags in the user profile as query conditions.

[0086] Subsequently, multi-dimensional matching calculations are performed on merchants in the candidate set to obtain matching scores for each dimension. Then, the preset weights for each tag dimension are obtained, and the matching scores are weighted and summed to obtain the content matching degree. Finally, the real-time geographical distance is converted into a distance matching score, which is then weighted and summed with the content matching degree according to a preset ratio to obtain the final recommendation score. The specific calculation formula is as follows:

[0087] ;

[0088] In the formula, FinalScore is the merchant's final recommendation score, m is the total number of tag dimensions participating in the weighted calculation, and Score is... i The sub-matching score for the i-th label dimension is calculated using the matching formulas for the four label categories mentioned above, W. i The preset weights for the i-th label dimension are obtained by combining the output of a pre-trained multilayer perceptron model with the scene attributes of the current request. The weights have fixed priorities. dist The distance matching score is calculated by converting the distance between the user's real-time geographic location and the merchant's geographic location according to a preset rule that the closer the distance, the higher the score. dist The distance dimension has a fixed weight, which is a preset hyperparameter for the algorithm, and its value is less than the sum of the weights of all label dimensions;

[0089] The distance matching score conversion rule is as follows: with the preset maximum effective distance (e.g., 3km) as the threshold, when the distance between the merchant and the user is less than or equal to the threshold, the score = 1 - (actual distance / maximum effective distance); when the distance is greater than the threshold, the score is 0.

[0090] The weighting follows a priority rule: the sum of the weights of cuisine preference and taste preference tags has the highest weight, the weight of price sensitivity tag is lower than that of cuisine and taste preference, the weight of time period preference tag is lower than that of price preference, and the weight of discount sensitivity tag is the lowest among all dimensions.

[0091] Finally, the candidate merchants are sorted from highest to lowest based on their final recommendation scores, and the top-ranked merchants are selected to generate a recommendation list.

[0092] This embodiment significantly reduces the computational load of the fine-ranking stage through the recall stage, the dynamic weight setting takes into account both business priority and scenario adaptability, and the quantitative weighting logic ensures the fairness and rationality of the final ranking result.

[0093] In this embodiment, the two-stage design of recall and fine ranking takes into account both the response speed and matching accuracy of the recommendation system. The dynamic weight setting can adapt to the differentiated needs in different scenarios, and the fixed weight priority setting ensures that the degree of satisfaction of core needs will not be affected by scenario adjustments, which greatly improves the rationality of recommendation results and user satisfaction.

[0094] For new users without historical behavioral data, the multi-dimensional matching logic is temporarily suspended and downgraded to the popularity recommendation rule, prioritizing the display of merchants with high positive reviews and close proximity. When a user makes their first interaction (click, search behavior), a temporary user tag is immediately generated, and the multi-dimensional matching rule can be applied on the next request, solving the recommendation matching problem in the cold start scenario.

[0095] Meanwhile, a differentiated update frequency is adopted based on the timeliness of the tags: highly time-sensitive geographic location tags are updated in streaming mode every 5 minutes, mid-term preference taste tags are updated daily, and long-term characteristic price sensitivity tags are updated in batch mode weekly. This corresponds to a tag update architecture that combines streaming and batch processing, taking into account both the timeliness of the tags and the reasonable allocation of computing resources.

[0096] Example 5, refer to Figure 1-3 As shown, the present invention also provides a merchant recommendation system based on a multi-dimensional tag system and a hybrid matching strategy, comprising four core components: a business data storage module, a profile feature calculation module, a recommendation request response module, and a tag data management module.

[0097] In this embodiment, the business data storage module first connects to all data ports of the front-end business system, collects and stores all raw business data and tag metadata, and provides a standardized basic data source for end-to-end computing;

[0098] The profile feature calculation module is based on a distributed computing engine. After a user triggers a recommendation request, it reads the corresponding user's structured profile data from the storage layer, including core tags such as category preferences, taste characteristics, price acceptance, and geographical location, providing basic input for the end-to-end calculation. Based on the inverted index mechanism of the full-text search engine, it sets hard filtering rules such as distance <3km and merchant being open for business, while superimposing scoring rules for matching cuisine and taste tags with high weight. It quickly filters out 50 candidate merchants that meet the basic conditions from the full 500 surrounding merchants, reducing the computational load of the subsequent fine ranking process. For the 50 recalled candidate merchants, it matches them one by one with the differentiated calculation rules of four types of tags: discrete, set, continuous value, and time period, to obtain the scores of each dimension. Then, it calculates the final recommendation score by weighting according to the weight rules. All candidate merchants are sorted in descending order of the final score, and the top 10 merchants are selected to generate a standardized recommendation list, which is returned to the client for rendering and display.

[0099] Furthermore, in this embodiment, the portrait feature calculation module is equipped with both a distributed batch processing engine and a distributed stream processing engine to handle the periodic update tasks of long-term tags and the real-time update tasks of short-term tags, respectively, to adapt to the calculation needs of different timeliness requirements. The tag data management module is equipped with both a relational database and a full-text search engine to support the precise query of structured data and the fuzzy retrieval needs of unstructured tags, respectively.

[0100] This embodiment constructs a full-link technical architecture that combines streaming and batch processing with hierarchical storage through the coordinated operation of four modules. It decouples offline computing, near-line updates, and online response capabilities, achieving low-latency recommendation responses in high-concurrency scenarios, while also supporting flexible expansion of the tag system and matching rules.

[0101] In this embodiment, a modular design is adopted to decouple the capabilities of each link. Each module has a clear responsibility and can be independently iterated and optimized, which greatly reduces the technical cost of subsequent function expansion and matching rule adjustment. The combined stream and batch computing architecture takes into account both the computing stability of long-term tags and the update timeliness of short-term tags. It can reasonably control the consumption of computing resources while ensuring the accuracy of profile data. The hierarchical storage mechanism takes into account the consistency of structured data storage and the response speed of unstructured tag retrieval, providing stable and efficient data support for end-to-end recommendation computing. The overall architecture has extremely high flexibility and reliability.

[0102] Example 6: The sixth embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the steps of the method in the above embodiments.

[0103] Example 7: This example provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the method described in the above example.

[0104] It should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should also be within the scope of protection of this invention.

Claims

1. A merchant recommendation method based on a multi-dimensional tagging system and a hybrid matching strategy, characterized by: Includes the following steps: Data collection: Collect registration information, historical order data, and click logs of target users, as well as basic attributes, menu data, and operational data of candidate merchants, as the raw data source for end-to-end calculation; Tag Construction: A three-tier tag system is built, comprising primary business dimensions, secondary subdivision dimensions, and tertiary calculation indicators. User profiles and merchant profiles are generated based on this system. The primary business dimensions include four categories: spending power, taste preferences, scenario attributes, and time attributes. The taste preference dimension is further subdivided into cuisine subdivisions, and the spending power dimension is further subdivided into price sensitivity and discount sensitivity subdivisions. The secondary dimensions are the business breakdown items of the primary dimensions, and the tertiary indicators are the specific quantitative values ​​of the tags. Matching calculation: For similar tags in user and merchant profiles, differentiated similarity calculation rules are applied according to four data types: discrete, set, continuous numerical, and time period, to obtain the sub-matching scores for each tag dimension. Comprehensive score: Obtain the contextual dynamic weights corresponding to dining time, user geographical location, and consumption scenario. The sub-item matching scores of each dimension are weighted and summed according to their corresponding weights to calculate the content matching degree. Real-time geographical distance is converted into distance matching score according to the rule that the closer the distance, the higher the score. The final recommendation score is obtained by weighting and summing the distance matching score with the content matching degree according to a preset ratio. The dynamic weights are generated by a pre-trained multilayer perceptron model, and the fusion ratio satisfies that the content matching degree accounts for a higher percentage than the distance score. List generation: Sort candidate merchants from highest to lowest based on their final recommendation score, and output the recommendation list.

2. The merchant recommendation method based on a multi-dimensional tagging system and a hybrid matching strategy according to claim 1, characterized in that: The specific methods for constructing user profiles in the tag construction step include: Using a preset fixed decay coefficient as the base, and the difference between the historical order placement time and the current time as the exponent, the time decay factor of the corresponding order is calculated. Based on the user's historical order data, the frequency of occurrence of each flavor keyword is counted. The flavor preference intensity is calculated in combination with the time decay factor. Keywords with intensity exceeding a preset threshold are selected to generate flavor tags. The preset threshold is a pre-set intensity threshold. Based on the user's historical order amount, calculate the mean and standard deviation of the amount, and generate price sensitivity labels and consumption stratification labels; Based on the distribution of users' order placement times, clustering algorithms are used to generate tags for users' active time periods.

3. The merchant recommendation method based on a multi-dimensional tagging system and a hybrid matching strategy according to claim 2, characterized in that: The specific methods for constructing merchant profiles in the tag construction step include: Natural language processing is used to extract keywords related to ingredients, flavors, and cooking methods from merchants' menu names and product descriptions. By combining merchants' historical order review data, sentiment analysis and weighting are performed on the extracted keywords to generate merchants' taste characteristic tags and signature dish tags.

4. The merchant recommendation method based on a multi-dimensional tagging system and a hybrid matching strategy according to claim 1, characterized in that: The matching calculation step includes differential similarity calculation algorithms for different data types, including: For discrete tags, determine whether the user tag value and the merchant tag value are completely consistent. If they are consistent, obtain the full score of the discrete tag matching item; if they are not consistent, no score is awarded. For set-type tags, the ratio of the number of elements in the intersection of the user tag set and the merchant tag set to the number of elements in the union is calculated and used as the Jaccard similarity coefficient score. For continuous numerical tags, the absolute value of the difference between the user preference value and the merchant attribute value is calculated, and normalization is performed based on a preset maximum tolerance threshold to obtain the corresponding matching score. The preset maximum tolerance threshold is a pre-set numerical threshold. For time-based tags, the overlap between the user's preferred time period and the merchant's business hours is calculated to obtain the corresponding matching score.

5. The merchant recommendation method based on a multi-dimensional tagging system and a hybrid matching strategy according to claim 1, characterized in that: In the comprehensive scoring step, the tag dimension weight allocation strategy corresponding to the content matching degree is as follows: The combined weight of the cuisine preference and flavor preference tags is the highest among all dimensions; The weighting of the price sensitivity tag is lower than that of the cuisine and taste preference dimensions; The weighting of the time-period preference tag is lower than that of the price preference dimension; The weighting of the discount sensitivity tag is the lowest among all dimensions; The sum of the weights of each dimension is one. Each weight can be adjusted according to the dining time, user's geographical location, and consumption scenario attributes corresponding to the recommendation request. The proportion of each dimension can be adjusted through the pre-trained multilayer perceptron model. The priority level of the weights remains unchanged. The priority level refers to the fixed sorting rule of the weights of each dimension.

6. The merchant recommendation method based on a multi-dimensional tagging system and a hybrid matching strategy according to claim 1, characterized in that: The method further includes a recall filtering step, specifically: Before calculating the final recommendation score, the inverted index mechanism of the search engine is used to initially filter out a set of candidate merchants from the full merchant database with high-weight tags in the user profile as query conditions. These high-weight tags are those whose weight ratio exceeds a preset classification threshold, which is a pre-set weight ratio judgment threshold. The basic matching degree is the matching rate between the candidate merchants and the user's high-weight tags, and the matching rate is the proportion of the number of matched tags to the total number of the user's high-weight tags. The preset recall threshold is a pre-set matching rate threshold.

7. The merchant recommendation method based on a multi-dimensional tagging system and a hybrid matching strategy according to claim 1, characterized in that: The method also includes a step of dynamically updating the tags, specifically: For users' short-term behavior tags, a streaming computing framework is used to update them in real time. The update frequency meets the timeliness requirements of short-term behavior. The short-term behavior tags refer to tags related to the user's behavior in the past seven days. For users' long-term preference tags and merchant tags, a batch processing calculation framework is used to update them periodically. The update frequency is adapted to the stability requirements of long-term features. The long-term preference tags refer to tags that are related to the user's behavior over the past three months or more.

8. A merchant recommendation system that implements the merchant recommendation method based on a multi-dimensional tagging system and a hybrid matching strategy as described in any one of claims 1-7, characterized in that: It includes a business data storage module, a profile feature calculation module, a recommendation request and response module, and a tag data management module; The business data storage module is configured to connect to the front-end business system data port, collect and store raw business data of user behavior, raw business data of merchant operation, and metadata of the tag system, and provide a standardized basic data source for end-to-end computing; The profile feature calculation module is configured to carry a distributed batch processing calculation engine and a distributed stream processing calculation engine to complete the extraction of tag keywords, the calculation of dimension weights, and the dynamic update calculation of user and merchant profiles, and output standardized profile feature data. The recommendation request response module is configured to connect to the user's recommendation request port. After receiving the recommendation query command, it executes a multi-dimensional tag matching algorithm, a comprehensive matching score calculation, and a candidate merchant ranking operation to generate a final recommendation result list and return it to the user. The tag data management module is configured to integrate a relational database and a full-text search engine to achieve persistent storage and fast query response for user profile data, merchant profile data, and tag inverted index data.

9. An electronic 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 a computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images