Information ranking method and device based on multi-dimensional data weighting and personalized model, and medium

Abstract

The application discloses a method and device for information sorting based on multi-dimensional data weighting and personalized model and a medium, belongs to the technical field of internet data processing, and comprises candidate information screening, basic comprehensive scoring, personalized preference modeling, recommended merchant screening, mixed insertion list generation and several steps. In the application, a service range is first determined based on the user position and the current time, and effective candidate merchants are screened out. After authorized operation data processing, a basic comprehensive score is calculated. A multi-dimensional preference model is constructed based on the user's full historical effective behavior, and the personalized preference score of the user for each merchant is obtained. After an initial sorting list is generated according to the basic score, high-matching-degree merchants are inserted into preset fixed positions to generate a final list. The application is matched with a new merchant supporting and user cold start adjustment mechanism, all rules are unified and disclosed, the user decision efficiency and consumption experience are improved, and the platform overall traffic distribution efficiency and conversion level are also improved.

Description

Technical Field

[0001] This invention relates to the field of Internet data processing technology, and in particular to an information sorting method, apparatus and medium based on multidimensional data weighting and personalized models. Background Technology

[0002] With the rapid development of local life service platforms, the ranking technology of merchant information for C-end users has become a core technology module affecting the efficiency of platform traffic distribution, user experience, and ecosystem health. Early ranking technologies were mostly based on static attribute weighting, usually using linear weighting based on dimensions such as geographical distance, merchant sales, and ratings. With the iteration of personalized recommendation technology, collaborative filtering and content matching algorithms based on user behavior profiles have been gradually introduced into the ranking process. By mining users' historical behavioral preferences, differentiated ranking results can be achieved, which has improved the matching degree of recommendation results to a certain extent and has become the mainstream technical path in the current industry.

[0003] However, current mainstream ranking schemes still have room for optimization in multiple dimensions. On the one hand, most schemes have a balance between fairness and adaptability. Some schemes that overemphasize personalization use dynamic weight adjustment mechanisms, and there are implicit differences in the scoring calculation rules for different merchants, making it difficult to guarantee the openness and fairness of the basic ranking. On the other hand, schemes that emphasize rule-based systems suffer from insufficient accuracy in personalized matching. On the other hand, existing schemes do not adequately support the balance of the platform ecosystem. Traffic resources tend to concentrate on top merchants, and the cold start cost for new merchants is high. In addition, for new users with insufficient historical behavioral data, most schemes have large biases in their preference fitting, and the traceability of the entire ranking process is weak, making it difficult to adapt to the flexible adjustment needs of different business scenarios. Summary of the Invention

[0004] To address the shortcomings mentioned in the background technology, we propose an information ranking method, apparatus, and medium based on multidimensional data weighting and personalized models.

[0005] The technical solution mainly consists of an information ranking method based on multidimensional data weighting and personalized models, including the following steps:

[0006] Obtain the target user's current geographic location coordinates and current system time, and determine the candidate information set within the service area based on the current geographic location coordinates. The candidate information set consists of all valid merchants that meet the platform's service access rules.

[0007] The system retrieves operational data for each merchant in the candidate information set. The collection of operational data is authorized by both the user and the merchant. The operational data includes distance data, which is the straight-line distance between the merchant's business address and the target user's current geographical location.

[0008] The operational data is normalized, and the basic comprehensive score of each merchant is calculated based on the preset weights set according to the core priority of the business scenario. The historical behavior data of the target users is obtained, which is all the valid redemption behavior data generated by the users within the platform.

[0009] Construct a multi-dimensional personalized preference model, calculate the personalized preference score of the target user for each merchant based on the multi-dimensional personalized preference model, generate an initial sorted list based on the basic comprehensive score, and filter and recommend merchants based on the personalized preference score;

[0010] The recommended merchants are inserted into the predetermined positions in the initial sorting list according to the preset hybrid insertion strategy, generating the final display list.

[0011] Preferably, the steps of normalizing operational data and calculating a basic comprehensive score for each merchant based on preset weights include:

[0012] For positive indicators, forward normalization is used, and for negative indicators, reverse normalization is used. The normalized scores of each dimension are multiplied by their corresponding preset weights and then summed to obtain the basic comprehensive score for each merchant. The higher the value of a positive indicator, the higher the normalized score; the higher the value of a negative indicator, the lower the normalized score. The sum of the preset weights for each dimension is one. The weight allocation is determined based on the core priority of the business scenario to ensure that the contribution of core indicators to the basic comprehensive score meets business expectations. The entire score calculation process uses standardized calculation logic, and the calculation rules for all candidate merchants are completely consistent to ensure the fairness and comparability of the basic comprehensive scores.

[0013] Preferably, the step of constructing a multi-dimensional personalized preference model includes:

[0014] Calculate the weights for category preference, content keyword preference, price range preference, and consumption time period preference respectively, and allocate the weights of each dimension according to a preset fixed priority. The sum of the weights of each dimension is one.

[0015] Category preference dimension corresponds to the highest priority weight;

[0016] The priority weight of the content keyword preference dimension is second only to the category preference dimension.

[0017] The price segment preference dimension has the second highest priority weight, only after the content keyword preference dimension.

[0018] The lowest priority weight corresponds to the consumption time period preference dimension;

[0019] Once the weights for each dimension are calculated, they are fixed parameters of the preference model. The priority level of the weights remains fixed and the priority order is not adjusted as the scenario changes.

[0020] The weight of the category preference dimension is determined by the proportion of the number of redemptions for the target category to the total number of redemptions;

[0021] The weight of the content keyword preference dimension is determined by the frequency of the content keywords in the total orders in the historical orders;

[0022] The weight of the price segment preference dimension is determined based on the proportion of consumption in the corresponding price range in the user's historical behavior;

[0023] The weight of the consumption time period preference dimension is determined based on the proportion of consumption during the corresponding time period in the user's historical behavior;

[0024] The weights for all dimensions are calculated based on the user's full historical behavioral data, ensuring that the preference model can fully reflect the user's true needs and tendencies.

[0025] Preferably, the matching logic for the content keyword preference dimension includes:

[0026] When the content name provided by the merchant is exactly the same as the content name in the user's historical behavior, it is judged as an exact match and assigned a first match score. When the content name provided by the merchant is not exactly the same as the content name in the user's historical behavior, the core nouns of the user's historical content name are extracted and matched with the content name provided by the merchant. A second match score is assigned based on the inclusion of the core nouns.

[0027] The first matching score is higher than the second matching score. The calculation rules for the matching score are consistent for all merchants to ensure the objectivity and reproducibility of the matching results. The extraction of core terms is completed based on standardized text processing rules. All merchants' content names adopt a unified text preprocessing process, including special character removal, unified character encoding conversion, and redundant content filtering steps. The preprocessed text is then used for keyword matching calculation to ensure the accuracy and consistency of the matching results.

[0028] Preferably, the method also includes traffic support strategies for newly joined merchants:

[0029] Obtain the total number of activity shares for the first time launched by a newly joined merchant. Set a threshold for the number of activity shares based on the platform's historical data and business strategy. When the total number of activity shares is greater than or equal to the threshold, the merchant will be displayed at the top within a preset distance range for a preset time period. When the total number of activity shares is less than the threshold, a preset weighting coefficient will be added to the merchant's basic comprehensive score within a preset period.

[0030] The traffic support strategy only applies to new merchants whose onboarding time has not exceeded the preset onboarding time threshold. Merchants that have passed the new merchant certification period are not eligible for the support strategy. The applicable conditions for the support strategy are the same for all new merchants. The onboarding time of a new merchant is calculated from the time when the merchant completes the platform onboarding review. The total number of activities is calculated based on the total number of redeemable activities for all valid activities published by the merchant after onboarding. The weighting coefficient is a fixed value and the same weighting coefficient is used for all eligible new merchants. The scope and duration of the featured display are based on a unified standard and are not differentiated for specific merchants.

[0031] Preferably, the step of inserting recommended merchants into predetermined positions in the initial sorting list according to a preset hybrid insertion strategy includes:

[0032] The final display list is divided into a first screen area, a second screen area, and subsequent areas. Multiple fixed insertion positions are set in the first screen area. Top merchants are selected from the merchants sorted by personalized preference scores to fill the insertion positions. When the original ranking of the selected merchant in the initial sorting list is better than the insertion position, its original ranking is retained and it is not inserted again. The number and position of the insertion positions are fixed and will not be adjusted with changes in merchants or users.

[0033] The number of top merchants selected is consistent with the number of insertion positions. The insertion process does not change the original order of merchants in the initial sort list that are not in the insertion positions. The personalized preference scores are sorted in descending order. The top merchants are the merchants whose ranking corresponds to the number of insertion positions in the sorting results. After the insertion operation is completed, the uniqueness of all merchants in the final list is checked. The list after the check is completed is the final sorted list that can be displayed.

[0034] Preferably, the method further includes user cold start processing, specifically:

[0035] The system detects the number of orders corresponding to the target user's historical behavior data. When the number of orders is zero, the personalized preference score calculation step is skipped, and the final display list is generated based solely on the basic comprehensive score.

[0036] When the number of orders is in the preset low-frequency range, the calculated personalized preference score is multiplied by the preset reduction coefficient. The reduction coefficient is a fixed value and the same reduction coefficient is used for all users in the low-frequency range. The triggering condition of the cold start processing logic is consistent for all users.

[0037] The statistical scope of historical behavior data includes all valid redeemed orders generated by users within the platform, excluding unredeemed canceled orders and abnormal orders. The value range of the low-frequency range is fixed, and the personalized preference score after reduction still participates in the subsequent recommended merchant selection and insertion process according to the original logic.

[0038] Preferably, the operational data also includes rebate amount and rebate difference. The rebate amount is the amount that can be returned to the user by the merchant for the activity. The rebate difference is the difference between the activity rebate amount marked by the merchant and the actual amount that can be returned. The rebate amount, rebate difference and distance data are all publicly queryable data of the merchant. The data collection process directly connects to the merchant's public data port. All merchants' operational data adopt a unified collection cycle and update frequency.

[0039] The present invention also provides an information sorting device based on a multidimensional data weighting and personalization model, including a data acquisition module, a comprehensive scoring module, a preference analysis module, and a list generation module;

[0040] The data acquisition module is configured to obtain user location, time information and merchant operation data. The data acquisition process uses standardized interface connection, and all data is checked for integrity and filtered for outliers before entering the subsequent processing flow.

[0041] The comprehensive scoring module is configured to calculate the merchant's basic comprehensive score based on a normalization algorithm.

[0042] The preference analysis module is configured to build a multi-dimensional preference model based on the user's historical behavior and calculate a personalized score. The model parameters are generated based on the user's full historical data.

[0043] The list generation module is configured to combine new merchant support strategies with a hybrid insertion strategy to generate the final sorted list. The entire list generation process is traceable, and all operations are logged to ensure the auditability of the sorting results.

[0044] 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.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] This invention employs a layered fusion mechanism that combines objective operational attribute quantification with multi-dimensional user preference modeling. While ensuring the openness and fairness of the basic ranking rules, it significantly improves the matching degree between recommended results and users' actual needs. The scoring calculation rules for all merchants are unified, avoiding interference from human factors in the ranking results. A fixed-priority preference model built based on users' full historical behavior can accurately depict users' multi-level consumption tendencies. Combined with a cold-start adaptive adjustment strategy, it covers demand adaptation scenarios for users with different behavioral volumes, effectively reducing users' information filtering costs, improving user decision-making efficiency and consumption experience, and also driving the overall traffic distribution efficiency and conversion level of the platform.

[0047] In this invention, a standardized differentiated support mechanism for new merchants provides them with a fair channel for traffic exposure, avoiding the problem of excessive concentration of traffic resources in the top merchants and ensuring a healthy competitive environment for both new and old merchants on the platform. The overall design adopts a modular decoupled architecture, with each module's calculation rules being independently adjustable. This not only ensures the stability and consistency of the ranking results but also provides flexible scenario adaptation capabilities to meet the needs of different business scenarios. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the overall operation of the present invention.

[0049] Figure 2 This is a flowchart illustrating the construction of the multi-dimensional personalized preference model in this invention;

[0050] Figure 3 This is a flowchart illustrating the execution process of the new merchant support strategy of this invention. Detailed Implementation

[0051] The present invention will now be described in further detail with reference to the accompanying drawings and preferred embodiments.

[0052] Example 1, refer to Figure 1-3 As shown, the information ranking method based on multidimensional data weighting and personalized models includes several steps: candidate information screening, basic comprehensive scoring, personalized preference modeling, recommended merchant screening, and hybrid insertion to generate a list.

[0053] Specifically, firstly, based on the obtained user location and current time, the service area is defined and all valid candidate merchants that meet the admission rules are selected. Then, the authorized operation data of all candidate merchants are collected, and after normalization processing, the basic comprehensive score of each candidate merchant is calculated according to the preset weight. At the same time, based on the user's full historical valid behavior data, a user personalized preference model is constructed from multiple fixed priority dimensions to calculate the user's personalized preference score for each candidate merchant. Then, an initial ordered list is generated according to the basic comprehensive score. Based on the personalized preference score, recommended merchants with high matching degree are selected. Finally, the recommended merchants are inserted into the fixed positions of the initial list according to the preset mixed insertion rules to generate the final display list.

[0054] Furthermore, this embodiment also includes traffic support for newly joined merchants and user cold start processing steps. Corresponding traffic support strategies are matched for eligible newly joined merchants, and corresponding preference score adjustment strategies are matched for users with different behavior levels. All calculation logic adopts a unified standard for all merchants and users, and there are no differentiated judgment rules.

[0055] This embodiment constructs a ranking system that balances fairness and adaptability by integrating basic objective scoring with personalized subjective preferences. While ensuring the fairness of the basic ranking, it effectively improves the matching degree between the ranking results and user needs. At the same time, it balances the traffic allocation between new and old merchants in the ecosystem through differentiated support logic, avoids excessive concentration of resources, and improves the overall traffic distribution efficiency.

[0056] Example 2, refer to Figure 1-3 As shown, after completing the screening of candidate merchants and the collection of operational data, the basic comprehensive score calculation step is entered to complete the standardization and quantification of objective operational data and obtain the basic ranking benchmark for all candidate merchants.

[0057] Specifically, the distance data for each candidate merchant is first obtained. This distance data is the straight-line distance between the merchant's business address and the target user's current geographical location. This data is a reverse indicator, where a lower value indicates better performance. A normalized score is calculated for the distance data using a reverse normalization formula, as follows:

[0058] Converse indicator normalization:

[0059] ;

[0060] = ;

[0061] In the formula, Score dist This is a standardized score for the distance dimension, ranging from 0 to 1. A higher score indicates better performance of the inverse indicator (i.e., a lower original value for the inverse indicator). V dist V represents the actual straight-line distance between the merchant and the user, calculated based on the user's current geographic location coordinates and the merchant's business address coordinates. dist_max V dist_min The maximum and minimum thresholds for the distance dimension are preset and matched with the service range rules. They are pre-set and fixed, and all merchants share the same threshold.

[0062] Score base The basic comprehensive score, ranging from 0 to 1, is the core benchmark for generating the initial sorting list. A higher score indicates better objective operational performance of the candidate merchants.

[0063] Furthermore, the calculation rules for all candidate merchants are completely uniform, with no differentiated weight allocation or normalization logic.

[0064] This embodiment ensures the fairness of the basic ranking by adopting a unified normalized calculation logic, ensuring that candidate merchants who are closer in distance can obtain higher basic scores, avoiding interference from human factors on the basic ranking results, and the standardized calculation logic greatly improves the efficiency and consistency of score calculation.

[0065] In this embodiment, the setting of a fixed threshold ensures the transparency of the basic sorting rules, providing a stable baseline for subsequent personalized preference matching and scenario-based adjustments, and effectively reducing the computational complexity of subsequent sorting steps.

[0066] Example 3, refer to Figure 1-3 As shown, after calculating the basic comprehensive scores of all candidate merchants, the process proceeds to the multi-dimensional personalized preference modeling and matching step. This step quantifies the matching between user needs and candidate merchants, obtaining a personalized preference score for each candidate merchant. Specifically, based on the user's full historical valid behavior data, the preference weights for four fixed priority dimensions are calculated. The formulas for calculating the preference weights for each dimension are as follows:

[0067] ;

[0068] In the formula, W d The user preference weights for the corresponding dimensions can be assigned to four dimensions in sequence: product category, content keywords, price range, and consumption time period, N. d N represents the corresponding statistical value of user's historical behavior under this dimension. total This is the total statistical value of user's historical behavior under this dimension. Both types of statistical values ​​are obtained based on the full set of valid historical user behavior data, without sampling or selective calculation.

[0069] After calculating the preference weights for each dimension, the matching score between the user and the current candidate merchant is calculated for each dimension. Then, the matching scores for each dimension are multiplied by the fixed priority weights of the corresponding dimensions and summed to obtain the final personalized preference score. The specific calculation formula is as follows:

[0070] ;

[0071] ;

[0072] In the formula, Score pref The total personalized preference score, ranging from 0 to 1, represents the degree of match between the user's overall needs and the current candidate merchants. It serves as the core basis for subsequent merchant recommendation selection. c The category dimension has a preset fixed priority weight, representing the contribution of category matching degree to the total preference score. It is the highest priority weight among the four dimensions. W k The content keyword dimension has a preset fixed priority weight, representing the contribution of content keyword matching degree to the total preference score. Its priority is second only to the category dimension. p The price segment dimension has a preset fixed priority weight, representing the proportion of the price segment's relevance to the overall preference score. Its priority is second only to the content keyword dimension.t The preset fixed priority weight for the consumption time period dimension represents the contribution ratio of the consumption time period matching degree to the total preference score, and is the lowest priority weight among the four dimensions;

[0073] S c The category matching score ranges from 0 to 1. When the category of the candidate merchant overlaps with the category preferred by the user, the score is 1. When the category of the candidate merchant does not overlap with the category preferred by the user, the score is 0. The higher the score, the higher the matching degree between the category of the candidate merchant and the category preferred by the user.

[0074] S k The content keyword matching score ranges from 0 to 1. A value of 1 indicates an exact match between the candidate merchant's content name and the user's historical content name; a value of 0 indicates a mismatch. k =m / N, where m is the number of elements in the intersection of the user's historical content core noun set and the candidate merchant's content core noun set. Core nouns are extracted using standardized text processing rules (removing special characters and extracting core content words). N is the total number of core nouns corresponding to all valid historical order content of the user, obtained based on user historical behavior data statistics. When there are no matching core nouns, S... k The value is 0;

[0075] S p The price range matching score ranges from 0 to 1. The score is 1 when the average order value of the candidate merchants falls within the user's preferred price range, and 0 when the average order value of the candidate merchants does not fall within the user's preferred price range.

[0076] S t The score is the consumption time period matching score, with a value range from 0 to 1. If the current request system time belongs to the user's preferred consumption time period range, the value is 1; if the current request system time does not belong to the user's preferred consumption time period range, the value is 0.

[0077] Furthermore, the model building and score calculation logic are completely unified for all users, with no differentiated dimension settings or weight allocations.

[0078] This embodiment achieves a quantitative characterization of user needs through multi-dimensional user behavior analysis. The fixed priority weight setting ensures that the core need dimension dominates the matching calculation. The keyword matching logic that prioritizes accurate matching effectively improves the accuracy of content matching and avoids irrelevant merchants being judged as having a high matching degree.

[0079] In this embodiment, the weight calculation based on the full historical behavior fully reflects the user's true needs and preferences. The hierarchical priority design takes into account the balance of multi-dimensional needs. The standardized matching rules ensure the objectivity and reproducibility of preference score calculation, providing a matching degree benchmark for subsequent adaptive adjustments in special scenarios, and effectively improving the fit between the ranking results and user needs.

[0080] Example 4, refer to Figure 1-3 As shown, after completing the full calculation of the basic comprehensive score and personalized preference score, the system enters the special scenario adaptive adjustment step to correct the scores of newly joined merchants and users with low activity levels, adapting to the ranking requirements under special scenarios. Specifically:

[0081] For newly registered merchants who meet the onboarding duration requirements, the total number of their initial campaigns is first determined. If the total number of campaigns is greater than or equal to a preset campaign count threshold, the campaign will be prominently displayed within a preset time period and range. If the total number of campaigns is less than the preset campaign count threshold, the merchant's original base score will be multiplied by a fixed weighting coefficient for weighting. The weighting formula is as follows:

[0082] ;

[0083] In the formula, Score base_now The weighted base score for new merchants, Score base Based on the overall score, W boost The preset fixed weighting coefficient applies uniformly to all eligible new merchants and is pre-set and fixed.

[0084] For users with insufficient historical behavior data, the total number of their valid historical behaviors is first counted. If the total number is zero, the personalized preference score calculation step is skipped, and only the basic comprehensive score is used for sorting. If the total number is in a preset low-frequency range, their original personalized preference score is multiplied by a fixed reduction factor and reduced. The reduction formula is as follows:

[0085] ;

[0086] In the formula, Score pref_cold Score is the reduced personalized preference score. pref W is the total score for personalized preferences. discount The preset fixed reduction factor applies uniformly to all users in the low-frequency range; it is pre-set and fixed.

[0087] This embodiment solves the matching deviation problem caused by the lack of exposure for new merchants and insufficient behavioral data for new users through differentiated scenario adaptation logic. Fixed threshold and coefficient settings ensure the fairness of the adaptation logic, avoid special adjustments for specific merchants or users, and protect the balance of rights and interests of all participants in the ecosystem.

[0088] In this embodiment, without compromising the fairness of the overall ranking rules, a reasonable traffic support channel is provided for newly joined merchants. At the same time, through a gradual preference score release logic, the impact of user preference model bias on the ranking results during the cold start phase is reduced, thereby improving the rationality of the ranking results and user experience in special scenarios.

[0089] Example 5, refer to Figure 1-3 As shown, after the scores of all candidate merchants are corrected, the process enters the hybrid insertion sorting step to complete the integration of basic sorting and personalized recommendation, generating the final displayable ordered list.

[0090] Specifically, the final display list is first divided into the first screen area, the second screen area, and subsequent areas according to the display order. A fixed number and position of insertion slots are set in the first screen area. Then, all candidate merchants are sorted in descending order according to their personalized preference scores. The top high-match merchants corresponding to the number of insertion slots are selected and filled into the preset insertion slots. If the original ranking of the merchant to be inserted in the initial sorting list generated by the basic comprehensive score is better than the position of the insertion slot, its original ranking is retained and it is not inserted again. After the insertion is completed, the uniqueness of the final list is checked.

[0091] During the verification process, each merchant's identifier is checked one by one according to the order of the list display. All merchants that have appeared are recorded. If a duplicate merchant that has already been recorded is encountered later, it is removed directly. The empty space after removal is filled by subsequent merchants in order. The original order of non-duplicate merchants is not changed throughout the process. Finally, all duplicate merchants are removed to obtain the final sorted list that can be displayed.

[0092] Furthermore, the insertion process only adjusts the merchants at the insertion position, without changing the original order of merchants not at the insertion position in the initial sorting list. The selection of top merchants is based solely on the sorting results of personalized preference scores, without any manual intervention or adjustment of the priority of specific merchants.

[0093] This embodiment, through the fixed-position insertion design, ensures the overall orderliness of the basic sorted list while achieving precise exposure of highly matched merchants, avoiding the unpredictability of results brought about by completely personalized sorting, and the uniqueness verification design avoids the decline in user experience caused by repeated display.

[0094] In this embodiment, the fairness of the basic ranking and the adaptability of personalized recommendations are taken into account. The fixed insertion position setting ensures the transparency and controllability of the traffic tilt rules, effectively balancing the traffic needs of different participants in the platform ecosystem, while avoiding the consumption of computing resources caused by full re-sorting and improving the running efficiency of the ranking.

[0095] Example 6, refer to Figure 1-3 As shown, the operational data also includes rebate amounts and rebate differences collected through standardized interfaces connecting to the public data ports of all candidate merchants. The rebate amount is the amount that merchants offer users in exchange for a rebate, and the rebate difference is the difference between the merchant's advertised rebate amount and the actual rebate amount. Both types of data are publicly available and searchable information from the merchants. The collection process directly connects to the merchants' public data ports, and all merchants' rebate data uses a completely uniform collection cycle and update frequency. There are no differentiated collection rules for specific merchants. After data collection, normalization is performed on the different attributes of the two types of rebate data, outputting a standardized score in the 0-1 range. The specific calculation formula is as follows:

[0096] Positive normalization formula for rebate amount:

[0097] ;

[0098] In the formula, Score rebate This is a standardized score for the rebate amount; a higher score indicates better performance. rebate The actual rebate amount for merchants is obtained directly through the merchant's public data portal. rebate_max V rebate_min The maximum and minimum thresholds for the rebate amount are preset and fixed based on business strategies, and all merchants share the same threshold.

[0099] Rebate difference reverse normalization formula:

[0100] ;

[0101] In the formula, Score diff This is a standardized score for the rebate difference dimension; a higher score indicates better performance. diff The actual rebate difference for merchants is obtained directly through the merchant's public data portal. diff_max V diff_min The maximum and minimum thresholds for the rebate difference are preset and fixed based on business strategies, and all merchants share the same threshold.

[0102] After the normalization calculation of the two types of rebate data is completed, the rebate amount score and the rebate difference score are not used as mandatory input items for the calculation of the basic comprehensive score, but only as auxiliary reference dimensions in the sorting process. They can be used as the basis for secondary sorting in the case of the same basic comprehensive score, as optional reference tags displayed on the user end, and as reference indicators for operators to make manual traffic fine-tuning. They can flexibly adapt to the dimension configuration requirements of different business scenarios.

[0103] In this embodiment, the standardized processing of rebate-related dimensions is implemented as an optional extension module. It can be flexibly selected as an auxiliary reference dimension according to business needs. The positioning of the auxiliary reference also avoids the problem of rebate-related factors excessively interfering with the fairness of the basic ranking, and provides richer dimension supplements without destroying the fairness of the ranking.

[0104] Example 7, refer to Figure 1-3 As shown, the present invention also provides an information sorting device based on a multidimensional data weighting and personalization model, including a data acquisition module, a comprehensive scoring module, a preference analysis module, and a list generation module;

[0105] Specifically, the data acquisition module first connects to the standardized data ports of the user and merchant ends to collect the required user location, time information, historical behavior data, and operational data of candidate merchants. After performing integrity verification and outlier filtering on all collected data, it distributes the data to the subsequent processing modules.

[0106] Subsequently, the comprehensive scoring module standardizes the multi-dimensional operational data and calculates the basic comprehensive score for each candidate merchant according to the preset weighting rules. The preference analysis module builds a multi-dimensional preference model based on the user's full historical behavior data and calculates the user's personalized matching score for each candidate merchant. Finally, the list generation module combines the basic score, matching score, new merchant support rules, and mixed insertion rules to generate the final display list. All operations in the entire process are logged to ensure the traceability of the sorting process.

[0107] Furthermore, the processing modules in this embodiment adopt a decoupled design, and the calculation rules of each module can be adjusted independently according to business needs without modifying the overall operation logic, thus adapting to sorting requirements in different scenarios.

[0108] This embodiment achieves standardization and traceability of the entire sorting process through modular design. The modules work together to complete the sorting calculation of the entire chain without manual intervention, which not only ensures the stability and consistency of the sorting results, but also has flexible scenario adaptability and reduces the iterative maintenance cost of sorting logic.

[0109] Example 8: 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.

[0110] 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. An information ranking method based on multidimensional data weighting and personalized models, characterized in that: Includes the following steps: Obtain the target user's current geographic location coordinates and current system time, and determine the candidate information set within the service area based on the current geographic location coordinates. The candidate information set consists of all valid merchants that meet the platform's service access rules. The system retrieves operational data for each merchant in the candidate information set. The collection of operational data is authorized by both the user and the merchant. The operational data includes distance data, which is the straight-line distance between the merchant's business address and the target user's current geographical location. The operational data is normalized, and the basic comprehensive score of each merchant is calculated based on the preset weights set according to the core priority of the business scenario. The historical behavior data of the target users is obtained, which is all the valid redemption behavior data generated by the users within the platform. Construct a multi-dimensional personalized preference model, calculate the personalized preference score of the target user for each merchant based on the multi-dimensional personalized preference model, generate an initial sorted list based on the basic comprehensive score, and filter and recommend merchants based on the personalized preference score; The recommended merchants are inserted into the predetermined positions in the initial sorting list according to the preset hybrid insertion strategy, generating the final display list.

2. The information ranking method based on multidimensional data weighting and personalization model according to claim 1, characterized in that: The steps for normalizing operational data and calculating a basic comprehensive score for each merchant based on preset weights include: For positive indicators, forward normalization is used, and for negative indicators, reverse normalization is used. The normalized scores of each dimension are multiplied by their corresponding preset weights and then summed to obtain the basic comprehensive score for each merchant. The higher the value of a positive indicator, the higher the normalized score; the higher the value of a negative indicator, the lower the normalized score. The sum of the preset weights for each dimension is one. The weight allocation is determined based on the core priority of the business scenario to ensure that the contribution of core indicators to the basic comprehensive score meets business expectations. The entire score calculation process uses standardized calculation logic, and the calculation rules for all candidate merchants are completely consistent to ensure the fairness and comparability of the basic comprehensive scores.

3. The information ranking method based on multidimensional data weighting and personalization model according to claim 1, characterized in that: The steps for constructing a multi-dimensional personalized preference model include: Calculate the weights for category preference, content keyword preference, price range preference, and consumption time period preference respectively, and allocate the weights of each dimension according to a preset fixed priority. The sum of the weights of each dimension is one. Category preference dimension corresponds to the highest priority weight; The priority weight of the content keyword preference dimension is second only to the category preference dimension. The price segment preference dimension has the second highest priority weight, only after the content keyword preference dimension. The lowest priority weight corresponds to the consumption time period preference dimension; Once the weights for each dimension are calculated, they are fixed parameters of the preference model. The priority level of the weights remains fixed and the priority order is not adjusted as the scenario changes. The weight of the category preference dimension is determined by the proportion of the number of redemptions for the target category to the total number of redemptions; The weight of the content keyword preference dimension is determined by the frequency of the content keywords in the total orders in the historical orders; The weight of the price segment preference dimension is determined based on the proportion of consumption in the corresponding price range in the user's historical behavior; The weight of the consumption time period preference dimension is determined based on the proportion of consumption during the corresponding time period in the user's historical behavior; The weights for all dimensions are calculated based on the user's full historical behavioral data, ensuring that the preference model can fully reflect the user's true needs and tendencies.

4. The information ranking method based on multidimensional data weighting and personalization model according to claim 3, characterized in that: The matching logic for the content keyword preference dimension includes: When the content name provided by the merchant is exactly the same as the content name in the user's historical behavior, it is judged as an exact match and assigned a first match score. When the content name provided by the merchant is not exactly the same as the content name in the user's historical behavior, the core nouns of the user's historical content name are extracted and matched with the content name provided by the merchant. A second match score is assigned based on the inclusion of the core nouns. The first matching score is higher than the second matching score. The calculation rules for the matching score are consistent for all merchants to ensure the objectivity and reproducibility of the matching results. The extraction of core terms is completed based on standardized text processing rules. All merchants' content names adopt a unified text preprocessing process, including special character removal, unified character encoding conversion, and redundant content filtering steps. The preprocessed text is then used for keyword matching calculation to ensure the accuracy and consistency of the matching results.

5. The information ranking method based on multidimensional data weighting and personalization model according to claim 1, characterized in that: The method also includes traffic support strategies for newly joined merchants: Obtain the total number of activity shares for the first time launched by a newly joined merchant. Set a threshold for the number of activity shares based on the platform's historical data and business strategy. When the total number of activity shares is greater than or equal to the threshold, the merchant will be displayed at the top within a preset distance range for a preset time period. When the total number of activity shares is less than the threshold, a preset weighting coefficient will be added to the merchant's basic comprehensive score within a preset period. The traffic support strategy only applies to new merchants whose onboarding time has not exceeded the preset onboarding time threshold. Merchants that have passed the new merchant certification period are not eligible for the support strategy. The applicable conditions for the support strategy are the same for all new merchants. The onboarding time of a new merchant is calculated from the time when the merchant completes the platform onboarding review. The total number of activities is calculated based on the total number of redeemable activities for all valid activities published by the merchant after onboarding. The weighting coefficient is a fixed value and the same weighting coefficient is used for all eligible new merchants. The scope and duration of the featured display are based on a unified standard and are not differentiated for specific merchants.

6. The information ranking method based on multidimensional data weighting and personalization model according to claim 1, characterized in that: The steps for inserting recommended merchants into predetermined positions in the initial sorting list according to a preset hybrid insertion strategy include: The final display list is divided into a first screen area, a second screen area, and subsequent areas. Multiple fixed insertion positions are set in the first screen area. Top merchants are selected from the merchants sorted by personalized preference scores to fill the insertion positions. When the original ranking of the selected merchant in the initial sorting list is better than the insertion position, its original ranking is retained and it is not inserted again. The number and position of the insertion positions are fixed and will not be adjusted with changes in merchants or users. The number of top merchants selected is consistent with the number of insertion positions. The insertion process does not change the original order of merchants in the initial sort list that are not in the insertion positions. The personalized preference scores are sorted in descending order. The top merchants are the merchants whose ranking corresponds to the number of insertion positions in the sorting results. After the insertion operation is completed, the uniqueness of all merchants in the final list is checked. The list after the check is completed is the final sorted list that can be displayed.

7. The information ranking method based on multidimensional data weighting and personalization model according to claim 1, characterized in that: The method also includes user cold start processing, specifically: The system detects the number of orders corresponding to the target user's historical behavior data. When the number of orders is zero, the personalized preference score calculation step is skipped, and the final display list is generated based solely on the basic comprehensive score. When the number of orders is in the preset low-frequency range, the calculated personalized preference score is multiplied by the preset reduction coefficient. The reduction coefficient is a fixed value and the same reduction coefficient is used for all users in the low-frequency range. The triggering condition of the cold start processing logic is consistent for all users. The statistical scope of historical behavior data includes all valid redeemed orders generated by users within the platform, excluding unredeemed canceled orders and abnormal orders. The value range of the low-frequency range is fixed, and the personalized preference score after reduction still participates in the subsequent recommended merchant selection and insertion process according to the original logic.

8. The information ranking method based on multidimensional data weighting and personalization model according to claim 1, characterized in that: The operational data also includes rebate amount and rebate difference. The rebate amount is the amount that can be returned to users by the merchant for the activity. The rebate difference is the difference between the activity rebate amount marked by the merchant and the actual amount that can be returned. The rebate amount, rebate difference and distance data are all publicly available data of the merchant. The data collection process directly connects to the merchant's public data port. All merchants' operational data adopt a unified collection cycle and update frequency.

9. An information sorting apparatus for implementing the information sorting method based on multidimensional data weighting and personalization model as described in any one of claims 1-8, characterized in that: It includes a data acquisition module, a comprehensive scoring module, a preference analysis module, and a list generation module; The data acquisition module is configured to obtain user location, time information and merchant operation data. The data acquisition process uses standardized interface connection, and all data is checked for integrity and filtered for outliers before entering the subsequent processing flow. The comprehensive scoring module is configured to calculate the merchant's basic comprehensive score based on a normalization algorithm. The preference analysis module is configured to build a multi-dimensional preference model based on the user's historical behavior and calculate a personalized score. The model parameters are generated based on the user's full historical data. The list generation module is configured to combine new merchant support strategies with a hybrid insertion strategy to generate the final sorted list. The entire list generation process is traceable, and all operations are logged to ensure the auditability of the sorting results.

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 8.

Images