Recommendation method based on multi-dimensional data of mesh points

Abstract

This invention discloses a recommendation method based on multi-dimensional data of service outlets, belonging to the field of data processing technology. The method includes: collecting basic service outlet data and performing multi-dimensional feature mining and quantification to generate multi-dimensional feature data of service outlets; receiving user demand query information, generating user demand feature elements through semantic parsing and feature extraction, and assigning weighted user demand feature vectors according to demand priority weights; calculating the similarity between the vectors and the multi-dimensional feature data of service outlets, and filtering to generate a candidate service outlet set; constructing a supply-demand matching model, performing supply-demand adaptation verification on the candidate service outlet set, eliminating unqualified service outlets to generate a suitable service outlet set; comprehensively ranking the suitable service outlet set, generating and outputting standardized service outlet recommendation results. This invention achieves accurate supply-demand matching between user needs and catering service outlets, and is applicable to service outlet recommendation needs in dining scenarios.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a recommendation method based on multi-dimensional data of network locations. Background Technology

[0002] In the technical field of connecting merchants with cooperative outlets, recommending suitable cooperative outlets for merchants based on their business needs is an important technical means to improve the efficiency of business matching. There are already a number of outlet recommendation methods based on data matching in the existing technology. The conventional implementation method is to collect the basic information of the outlets and perform simple structured processing. After receiving the user's demand query information, the demand information is extracted in a basic way. Then, the extracted demand information is directly matched with the basic information of the outlets, and the outlet recommendation results are generated by combining the results of simple similarity calculation.

[0003] However, in practical applications, existing technologies only perform superficial structuring processing on basic branch data, without conducting multi-dimensional feature mining and quantification. Furthermore, for user-requested information, only basic information extraction is performed, without standardized feature element transformation, weight allocation, or targeted supply-demand matching verification processes. As a result, the entire branch recommendation process can only achieve a superficial match between demand information and basic branch information, failing to build an accurate supply-demand matching relationship based on multi-dimensional feature data, and making it difficult to ensure that the recommendation results closely match the user's actual needs. Summary of the Invention

[0004] To address the technical problem that existing technologies often fail to provide branch recommendation results that accurately reflect users' actual needs, this invention provides a recommendation method based on multi-dimensional branch data.

[0005] The technical solution adopted in this invention is a recommendation method based on multi-dimensional data of network outlets, characterized by the following steps:

[0006] Step 100: Collect basic data of the outlets, perform multi-dimensional feature mining and quantification on the basic data of the outlets, and generate multi-dimensional feature data of the outlets.

[0007] Step 200: Receive user query information, perform semantic parsing and feature extraction on the user query information, and generate user query feature elements;

[0008] Step 300: Based on the demand priority of user demand feature elements, perform weight allocation processing on each user demand feature element to generate a weighted user demand feature vector.

[0009] Step 400: Perform similarity calculation between the weighted user demand feature vector and the multi-dimensional feature data of the network points, and generate a candidate network point set based on the similarity calculation results;

[0010] Step 500: Construct a supply and demand matching model, input the candidate site set into the supply and demand matching model to perform supply and demand adaptation verification, remove sites that fail the supply and demand adaptation verification, and generate a set of adapted sites;

[0011] Step 600: Perform comprehensive sorting on the adapted site set, generate standardized site recommendation results, and complete the result output.

[0012] The beneficial effects of this invention are as follows: This solution performs multi-dimensional feature mining and quantification processing on the collected network point basic data. Compared with the surface-level structured processing of network point data in existing technologies, it can fully mine the multi-dimensional feature information in the network point data, transforming the network point basic data into multi-dimensional feature data that can be used for accurate matching. By performing semantic parsing and feature extraction on user demand query information, and combining demand priority for weight allocation, a weighted user demand feature vector is generated. This transforms the scattered user demand information into a standardized feature vector with priority distinction. Compared with the basic extraction of demand information in existing technologies, it can more accurately capture the core needs of users, making the matching of demand features and network point features more targeted. By screening the candidate network point set through similarity calculation, and then performing supply and demand adaptation verification and comprehensive ranking through the supply and demand matching model, the output results can be effectively improved, making the adaptation degree between candidate network points and user needs more in line with the actual needs of users. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the method flow of Embodiment 1 of the present invention;

[0014] Figure 2 This is a schematic diagram of the fuzzy demand determination method in Embodiment 2 of the present invention;

[0015] Figure 3 This is a schematic diagram of the method for determining and processing fuzzy requirements in Embodiment 2 of the present invention. Detailed Implementation

[0016] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] Example 1

[0018] Considering that existing technologies, in practical applications, only perform superficial structuring processing on basic network point data, the entire network point recommendation process can only achieve a superficial match between demand information and basic network point information. It cannot build an accurate supply-demand matching relationship based on multi-dimensional feature data, making it difficult to ensure that the recommendation results closely match the user's actual needs. To solve this problem, this embodiment provides a recommendation method based on multi-dimensional network point data, including the following steps:

[0019] Step 100: Collect basic data of the network outlets, perform multi-dimensional feature mining and quantification on the basic data of the network outlets, and generate multi-dimensional feature data of the network outlets.

[0020] It should be noted that the basic data of the outlets refers to various raw data generated during the operation of catering outlets, including location data, such as the address of the catering outlet, distance from various business districts, and radiation range; business category data, such as Sichuan hot pot / clear soup hot pot / Thai hot pot, and soup base type; cost-effectiveness data, such as average consumption per person, discount activities, and group purchase packages; experience rating data, such as taste rating, environment rating, service rating, and average user reviews on the platform; operational capability data, such as operating hours, number of tables, and average daily reception; and supporting service data, such as free parking, provision of private rooms, and complimentary desserts. The multi-dimensional feature data of the outlets refers to the standardized multi-dimensional feature set that can be used for feature matching after the basic data of the catering outlets is mined and quantified. Quantification refers to mapping the raw data to standardized values ​​in the [0,1] interval through a normalization algorithm.

[0021] In the specific implementation process, basic data of various catering outlets are collected in batches through local life service platforms, catering outlet management systems, and other channels. Invalid data with ambiguous addresses, unclear categories, and abnormal ratings are cleaned and removed. Feature mining is performed from six dimensions: location, operating categories, cost-effectiveness, experience rating, operational capabilities, and supporting services. The min-max normalization algorithm is used to quantize the original data of each dimension to the same numerical range of [0,1]. The calculation formula is as follows:

[0022]

[0023] in, Let be the quantified value of the j-th feature dimension of the i-th catering outlet. Let be the original value of the j-th feature dimension of the i-th catering outlet. This represents the original minimum value for all catering outlets under this feature dimension. This represents the original maximum value for all catering outlets under this feature dimension; finally, the quantified values ​​of each dimension are integrated to form multi-dimensional feature data of the outlets.

[0024] For example, taking a Sichuan-style hot pot restaurant as an example, the basic data collected is as follows: 2.5 kilometers away from the user's business district, specializing in Sichuan-style hot pot (beef tallow broth / clear oil broth), average spending per person is 75 yuan, taste rating is 4.6, environment rating is 4.5, service rating is 4.4, business hours are 10:00-24:00, there are 30 tables, and free parking is available.

[0025] After cleaning, outlier data such as those without specific locations or those categorized as "hot pot" without further subdivision were removed. For the location dimension, if the core is within 3 kilometers, the original value is 0-3 kilometers, and outside 3 kilometers is denoted as 3. The quantified value for this restaurant location within 2.5 kilometers is... In terms of cost-effectiveness, such as an average spending range of 50-150 yuan per person (which is considered an effective range), 75 yuan is a quantifiable value. Taste rating dimensions, such as 0-5 points, with 4.6 points being a quantitative value. The remaining dimensions are quantified using the same formula; after integration, multi-dimensional feature data of this Sichuan-style hot pot restaurant network is formed.

[0026] Step 200: Receive user query information, perform semantic parsing and feature extraction on the user query information, and generate user query feature elements.

[0027] It should be noted that user demand query information refers to the dining needs expressed by ordinary users, which can be in natural language forms such as text and voice-to-text, such as "I want to find a Sichuan hot pot restaurant with a high cost-performance ratio within 3 kilometers" or "I want to eat clear broth hot pot with a taste rating of 4.5 or higher". User demand feature elements refer to the core feature items extracted from user dining demand information that can be matched with the multi-dimensional feature data of the outlets, and are divided into demand location, such as within 3 kilometers / around the business district / next to the subway station; business category, such as Sichuan hot pot / clear broth hot pot / fast food; cost-performance requirements, such as less than 80 yuan per person / group purchase packages available; experience requirements, such as taste rating ≥ 4.5 / environment rating ≥ 4.0; and supporting service requirements, such as free parking / private rooms available, etc.

[0028] In the specific implementation process, the system receives dining request information sent by users. If the information is in voice format, it first performs voice-to-text processing, performs semantic parsing on the request text, and splits the request keywords and semantic logic based on a preset dictionary of dining consumption demand characteristics. It removes colloquial expressions such as "delicious" and "good" that have no actual matching meaning. It extracts the core features corresponding to the keywords and integrates them into user demand feature elements in the order of location, business category, cost performance, experience requirements, and supporting services.

[0029] For example, a user's query is "I want to find a Sichuan hot pot restaurant within 3 kilometers that is cost-effective and has a taste rating of 4.5 or higher". After semantic parsing, four core requirements are extracted: within 3 kilometers, Sichuan hot pot, cost-effective, and taste rating of 4.5 or higher. Based on the dictionary of catering consumption demand characteristics, "cost-effective" is concretized as a quantifiable feature of "average price of 80 yuan per person". The corresponding features are extracted to form user demand feature elements, including the demand location being within 3 kilometers, the business category being Sichuan hot pot, the cost-effectiveness being 80 yuan per person, and the experience requirement being a taste rating of ≥4.5.

[0030] Step 300: Based on the demand priority of user demand feature elements, perform weight allocation processing on each user demand feature element to generate a weighted user demand feature vector.

[0031] It should be noted that demand priority refers to the importance of each user demand feature to the user's dining experience. For example, users often set the matching of business categories to a higher priority than the supporting services. The weighted user demand feature vector refers to the feature vector with priority distinction formed after assigning unique weight values ​​to each user's dining demand feature, and the sum of the weight values ​​is 1.

[0032] Because different dining needs significantly affect the user's dining experience, for example, users who want to eat Sichuan hot pot prioritize the type of food offered over whether there is a complimentary dessert. If indiscriminate matching is used, the core needs will be weakened. This step highlights the user's core dining needs through weight allocation, making the subsequent similarity calculation more targeted. At the same time, the weight allocation rules reserve a self-optimization interface, which can be iteratively adjusted based on the user's actual dining evaluation.

[0033] In the specific implementation process, based on the user's preset priority rules for dining needs, a unique weight value is assigned to each user's demand characteristic element. The sum of the weights of all feature elements is 1; each demand feature element is bound to its corresponding weight value and arranged in the order of preset dimensions such as location, business category, cost performance, experience requirements, and supporting services to generate a weighted user demand feature vector.

[0034] For example, a user sets the product category, such as Sichuan hot pot, as the highest priority, and assigns a weight of 100%. Demand location, such as within 3 kilometers, is the second priority, with weight allocation as follows: Cost-effectiveness, such as an average cost of 80 yuan per person, is the third priority, with the following weight allocation: Experience requirements, such as a taste rating of ≥4.5, are given fourth priority, with the following weighting: The total weight of all elements is 1. Combining the user demand feature elements generated in step 200, a weighted user demand feature vector is generated, in the following order: within 3 kilometers (0.25), Sichuan hot pot (0.4), average cost per person within 80 yuan (0.2), and taste rating. (0.15).

[0035] Step 400: Perform similarity calculation between the weighted user demand feature vector and the multi-dimensional feature data of the network points, and generate a candidate network point set based on the similarity calculation results.

[0036] It should be noted that similarity calculation refers to the degree of fit between the weighted user demand feature vector and the multi-dimensional feature data of each catering outlet, calculated by an algorithm; the candidate outlet set refers to the set of catering outlets whose similarity calculation results reach a preset threshold.

[0037] By quantifying the degree of fit between dining needs and the characteristics of service outlets through similarity calculation, outlets that meet the basic dining needs of users are selected, providing alternative targets for subsequent accurate verification and reducing the computational cost of invalid verification.

[0038] In the specific implementation process, the cosine similarity algorithm is used to perform similarity calculation, and the calculation formula is as follows:

[0039]

[0040] in, Let n be the similarity score of the i-th restaurant, and n be the number of feature dimensions. Let J be the weight value of the demand feature element in the j-th dimension. Quantify the user demand feature for the j-th dimension. The feature quantification value of the j-th dimension of the i-th outlet is used; a preset similarity threshold is set to filter out catering outlets with similarity higher than the threshold and integrate them into a candidate outlet set.

[0041] For example, 15 hot pot restaurants are selected, and the similarity between the multi-dimensional feature data of each restaurant and the weighted user demand feature vector generated in step 300 is calculated. The calculation results are 0.88, 0.75, 0.62, 0.82, 0.58, 0.65, 0.70, 0.55, 0.68, 0.78, 0.52, 0.61, 0.72, 0.48, and 0.66, respectively. A similarity threshold of 0.6 is set, and 11 hot pot restaurants with similarity higher than the threshold are selected. These 11 restaurants are integrated into a candidate restaurant set, and the similarity values ​​of these 11 restaurants are retained as the positive adaptation results of the subsequent supply and demand matching model.

[0042] Step 500: Construct a supply and demand matching model, input the candidate site set into the supply and demand matching model to perform supply and demand adaptation verification, remove sites that fail the supply and demand adaptation verification, and generate a set of adapted sites.

[0043] In one possible implementation, the construction of the supply-demand matching model, which involves inputting the candidate site set into the supply-demand matching model to perform supply-demand adaptation verification, eliminating sites that fail the supply-demand adaptation verification, and generating an adapted site set includes the following:

[0044] The operational capability features of candidate outlets are extracted from multi-dimensional feature data of outlets. The operational capability features are then matched with user demand features to generate a demand feasibility verification result.

[0045] The results of the demand feasibility verification and the positive adaptation results of the user demand feature elements corresponding to the candidate sites are used together as the basis for determining the supply and demand adaptation verification. The supply and demand adaptation verification is performed and an adapted site set is generated.

[0046] It should be noted that the supply and demand matching model refers to a five-layer architecture model used to verify the suitability of candidate catering outlets with users' dining needs; the supply and demand adaptation verification refers to a comprehensive verification of whether candidate catering outlets can actually meet users' dining needs, including two core dimensions: positive adaptation and reverse feasibility verification.

[0047] The adapted network set refers to the set of catering outlets that have passed the positive and negative dual-dimensional verification of the supply and demand matching model; the positive adaptation result is the similarity of each candidate catering outlet generated in step 400. This is a result that quantifies the degree of fit between the characteristics of catering outlets and user needs from the perspective of user needs. The value range is [0,1], and the higher the value, the stronger the fit.

[0048] Operational capability characteristics refer to the core capability characteristics of catering outlets that are actually able to meet users' dining needs, extracted from multi-dimensional feature data of the outlets. These characteristics are adapted to user scenarios, including table capacity, operating hours adaptability, and reception efficiency. Reverse adaptation calculation refers to the process of verifying whether the user's dining needs are within the actual capacity of the catering outlets, based on the operational capability characteristics of the outlets. The results of the demand feasibility verification are also included. This refers to the result value generated after reverse adaptation calculation, which quantitatively represents whether the user's dining needs can be actually undertaken by catering outlets. The value range is [0,1], and the higher the value, the stronger the feasibility.

[0049] Existing technologies only verify the matching between catering outlets and user needs through a single positive fit, which cannot avoid the problem of catering outlets that match the features but do not actually have the capacity to accommodate customers. For example, a Sichuan hot pot restaurant may meet the characteristics of being within 3 kilometers and costing less than 80 yuan per person, but it is full and the waiting time is more than 2 hours, which cannot meet the user's immediate dining needs.

[0050] In practical implementation, the hierarchical architecture of the supply and demand matching model is as follows:

[0051] The data input layer is the bottom layer of the model. The input data includes candidate site sets and quantified values ​​of user demand feature elements. Multi-dimensional feature data of outlets Simultaneously, retrieve the similarity scores of each candidate network point generated in step 400. This provides complete and standardized basic data for subsequent model processing.

[0052] The feature extraction layer connects to the data input layer and is used to extract the operational capability features corresponding to each candidate branch from the multi-dimensional feature data of the branches. The extracted operational capability features are then processed by the normalization algorithm in step 100 to generate a standardized operational capability feature set. Ensure that the quantitative dimensions and units are consistent with the characteristics of user needs.

[0053] The dual-dimensional verification layer connects to the feature extraction layer to complete forward and reverse dual-dimensional verification. The forward and reverse dual-dimensional verifications are processed in parallel by two verification branches.

[0054] The reverse adaptation verification branch is to standardize the operational capability feature set. Quantitative values ​​of user demand characteristics To perform reverse adaptation calculation, first calculate the single-dimensional reverse adaptation matching degree using a numerical interval matching algorithm. Then, weighted and integrated matching results from various dimensions are generated to verify the feasibility of the requirements. .

[0055] Single-dimensional reverse adaptation matching degree The calculation formula is:

[0056]

[0057] in, Let the reverse adaptation matching degree be the j-th demand dimension of the i-th candidate site. Let j be the quantified value of the j-th operational capability characteristic of the i-th candidate site. This is the quantified value of the user's j-th demand feature element.

[0058] Weighted integration of matching scores across various dimensions generates a feasibility verification result for the requirements. The calculation formula is:

[0059]

[0060] in, This represents the feasibility verification result of the requirements for the i-th candidate site, where n is the total number of dimensions of the requirement feature elements. The weight value of the j-th user demand feature element is taken from the weighted user demand feature vector in step 300. Let be the reverse adaptation matching degree of the i-th candidate site in the j-th dimension.

[0061] The forward adaptation result retrieval branch directly retrieves the similarity from the data input layer. As a positive adaptation result, the two-dimensional verification result is obtained: because the cosine similarity calculation in step 400 is based on user needs as the core benchmark, it quantifies the degree of fit between the network point features and user needs. This calculation logic is consistent with positive adaptation, that is, verifying whether the network point fits the needs based on user needs. There is no need to redesign the calculation logic, which can greatly reduce computing power consumption and improve algorithm efficiency.

[0062] The result determination layer connects to the two-dimensional verification layer to perform threshold comparison and result determination. The model presets a positive adaptation result threshold. Threshold for verification of feasibility of requirements The two results are compared with their corresponding thresholds. Only when the candidate site shows a positive fit will the result be considered. And the results of the feasibility verification of the requirements. If the supply and demand matching verification of the outlet is passed, it is determined that the verification is failed if any result fails.

[0063] The output layer connects to the result determination layer to output the adapted point set, integrate the determination results of all candidate points, remove points that fail the verification, integrate the points that pass the verification to generate the adapted point set, and complete the output of the model.

[0064] For example, the candidate location set includes 11 Sichuan hot pot restaurants, which are input into the supply and demand matching model. The data input layer simultaneously receives the quantified values ​​of user demand feature elements: 0.17 for within 3 kilometers, 0.95 for Sichuan hot pot, 0.7 for an average cost of less than 80 yuan per person, and 0.9 for a taste rating of ≥4.5. It also directly retrieves the positive adaptation results of the 11 restaurants in step 400. (0.88, 0.75, 0.62, 0.82, 0.65, 0.70, 0.68, 0.78, 0.61, 0.72, 0.66).

[0065] The feature extraction layer targets the core practical needs of users dining out, extracting three types of operational capability features for each restaurant: table capacity, operating hours adaptability, and reception efficiency. After standardization and quantification, the table capacity of a certain Sichuan hot pot restaurant is quantified as follows: (15 available tables), quantitative value of business hours adaptability (Open until 24:00, covering users' planned dining time of 20:00), Quantitative value of reception efficiency (queue time) minute).

[0066] The two-dimensional verification layer performs reverse adaptation calculations on the restaurant outlet to determine the table capacity dimension matching degree. Matching degree of business hours Matching degree of reception efficiency dimension The overall feasibility verification result is calculated by combining the weight values. The result is ≥0.5 threshold, and its positive adaptation result is... The threshold was met, and the verification was deemed successful; another restaurant outlet showed a positive adaptation result. The threshold was met, but the quantitative values ​​of the operational capacity characteristics were all 0 (tables are full, reception is suspended), and the results of the feasibility verification of the demand were obtained. A threshold is set to determine if the verification fails, and the affected element is directly removed.

[0067] Ultimately, 7 out of the 11 candidate catering outlets met the positive adaptation and demand feasibility verification results, generating an adaptation outlet set that includes these 7 Sichuan hot pot catering outlets.

[0068] Step 600: Perform comprehensive sorting on the adapted site set, generate standardized site recommendation results, and complete the result output.

[0069] It should be noted that comprehensive ranking processing refers to prioritizing the set of suitable outlets based on preset dimensions; standardized outlet recommendation results refer to standardized output results that include core information of catering outlets, user demand suitability scores, and ranking positions.

[0070] Because the compatibility of different restaurants within the matching network set still varies, directly outputting an unsorted set of restaurants would increase the user's filtering costs and reduce the efficiency of finding restaurants. This step highlights the better restaurants through comprehensive sorting, providing users with a clear and orderly basis for selection.

[0071] In the specific implementation process, the similarity of step 400 is used. (i.e., positive adaptation results) is the core dimension for ranking; the adapted network of outlets is sorted in descending order; core operational information of each restaurant outlet is supplemented. The results are integrated to form standardized outlet recommendation results containing restaurant name, core operational information, comprehensive similarity score, and ranking position, which are then delivered to users through local life service platforms, restaurant recommendation apps, and other channels.

[0072] The quantitative evaluation data collection for recommendation results involves setting up an evaluation entry point at the user's dining completion stage after the standardized recommendation results are output, and collecting effective quantitative evaluation data of the recommendation results from users. The evaluation indicators are three dimensions that fit the user's dining experience and are all mapped to the interval [0, 1]. The three dimensions of the user's dining experience include: demand satisfaction: the user's actual experience score of the recommended catering outlet matching their core dining needs, such as whether it is Sichuan hot pot, whether it is within 3 kilometers, and whether the per capita value meets expectations; experience fit: the user's comprehensive score of the actual dining experience of the recommended catering outlet, such as whether the taste, environment, and service meet psychological expectations; and dining success rate: the quantitative value of the user's actual dining situation at the recommended catering outlet, with 1 point for successful dining without queuing, 0.5 points for queuing for more than a minute, and 0 points for no seat and no dining.

[0073] When a single user accumulates 5 or more valid reviews, or when all users on the platform accumulate 100 or more valid reviews, the weight self-optimization process in step 300 is automatically triggered; 2. Optimization algorithm execution: The gradient descent algorithm is used, with the goal of maximizing the mean of the three-dimensional evaluation indicators. The weight allocation rules for the demand features in step 300 are iteratively adjusted in reverse. The optimization formula is as follows:

[0074]

[0075] in, For the first Optimized weight values ​​for each demand feature dimension. For the first The unoptimized weight values ​​for each demand feature dimension. The learning rate is set to 0.01 in this embodiment, which is unique and fixed to avoid large fluctuations in the weights. The mean of the three-dimensional evaluation indicators. This is the partial derivative of the mean of the evaluation index with respect to the weights before optimization.

[0076] The optimized weight values ​​must satisfy the non-negativity constraint and the sum of 1 constraint. If the weight value of a certain dimension is negative after optimization, it is set to 0 and the weight values ​​of the other dimensions are renormalized to ensure that the sum of the weight values ​​of all demand feature dimensions is still 1. The optimized weight allocation rule is updated to the weight allocation in step 300 as the weight allocation benchmark for subsequent new user dining needs and subsequent dining needs of the same user, realizing the self-iteration of weights. At the same time, a weight iteration log is retained to record the time of each optimization, the amount of evaluation data, and the magnitude of weight adjustment, supporting manual backtracking and manual adjustment.

[0077] In summary, this embodiment effectively solves the problem of insufficient accuracy in restaurant recommendation in existing technologies. The recommended results not only align with users' core dining needs but also possess good practical applicability. It can meet the basic application needs of efficient and accurate connection between users in the commercial sector and restaurant outlets. Furthermore, based on the weighted self-optimization process of actual restaurant evaluations, the algorithm achieves self-iterative upgrades, allowing the recommendation effect to continuously improve with actual use. It perfectly adapts to the recommendation needs of cooperative outlets in restaurant promotion scenarios, significantly improving the efficiency and success rate of connection between restaurant outlets and cooperative outlets.

[0078] Example 2

[0079] Considering that the above embodiments do not take into account scenarios where users have vague dining needs in practical applications, when a user's need is expressed in vague terms such as "I want to find a good hot pot restaurant nearby" or "I want to eat a hot pot restaurant with good value for money," it is difficult to generate accurate user need feature elements through conventional semantic parsing and feature extraction alone, leading to deviations in subsequent restaurant matching. To solve this technical problem, in this embodiment, before performing semantic parsing and feature extraction on the user need query information, such as... Figure 2 As shown, it also includes the following steps:

[0080] Step 201: Retrieve the user's historical dining data, perform structured and feature quantification processing on the historical dining data, and generate a basic consumption profile of the user.

[0081] Among them, historical dining data refers to data such as users' past dining records, favorites, and reviews on local life service platforms, including dining location preferences, category preferences, average spending range per person, taste rating requirements, and preferences for supporting services; user consumption profile refers to a standardized set of user dining preference characteristics formed after cleaning, structuring, and quantifying users' historical dining data.

[0082] Since the core of users' vague dining needs is that the expression of needs is disconnected from their own dining preferences, generating standardized consumption profiles based on historical dining data can serve as the core reference for judging and analyzing vague needs, thus solving the problem of vague dining needs having no reference and no direction.

[0083] In the specific implementation process, the historical dining data of users is retrieved through the local life service platform, and the data is cleaned and structured. It is then broken down into five core dimensions: dining location preference, business category preference, average consumption range per person, taste rating requirements, and supporting service preferences. The minimum-maximum normalization algorithm in step 100 of Example 1 is used to perform feature quantification on the data of each dimension and generate quantified values ​​for each dimension. The quantified values ​​of each dimension are then integrated to form a user consumption profile.

[0084] Step 202: Perform shallow semantic feature extraction on the user's query information to generate a shallow feature set.

[0085] It should be noted that shallow semantic feature extraction refers to the extraction method that only extracts unambiguous basic keywords from the user's dining needs information without in-depth analysis of semantic logic; shallow feature set refers to the feature set formed by integrating all unambiguous basic keywords from the user's needs information.

[0086] Because the core problem of users' ambiguous dining needs is unclear semantics and incomplete elements, over-analysis will lead to feature extraction bias. Shallow extraction can avoid this problem and retain only the reliable basic features in the needs.

[0087] In the specific implementation process, shallow semantic parsing is performed on the user's fuzzy demand query information, unambiguous basic keywords are extracted based on the dictionary of catering consumption demand characteristics, and vague expressions such as "delicious", "good", and "suitable" are eliminated; all unambiguous basic keywords are integrated to generate a shallow feature set.

[0088] For example, a user's fuzzy query information is "I want to find a good hot pot restaurant nearby"; after shallow semantic extraction, only the two unambiguous keywords "nearby" and "hot pot" are retained to generate a shallow feature set.

[0089] Step 203: Perform feature matching degree calculation between the shallow feature set and the user's basic consumption profile to generate profile correlation degree.

[0090] It should be noted that the profile correlation R refers to the quantitative value of the degree of fit between the shallow feature set and the user's basic consumption profile. The value range is [-1, 1]. The closer the value is to 1, the stronger the correlation, and the closer the value is to 0, the weaker the correlation.

[0091] In the specific implementation process, the Pearson correlation coefficient algorithm is used to calculate the profile correlation degree. The calculation formula is as follows:

[0092]

[0093] in, Let be the quantized value of the j-th dimension of the shallow feature set. The mean of the quantized values ​​of the shallow feature set. The quantified value of the i-th dimension of the user's basic consumption profile. The mean of the quantitative values ​​for the user's basic consumption profile, where n is the number of feature dimensions.

[0094] For example, the quantification values ​​of "nearby" and "hot pot" in the shallow feature set are [0.5, 0.8], and the quantification values ​​of the user's basic consumption profile are [0.9, 0.95, 0.8, 0.9, 0.7]. The calculated profile correlation degree R is 0.35, indicating that the correlation between demand and merchant operating attributes is relatively weak.

[0095] Step 204: Perform global feature matching calculation on the shallow feature set and the multi-dimensional feature data of the network points to generate the feature orientation of the network points.

[0096] It should be noted that the network feature orientation degree P refers to the quantitative value of the effective matching degree between the shallow feature set and the multi-dimensional feature data of catering network. The value range is [0,1]. The closer the value is to 1, the more specific the demand is for the network feature. The closer the value is to 0, the more ambiguous the orientation is.

[0097] Another core characteristic of users' vague dining needs is that they do not have a clear target for restaurant features. By calculating the target of restaurant features, the clarity of the needs can be quantitatively determined. This, along with the correlation with the user profile, forms a two-way basis for judgment, thus improving the dimensions for judging vague needs.

[0098] In the specific implementation process, the cosine similarity algorithm of Example 1 is first used to calculate the matching degree between the shallow feature set and the multi-dimensional feature data of a single catering outlet. t is the node number; then calculate the average matching degree of all participating nodes as the node feature orientation degree. The calculation formula is:

[0099]

[0100] Where m represents the total number of restaurants participating in the global matching. Let be the cosine similarity value between the t-th catering outlet and the shallow feature set;

[0101] For example, a global matching process is performed between the shallow feature set and the multi-dimensional feature data of 100 hot pot restaurants, and the cosine similarity value of each restaurant is calculated, with a mean of 0.28, generating the feature orientation of each restaurant. This indicates that user demand has an ambiguous direction regarding the characteristics of catering outlets.

[0102] Step 205: Based on the preset portrait relevance threshold and network feature orientation threshold, perform threshold comparison on the portrait relevance and network feature orientation respectively, output the fuzzy demand judgment result, and output the user basic consumption portrait and shallow feature set at the same time; among them, when the portrait relevance is lower than the portrait relevance threshold and the network feature orientation is lower than the network feature orientation threshold, it is judged as a fuzzy demand; the rest are judged as non-fuzzy demands.

[0103] It should be noted that the profile correlation threshold , Point feature orientation threshold The reference is to a preset threshold for determining users' fuzzy dining needs, generated through training on a large amount of historical user dining consumption data. This embodiment sets... , The fuzzy requirement determination result only includes fuzzy requirements and non-fuzzy requirements.

[0104] By comparing and judging with dual thresholds, ambiguous dining needs can be accurately defined, avoiding misjudgments caused by a single threshold, and ensuring that subsequent targeted analysis processes are only performed on genuine ambiguous dining needs.

[0105] In the specific implementation process, the preset [data] is retrieved. and Relationship between portraits and Comparison, point feature orientation and Comparison; if and If the condition is fuzzy, it is determined to be a fuzzy requirement; otherwise, it is determined to be a non-fuzzy requirement. The fuzzy requirement determination result is output, and the user's basic consumption profile and shallow feature set are output simultaneously for subsequent process calls.

[0106] In one possible implementation, such as Figure 3 As shown, when a user's query is determined to be ambiguous, semantic parsing and feature extraction are performed on the user's query information to generate user demand feature elements, including the following:

[0107] Step 206: Perform feature quantification and structuring processing on the basic user consumption profile to generate a standardized user consumption profile.

[0108] It should be noted that the standardized user consumption profile refers to a standardized profile formed by further refining and quantifying the basic user consumption profile, supplementing feature dimensions, and reconstructing the structure, which meets the matching requirements of the rule library for mapping the needs of catering consumption scenarios and can be used for cluster analysis.

[0109] In the specific implementation process, the characteristics of each dimension of the user consumption profile are refined and quantified, and feature dimensions that are strongly related to the user's dining experience (such as dining time preference, whether or not to queue, and preference for soup base type) are added; the quantitative deviation of the original feature dimensions is corrected; and the profile is restructured according to the preset catering consumption scenario feature specifications to generate a standardized user consumption profile.

[0110] Step 207: Perform lightweight semantic parsing on the shallow feature set, extract parsable scattered feature elements and fuzzy evaluation weak labels, and integrate them to generate a requirement calibration set.

[0111] It should be noted that analyzable scattered feature elements refer to the scattered features that can be clearly defined in the shallow feature set; fuzzy evaluation weak labels refer to the standardized labels transformed from fuzzy evaluation expressions in users' dining needs; and the demand calibration set refers to the set of bases for subsequent calibration of real needs, formed by integrating analyzable scattered feature elements and fuzzy evaluation weak labels.

[0112] Since users' vague dining needs are not entirely devoid of usable information, extracting the parsable parts and vague evaluation tags can serve as the core basis for subsequent calibration of real needs, improving the accuracy of reverse inference of real dining needs and avoiding the waste of demand information.

[0113] For example, the shallow feature set is "nearby, hot pot"; after secondary parsing, the parsable scattered feature elements are extracted as "hot pot", and the fuzzy evaluation weak labels are "dining nearby, good taste"; and the requirements calibration set is generated by integrating them.

[0114] Step 208: Perform feature matching between the standardized user consumption profile and the preset catering consumption scenario demand mapping rule library to generate a basic demand element set; at the same time, perform clustering processing on the standardized user consumption profile input network multi-dimensional feature data, filter to form candidate network clusters and extract common features, perform cross-validation and feature integration with the basic demand element set to generate the initial real demand element set of users.

[0115] It should be noted that the catering consumption scenario demand mapping rule base refers to a pre-set set of rules containing dining needs corresponding to different user consumption profiles, which is generated by training a large amount of historical user catering consumption data; the basic demand element set refers to the preliminary dining demand set derived by back-inferring through rule base matching; the candidate outlet cluster group refers to the set of catering outlets that are matched with user consumption profiles and selected by the K-means clustering algorithm; and the initial real user demand element set refers to the preliminary real dining demand set formed after back-inferring through the rule base and cross-validating with the common features of clustering.

[0116] In the specific implementation process, standardized consumer profiles are matched with a rule base for mapping dining consumption scenarios, and a set of basic demand elements is generated based on the matched rules. The K-means clustering algorithm is then used to perform clustering processing on the standardized user profiles input into the multi-dimensional feature data of the outlets. The core of the clustering is the calculation of Euclidean distance, with the formula as follows:

[0117]

[0118] in, Let p be the Euclidean distance between the p-th node and the cluster center. To quantify the j-th feature dimension of the standardized user consumption profile, Let be the quantized value of the j-th feature dimension of the p-th network point, and n be the number of feature dimensions.

[0119] Candidate network clusters with a feature matching degree reaching a preset threshold are selected based on Euclidean distance. Common features of these clusters are then extracted. Cross-validation is performed between these common features and the basic requirement element set to eliminate conflicting features, supplement missing features, and integrate them to generate the initial set of real user requirement elements.

[0120] For example, a standardized consumer profile is defined as "within 3 kilometers, Sichuan hot pot, average price 50-80 yuan, taste rating ≥ 4.5, dining hours 18:00-22:00". After matching with the catering consumption scenario demand mapping rule library, the basic demand element set is deduced as "within 3 kilometers, Sichuan hot pot, average price within 80 yuan, taste rating ≥ 4.5". The common characteristics of the candidate catering outlet clusters formed by K-means clustering are "within 3 kilometers, Sichuan hot pot, average price 50-80 yuan, taste rating ≥ 4.5, business hours ≥ 22:00, free parking available". After cross-validation, the demand for "business hours ≥ 22:00, free parking available" is added, and non-conflicting features are eliminated to generate the initial real demand element set of consumers.

[0121] Step 209: Perform fusion calibration on the demand calibration set and the user's initial real demand element set. Optimize the user's initial real demand element set through resolvable scattered feature elements. Associate and bind the weak labels of fuzzy evaluation with the objective quantitative features in the multi-dimensional feature data of the network points and supplement them to the calibrated element set. Label each feature element with hard demand attributes or soft demand attributes to generate standardized user demand feature elements.

[0122] It should be noted that fusion calibration refers to the process of integrating the demand calibration set with the user's initial real demand element set to optimize and improve the dining demand elements; attribute self-calibration refers to the process of verifying and correcting the hard and soft demand attribute labeling results through historical attribute labeling data; hard demand attributes refer to the necessary prerequisite characteristics for users to dine; soft demand attributes refer to non-essential characteristics that affect the dining experience.

[0123] In the specific implementation process, the demand calibration set and the user's initial real demand element set are fused and calibrated. The initial real demand element set is refined and optimized in the same dimension by resolvable scattered feature elements. Objective quantitative features corresponding to the weak labels of fuzzy evaluation are extracted from the multi-dimensional feature data of the network points, and the two are associated and bound and supplemented into the calibrated element set.

[0124] In one possible implementation, historical attribute annotation data is extracted from the catering consumption scenario demand mapping rule base, and feature matching verification is performed between the initial attribute annotation results of the current feature elements and the historical annotation data; if there is an attribute matching conflict, the initial attribute annotation results are corrected based on the scenario features of the standardized user consumption profile; and the hard demand attributes and soft demand attributes of each feature element are annotated.

[0125] For example, the resolvable fragmented feature element of the demand calibration set is "hot pot". The "hot pot" in the initial demand element set is refined into "Sichuan-style hot pot (beef tallow broth)". The fuzzy evaluation weak label is "dining nearby, good taste". The "within 3 kilometers, taste rating ≥ 4.5" feature in the multi-dimensional feature data of the network points is associated and supplemented. The initial label "supports free parking" is a hard demand, which conflicts with the historical label data. It is corrected to a soft demand based on the consumer's dining preferences. Finally, a standardized user demand feature element containing 5 feature elements and corresponding hard and soft attributes is generated: within 3 kilometers (hard), Sichuan-style hot pot (beef tallow broth) (hard), average price per person within 80 yuan (hard), taste rating ≥ 4.5 (hard), support for free parking (soft).

[0126] This embodiment uses a dual-dimensional quantitative judgment based on profile correlation and network feature orientation to accurately distinguish between fuzzy and non-fuzzy dining needs, avoiding misjudgments and omissions of fuzzy needs. For fuzzy needs, it combines merchant business profiles, a catering consumption scenario demand mapping rule library, and K-means clustering analysis. Through a process of bidirectional back-inference, cross-validation, and fusion calibration, it transforms fuzzy natural language needs into precise and standardized user demand feature elements. At the same time, it optimizes the annotation results of hard and soft demand attributes through attribute self-calibration, solving the technical problem that fuzzy dining needs are difficult to accurately transform into standardized feature elements, further improving the recommendation accuracy under different dining demand scenarios.

[0127] Example 3

[0128] Considering that in practical applications, the above embodiments use an equal weighting approach to process the various feature dimensions of the multi-dimensional feature data of the network points, without taking into account the actual contribution differences of each feature dimension under different user dining consumption scenarios, it is easy for irrelevant feature dimensions with low contribution to interfere with the matching results, resulting in distortion of the network point feature orientation P, and thus increasing the misjudgment rate of the user's ambiguous dining needs. In order to solve this technical problem, in this embodiment, the step of performing a full-domain feature matching calculation between the shallow feature set and the multi-dimensional feature data of the network points to generate the network point feature orientation includes the following:

[0129] A scenario feature contribution database is constructed, which categorizes the contribution coefficients of each feature dimension according to the user's dining consumption scenario. Business scenario features are extracted from the user's basic consumption profile. Based on these features, the contribution coefficients corresponding to each feature dimension of the network's multi-dimensional feature data are matched from the scenario feature contribution database. Weighting is then applied to each feature dimension of the network's multi-dimensional feature data. Finally, a global feature matching calculation is performed between the shallow feature set and the weighted network multi-dimensional feature data to generate network feature orientation.

[0130] It should be noted that the scenario feature contribution library refers to a rule library that is divided according to the user's catering consumption scenario and contains the contribution coefficient of each catering outlet's feature dimension.

[0131] Contribution coefficient This refers to the weight value assigned to the importance of each network feature dimension based on the catering consumption scenario. The value range is [0,1]. The sum of the contribution coefficients of each dimension is 1. The higher the value, the stronger the importance of that feature dimension to the catering consumption scenario.

[0132] The importance of various feature dimensions of a restaurant varies significantly across different consumer dining scenarios. For example, in the scenario of a consumer dining at a Sichuan hot pot restaurant, the importance of "category of business, taste rating, and average spending per person" is far higher than irrelevant features such as "store decoration rating and number of private rooms". By weighting the contribution of a scenario, the core consumption feature dimensions can be highlighted and the influence of irrelevant feature dimensions can be weakened, avoiding the distortion of store feature orientation caused by indiscriminate matching and improving the accuracy of matching calculation.

[0133] For example, the database structure is divided according to catering consumption scenarios such as Sichuan hot pot, clear broth hot pot, fast food, and Chinese food, with each scenario corresponding to a set of contribution coefficients for catering store characteristic dimensions. The coefficients are generated by training with a large amount of historical data from catering scenarios, and the sum of the coefficients of each dimension is 1.

[0134] We extract users' dining consumption scenario characteristics from their basic consumption profiles, such as Sichuan hot pot consumption. Based on these characteristics, we match the contribution coefficients of each feature dimension of the network's multi-dimensional feature data from the scenario feature contribution database. .

[0135] Quantify the characteristic dimensions of each branch. With corresponding contribution coefficient Perform product calculation to complete the weighted processing of each feature dimension of the multi-dimensional feature data of the network points. The calculation formula is as follows:

[0136]

[0137] in, Let be the weighted quantization value of the j-th feature dimension of the i-th network point. This represents the original quantized value of the j-th feature dimension of the i-th point. is the contribution coefficient of the j-th feature dimension.

[0138] Using the cosine similarity algorithm and mean formula from Example 2, the shallow feature set is combined with the weighted multi-dimensional feature data of the dots. Perform global feature matching calculation to generate point feature orientation. .

[0139] This embodiment significantly weakens the interference of irrelevant feature dimensions on matching results and strengthens the influence of core feature dimensions by using scenario-based differentiated contribution weighting. This improves the accuracy of full-domain feature matching calculation, making the network feature orientation P more in line with the actual needs of catering business scenarios, thereby reducing the misjudgment rate of fuzzy dining needs judgment of catering network locations; the weighting logic is deeply bound to the user's catering consumption scenario.

[0140] Example 4

[0141] Considering that in practical applications, the above embodiments, which perform single forward clustering on catering outlets, are prone to feature bias due to the personalized characteristics of user consumption profiles, resulting in the common characteristics of candidate outlet clusters not matching the actual dining needs of similar users and affecting the accuracy of the back-inference of the basic demand element set, this embodiment inputs standardized user consumption profiles into the multi-dimensional feature data of the outlets for clustering processing, and filters to form candidate outlet clusters including the following:

[0142] Based on standardized user consumption profiles, forward clustering is performed on the outlets in the multi-dimensional data to generate a first candidate outlet cluster. Similarity calculation is then performed on historical merchant data based on the standardized user consumption profiles to generate a set of similar merchants. Based on the characteristics of cooperating outlets within this set of similar merchants, reverse clustering is performed on the outlets in the multi-dimensional data to generate a second candidate outlet cluster. The intersection of common features between the first and second candidate outlet clusters is extracted, and a final candidate outlet cluster is formed based on this intersection.

[0143] It should be noted that forward clustering refers to the process of directly performing K-means clustering on catering outlets based on the standardized user consumption profile of the current user; backward clustering refers to the process of performing K-means clustering on catering outlets based on the dining characteristics of similar consumer groups; similar consumer groups refer to the set of historical dining users whose similarity to the standardized user consumption profile of the current user reaches a preset threshold; common feature intersection refers to the overlapping part of the common features of the candidate outlet clusters generated by forward clustering and backward clustering.

[0144] In the specific implementation process, forward clustering adopts the K-means clustering algorithm and Euclidean distance formula of Example 2. Based on the standardized user consumption profile, forward clustering is performed on the catering outlets in the multi-dimensional data of the outlets to generate the first candidate outlet cluster group, and the common features of the cluster group are extracted. It covers dimensions that are strongly related to consumers' personalized needs, such as location, product category, taste, and average per capita spending.

[0145] Reverse clustering retrieves standardized consumer profiles from the platform's full consumer base and uses a cosine similarity algorithm to calculate the similarity between each consumer and the current consumer profile. Consumers with a similarity ≥ 0.8 are selected to form similar consumer groups. The characteristics of the restaurants frequented by these similar consumer groups are extracted, and based on these characteristics, K-means clustering is performed on the restaurant locations in the multi-dimensional data of the network, while maintaining the Euclidean distance formula. This generates a second candidate cluster of locations, and the common features of this cluster are extracted. ;

[0146] Cross-validation screening extracts common features from the clusters of the first candidate nodes. Common characteristics of clusters with the second candidate network intersection Based on the intersection of these common features, network points are selected to form the final candidate network point cluster.

[0147] This embodiment effectively solves the problem of candidate site cluster feature bias caused by single clustering by using forward and reverse dual clustering cross-validation, improves the reliability and accuracy of common features of candidate site clusters, and optimizes the back-inference accuracy of basic demand element set, making the analysis results of fuzzy dining demand more in line with the actual cooperation needs of business scenarios.

[0148] Example 5

[0149] Considering that the weight allocation in the above embodiments does not take into account the differences between hard and soft attributes of dining demand feature elements in practical applications, and only allocates weights according to the user's preset priority, it is easy for the weight of hard demand attribute feature elements to be diluted by soft demand attribute feature elements, affecting the accuracy of similarity calculation and matching results. In this embodiment, the weight allocation processing for each user demand feature element based on the demand priority of user demand feature elements includes the following:

[0150] Feature elements labeled as hard demand attributes are divided into a first weight level, and feature elements labeled as soft demand attributes are divided into a second weight level. Based on the preset weight ratio of each level, the feature elements in the first and second weight levels are assigned weights within each level to generate a weighted user demand feature vector. The overall weight ratio of the first weight level is higher than that of the second weight level.

[0151] It should be noted that the first weight level refers to the weight level of hard demand attribute features, which is the necessary prerequisite level for dining; the second weight level refers to the weight level of soft demand attribute features, which is the level for enhancing the dining experience; the overall weight ratio of the level refers to the preset total weight ratio of each weight level, with a value range of (0,1), and the sum of the two levels is 1.

[0152] Since essential attributes are necessary prerequisites for users to dine, if their weight is diluted by soft attributes, it will lead to the similarity calculation overemphasizing experience features and ignoring the necessary prerequisites for cooperation. By dividing the attributes into levels, the overall weight of essential attributes can be strengthened, ensuring the core priority of the necessary prerequisites, making the weight allocation more in line with the actual logic of dining, and improving the targeting of similarity calculation.

[0153] For example, the standardized consumer demand characteristics are: within 3 kilometers, Sichuan-style hot pot, average cost per person under 80 yuan, taste rating of 24.5, free parking, and availability of desserts. Among these, the hard demand attributes are: within 3 kilometers, Sichuan-style hot pot, and average cost per person under 80 yuan; the soft demand attribute is: taste rating. Free parking and desserts are available.

[0154] The first weighting tier is tied to locations within 3 kilometers, Sichuan-style hot pot, and an average cost of 80 yuan per person; the second weighting tier is tied to taste ratings. Free parking and desserts are among the features; consumers prioritize Sichuan hot pot the most (0.35), followed by areas within 3 kilometers (0.25), and then areas within 80 yuan per person (0.1), making it the top priority. =0.35 + 0.25 + 0.1 = 0.7. Consumers will rate the taste. Set to 0.15), free parking to 0.1, dessert availability to 0.05, second weight level. =0.15+0.1+0.05=0.3. The feature vectors are: Sichuan hot pot (0.35), within 3 kilometers (0.25), average cost per person within 80 yuan (0.1), and taste rating. (0.15), free parking (0.1), dessert available (0.05), the total weight is 1.

[0155] This embodiment effectively solves the problem of the dilution of the weight of hard dining needs by soft needs through attribute-level driven weight allocation, making the weight allocation more in line with the actual underlying logic of dining and improving the targeting and accuracy of similarity calculation.

[0156] Example 6

[0157] Considering that the above embodiments use a mixed calculation method of hard and soft demand feature elements in the actual application process, which incorporates the feature elements of hard and soft demand attributes into the same formula, it is easy to cause the problem of "hard demand not meeting the standard but soft demand high score" leading to an inflated overall similarity. For example, a store is a clear soup hot pot (hard demand not meeting the standard), but because of its low average price, free parking, and high taste rating, the similarity after mixed calculation reaches 0.75, and it is mistakenly included in the candidate store set. In this embodiment, the similarity calculation is performed on the weighted user demand feature vector and the multi-dimensional feature data of the network point, and the candidate network point set is generated based on the similarity calculation results, including the following:

[0158] User demand features are divided into hard demand attributes and soft demand attributes. Precise matching calculation is performed on the feature elements corresponding to hard demand attributes to generate the first similarity. Fuzzy matching calculation is performed on the feature elements corresponding to soft demand attributes to generate the second similarity. The first similarity and the second similarity are weighted and superimposed based on a preset ratio to generate a comprehensive similarity. Candidate site sets are generated based on the comprehensive similarity.

[0159] It should be noted that the first similarity refers to the exact matching calculation result of hard requirement attribute features, using the Jaccard similarity coefficient algorithm, with a value ranging from [0,1], where a value of 1 indicates a perfect match of hard requirements. The second similarity refers to the fuzzy matching calculation result of soft requirement attribute features, using the cosine similarity algorithm, with a value ranging from [0,1], where a higher value indicates a stronger fit of soft requirements. Overall similarity... It is a weighted sum of two levels of similarity, with a value range of [0,1]. This result is a positive adaptation result.

[0160] In practice, the first similarity calculation refers to performing an exact match calculation on the feature elements corresponding to the hard requirement attributes, using the Jaccard similarity coefficient algorithm. The calculation formula is:

[0161]

[0162] Where A is the set of user hard demand features, B is the set of features corresponding to the hard demand of catering outlets, and J is the first similarity.

[0163] The second similarity calculation involves performing fuzzy matching on the feature elements corresponding to the soft demand attributes, using the cosine similarity algorithm from Example 1 to generate the second similarity. .

[0164] Retrieve preset proportions, first similarity proportions The second similarity percentage The sum of the two levels of similarity is 1; the first similarity and the second similarity are weighted and summed according to their respective proportions to generate a comprehensive similarity. The calculation formula is:

[0165]

[0166] Set a comprehensive similarity threshold, filter out network points with a comprehensive similarity higher than the threshold, integrate them into a candidate network point set, and retain the comprehensive similarity of this layer. This serves as a positive adaptation result for subsequent supply and demand matching models.

[0167] This embodiment effectively solves the problem of inflated overall similarity caused by the mixed calculation of hard and soft requirements through hierarchical and progressive similarity calculation. It strengthens the veto power of hard dining requirements, improves the screening accuracy of the candidate network set, and provides higher-quality alternatives that are more in line with the core needs of users for subsequent supply and demand matching verification. At the same time, the optimized hierarchical overall similarity is still directly used as a positive matching result.

[0168] Example 7

[0169] Considering that the above embodiments, in practical applications, only use similarity or comprehensive similarity as a single ranking dimension, without taking into account the actual reception capacity of restaurants and the adaptability to consumers' dining needs, it is easy to produce invalid recommendations with high similarity but no actual dining available. For example, a Sichuan hot pot restaurant may have a layered comprehensive similarity of 0.95, but the tables are full and the waiting time exceeds 2 hours, failing to meet consumers' immediate dining needs. In this embodiment, the comprehensive ranking process performed on the adapted network set to generate standardized network recommendation results includes the following:

[0170] The supply and demand saturation features of each location in the matching location set are extracted from the multi-dimensional data of the locations. The supply and demand saturation features are used as the first ranking dimension to perform an initial ranking on the matching location set. The comprehensive similarity is used as the second ranking dimension to perform a second ranking on the matching location set after the initial ranking, generating standardized location recommendation results.

[0171] It should be noted that supply and demand saturation This refers to the ratio of the number of current partner merchants to the maximum dining capacity of a catering outlet. The value ranges from [0,1]. The closer the value is to 0, the more sufficient the partner capacity of the outlet is. The closer the value is to 1, the more saturated the outlet is. The first sorting dimension refers to the highest priority sorting criterion, which is used to filter outlets with sufficient capacity. The second sorting dimension refers to the secondary sorting criterion, which is used to filter outlets with high suitability.

[0172] The supply and demand saturation of catering outlets directly determines the likelihood of dining. If an outlet has no capacity, even with a high overall similarity score, actual dining cannot be arranged, making it an invalid recommendation. Using supply and demand saturation as the primary ranking dimension can fundamentally avoid invalid recommendations and improve the practicality of the recommendation results. At the same time, combining it with a secondary ranking based on overall similarity ensures the suitability of the recommended outlets.

[0173] In the specific implementation process, the current actual number of customers received by each catering outlet within the appropriate outlet set is extracted from multi-dimensional data of the outlets. Maximum dining capacity The formula for calculating supply and demand saturation is as follows:

[0174]

[0175] in, Let i be the supply and demand saturation level of the i-th network point. Let i be the actual number of customers received at the i-th branch. This represents the maximum dining capacity of the i-th location.

[0176] Supply and demand saturation As the first ranking dimension, the adapted network point set is initially sorted in ascending order of supply and demand saturation, with lower supply and demand saturation ranking higher. The overall similarity is then considered. or As a second sorting dimension, the adapted point set after the initial sorting is sorted again in descending order of comprehensive similarity.

[0177] By integrating the dual-dimensional ranking results with the core operational information of the outlets, standardized outlet recommendation results are generated, which include ranking position, supply and demand saturation, comprehensive similarity, and core operational information, and then output to the user.

[0178] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A recommendation method based on multi-dimensional data of service outlets, characterized in that, Includes the following steps: Step 100: Collect basic data of the outlets, perform multi-dimensional feature mining and quantification on the basic data of the outlets, and generate multi-dimensional feature data of the outlets. Step 200: Receive user query information, perform semantic parsing and feature extraction on the user query information, and generate user query feature elements; Step 300: Based on the demand priority of user demand feature elements, perform weight allocation processing on each user demand feature element to generate a weighted user demand feature vector. Step 400: Perform similarity calculation between the weighted user demand feature vector and the multi-dimensional feature data of the network points, and generate a candidate network point set based on the similarity calculation results; Step 500: Construct a supply and demand matching model, input the candidate site set into the supply and demand matching model to perform supply and demand adaptation verification, remove sites that fail the supply and demand adaptation verification, and generate a set of adapted sites; Step 600: Perform comprehensive sorting on the adapted site set, generate standardized site recommendation results, and complete the result output; Before performing semantic parsing and feature extraction on the user's query information, the following content is also included: Retrieve users' historical dining data, perform structured and feature quantification processing on the historical dining data, and generate basic user consumption profiles; Perform shallow semantic feature extraction on the user's query information to generate a shallow feature set; The shallow feature set is compared with the user's basic consumption profile to calculate the feature matching degree and generate the profile correlation degree. Perform global feature matching calculations on the shallow feature set and the multi-dimensional feature data of the network points to generate network point feature orientation. Based on preset thresholds for profile relevance and network feature orientation, threshold comparisons are performed on profile relevance and network feature orientation respectively, and the fuzzy demand judgment result is output. At the same time, the user's basic consumption profile and shallow feature set are also output. Among them, when the profile relevance is lower than the profile relevance threshold and the network feature orientation is lower than the network feature orientation threshold, it is judged as a fuzzy demand; the rest are judged as non-fuzzy demands.

2. The recommendation method based on multi-dimensional data of network points according to claim 1, characterized in that, The step of performing global feature matching calculation between the shallow feature set and the multi-dimensional feature data of the network points to generate network point feature orientation includes the following: Construct a scenario feature contribution database, which is divided into contribution coefficients for each feature dimension according to the user's dining consumption scenario; Extract business scenario features from the user's basic consumption profile, match the contribution coefficients of each feature dimension of the multi-dimensional feature data of the outlets from the scenario feature contribution database based on the business scenario features, and perform weighted processing on each feature dimension of the multi-dimensional feature data of the outlets. Perform global feature matching calculations on the shallow feature set and the weighted multi-dimensional feature data of the network points to generate network point feature orientation.

3. The recommendation method based on multi-dimensional data of network points according to claim 2, characterized in that, When the user's query information is determined to be ambiguous, semantic parsing and feature extraction are performed on the user's query information to generate user demand feature elements, including the following: Perform feature quantification and structuring processing on the basic user consumption profile to generate a standardized user consumption profile; Lightweight semantic parsing is performed on the shallow feature set to extract parsable scattered feature elements and weak labels of fuzzy evaluation, and then integrated to generate a demand calibration set. The standardized user consumption profile is matched with the pre-defined catering consumption scenario demand mapping rule library to generate a basic demand element set; at the same time, the standardized user consumption profile is input into the multi-dimensional feature data of the outlets to perform clustering processing, and candidate outlet clusters are selected and common features are extracted. The common features are cross-validated and integrated with the basic demand element set to generate the initial real demand element set of users. The demand calibration set is fused and calibrated with the user's initial real demand element set. The user's initial real demand element set is optimized by resolvable scattered feature elements. The weak labels of fuzzy evaluation are associated and bound with the objective quantitative features in the multi-dimensional feature data of the network points and supplemented to the calibrated element set. Hard demand attributes or soft demand attributes are labeled for each feature element to generate standardized user demand feature elements.

4. The recommendation method based on multi-dimensional network data according to claim 3, characterized in that, Standardized user consumption profiles are input into the multi-dimensional feature data of service outlets, and clustering is performed to filter and form candidate outlet clusters, including the following: Based on standardized user consumption profiles, positive clustering is performed on the outlets in the multi-dimensional data to generate the first candidate outlet cluster group; Based on standardized user consumption profiles, similarity calculations are performed on historical merchant data to filter and generate a set of similar merchants. Based on the characteristics of cooperative outlets in the set of similar merchants, reverse clustering is performed on outlets in multi-dimensional data to generate a second candidate outlet cluster. Extract the common feature intersection between the first candidate network cluster and the second candidate network cluster, and select candidate network clusters based on the common feature intersection.

5. The recommendation method based on multi-dimensional data of network points according to claim 3, characterized in that, The labeling of each feature element with hard or soft requirement attributes includes the following: Historical attribute-labeled data is extracted from the demand mapping rule base for catering consumption scenarios. Perform feature matching verification between the initial attribute labeling results of the current feature elements and the historical attribute labeling data. If there is an attribute matching conflict, correct the initial attribute labeling results based on the scenario features of the standardized user consumption profile, and complete the labeling of the hard requirement attributes and soft requirement attributes of each feature element.

6. The recommendation method based on multi-dimensional data of network points according to claim 5, characterized in that, The weight allocation process for each user demand feature element based on the demand priority includes the following: Feature elements labeled as hard demand attributes are divided into a first weight level, and feature elements labeled as soft demand attributes are divided into a second weight level. Based on the preset weight ratio of each level, the feature elements in the first and second weight levels are assigned weights within each level to generate a weighted user demand feature vector. The overall weight ratio of the first weight level is higher than that of the second weight level.

7. The recommendation method based on multi-dimensional data of network points according to claim 6, characterized in that, The step involves performing similarity calculations between weighted user demand feature vectors and multi-dimensional feature data of service points, and then generating a candidate service point set based on the similarity calculation results. This includes the following: The user demand feature elements are divided into hard demand attributes and soft demand attributes. The feature elements corresponding to the hard demand attributes are subjected to precise matching calculation to generate the first similarity. Perform fuzzy matching calculations on the feature elements corresponding to soft demand attributes to generate a second similarity; Based on a preset ratio, the first similarity and the second similarity are weighted and superimposed to generate a comprehensive similarity. Candidate network point sets are then selected based on the comprehensive similarity.

8. The recommendation method based on multi-dimensional data of network points according to claim 1, characterized in that, The process of constructing a supply-demand matching model involves inputting the candidate site set into the model to perform a supply-demand fit verification, removing sites that fail the verification, and generating a set of adapted sites, including the following: The operational capability features of candidate outlets are extracted from multi-dimensional feature data of outlets. The operational capability features are then matched with user demand features to generate a demand feasibility verification result. The results of the demand feasibility verification and the positive adaptation results of the user demand feature elements corresponding to the candidate sites are used together as the basis for determining the supply and demand adaptation verification. The supply and demand adaptation verification is performed and an adapted site set is generated.

9. The recommendation method based on multi-dimensional data of network points according to claim 7, characterized in that, The comprehensive ranking process performed on the adapted site set to generate standardized site recommendation results includes the following: The supply and demand saturation features of each point in the matching point set are extracted from the multi-dimensional data of the points. The supply and demand saturation features are used as the first ranking dimension to perform an initial ranking on the matching point set. The comprehensive similarity is used as the second ranking dimension to perform a second ranking on the matching point set after the initial ranking, generating standardized point recommendation results.

Images