Big data-based personalized offer recommendation system for user consumption and merchant habits

Abstract

This invention discloses a personalized discount recommendation system based on big data user consumption and merchant habits, belonging to the field of computer information management technology. It includes a merchant service platform, as well as an operation management component and a data processing component. The operation management component acquires relevant operational data from the merchant service platform, cleans it, and categorizes it into interest-related data, discount-related data, and comprehensive data, which is then sent to the operation identification component. This invention constructs a full-link computational logic from basic user characteristics to discount adaptation and comprehensive decision-making, solving the common problems of inaccurate user profiling and weak scenario adaptation capabilities in traditional recommendation systems. It achieves the core objective of "pushing the right business discounts to the right users at the right time," providing a complete and implementable technical solution for personalized discount recommendations in multiple business scenarios, and promoting the development of this field from extensive push to refined, dynamic, and multi-objective balanced approaches.

Description

Technical Field

[0001] This invention relates to the field of computer information management technology, specifically to a personalized discount recommendation system based on big data and user consumption and merchant habits. Background Technology

[0002] With the rapid development of local life services and e-commerce retail, merchants generally have multiple business lines, and promotional marketing has become a core operational method for merchants to increase user activity and revenue. Currently, various personalized recommendation systems are widely used in promotional push scenarios. These systems typically use algorithms such as collaborative filtering and content recommendation based on users' historical consumption data to initially identify user preferences, and then combine this with basic information about promotional activities to complete the push notification. This improves the accuracy of promotional marketing to a certain extent. Compared to the earlier indiscriminate push notifications, it significantly reduces the waste of marketing resources and provides users with consumption information that better meets their basic needs, becoming an important technological support connecting merchants and users.

[0003] Existing personalized discount recommendation systems still have room for improvement in multiple dimensions during practical application: At the user interest identification level, they mostly calculate preferences based on the absolute value of users' historical spending in a single business, without designing correction logic for specific scenarios. This easily leads to misjudgments of users' actual spending power and spending habits during specific times. Furthermore, they fail to consider the dynamic changes in users' spending interests and lack the ability to explore the interest migration effect across different businesses. At the discount and scenario adaptation level, most systems use the discount amount as a uniform value measure, failing to consider the differences in perception of the same discount among users with different spending power, and failing to adjust to dynamic factors such as the urgency of promotional activities. This makes it difficult to adapt to the differentiated operational needs of merchants at different stages of business. At the comprehensive recommendation architecture level, it is difficult to balance the multi-dimensional relationships between user needs, merchant demands, and platform operational goals. It lacks a mechanism to identify deeper behaviors during the decision-making hesitation period, such as browsing without placing an order or adding items to the cart without payment. The recommendation results are prone to becoming overly rigid or deviating from users' actual needs. Therefore, it is clear that existing discount recommendation systems still need improvement. Summary of the Invention

[0004] The purpose of this invention is to provide a personalized discount recommendation system based on big data user consumption and merchant habits, thereby solving the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a personalized discount recommendation system based on big data user consumption and merchant habits, including a merchant service platform, an operation management component, and a data processing component. The operation management component acquires relevant operation data from the merchant service platform and transmits the collected operation data to the data processing component. The data processing component cleans the operation data. The discount recommendation system also includes an operation identification component. The cleaned operation data is divided into interest-related data, discount-related data, and comprehensive related data, and then transmitted to the operation identification component. The operation identification component:

[0006] Based on interest-related data such as adjusted spending amount, total user spending amount, time preference weight, month-on-month change in spending amount, spending amount in the previous 30 days, smoothing constant, interest decay coefficient, migration interest degree, and migration impact coefficient, the user dynamic interest degree is output, which measures the user's preference for specific services.

[0007] Based on user dynamic interest and discount-related data such as remaining discount days, maximum activity days, discount perception intensity, maximum perception intensity, inventory weight, corrected effective push count, push frequency limit, push fatigue penalty coefficient, time period matching degree, and time period matching influence coefficient, the user discount recommendation priority is output, and the user discount recommendation priority is used as the recommendation ranking basis for user interests and discount features.

[0008] The final recommendation score is output based on the base score, the weighted sum of the enhancement factor and the penalty factor, and is used as the final basis for recommendation ranking.

[0009] Optionally, the operation identification component includes a dynamic interest judgment module, a discount recommendation judgment module, and a final score calculation module.

[0010] Optionally, the processing logic of the dynamic interest judgment module is as follows: starting from the user's historical consumption behavior, first correct the interference factors of consumption amount and consumption time, then superimpose the consumption trend change and cross-business interest migration effect, and finally output the user's comprehensive preference for a single business, and use the user's comprehensive preference for a single business as the user-side basic feature of the entire recommendation system.

[0011] Optionally, the processing logic of the discount recommendation judgment module is based on the user's dynamic interest, and superimposed with adjustment factors of five dimensions: the timeliness and urgency of the discount itself, the user's perceived value of the discount, inventory operation strategy, user push fatigue, and current time matching degree, to output the real-time recommendation priority of a single discount for a single user, thus serving as an intermediate layer connecting user interests and operation strategies.

[0012] Optionally, the processing logic of the final score calculation module is as follows: a layered weighted architecture is adopted, the basic layer retains the core weights of user interest and discount timeliness, the enhancement layer adds five positive dimensions of merchant operation needs, user value, market trends, social influence and user intention, the penalty layer adds the negative dimension of geographical distance, and finally normalizes to a fixed interval output, so that the score calculation takes into account recommendation accuracy, operational flexibility and result stability.

[0013] Optionally, the dynamic interest judgment module includes: a basic consumption ratio item obtained by dividing the user's corrected consumption amount in business b by the user's total consumption amount across all businesses; a time preference weight item obtained by dividing the user's peak-hour consumption ratio by the maximum value of the consumption ratios across all time periods; a consumption trend gain item obtained by dividing the difference between the consumption amount in the last 30 days and the consumption amount in the previous 30 days by the consumption amount in the previous 30 days, multiplying by the interest decay coefficient, and adding 1; and a cross-business migration gain item obtained by multiplying the cross-business migration interest degree by the migration impact coefficient and adding 1.

[0014] Optionally, the discount recommendation judgment module includes: a timeliness factor obtained by subtracting the remaining days of the discount from the maximum number of activity days and dividing by the maximum number of activity days; a discount perception factor obtained by dividing the user's perception intensity of the discount by the platform's maximum perception intensity; an inventory weight item calculated for three scenarios: hot-selling, clearance, and normal; a push fatigue penalty item obtained by dividing the corrected number of effective pushes by the upper limit of push frequency, multiplying by the fatigue penalty coefficient, and subtracting the value from 1; and a time period matching gain item obtained by multiplying the matching degree between the current time and the user's consumption time period by the time period matching influence coefficient and adding 1.

[0015] Optionally, the final score calculation module includes: a weighted sum of enhancement factors obtained by normalizing the five enhancement dimensions to intervals, multiplying them by their corresponding weights, and summing the results; a weighted sum of penalty factors obtained by dividing the distance between the user and the service by the maximum service radius, taking the minimum value of the result and 1, multiplying it by the distance decay weight; and a normalized final score obtained by adding the base score, enhancement score, and penalty score, and then using a truncation function to restrict the result to an interval.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] I. This invention constructs a multi-dimensional correction system for calculating dynamic user interests. First, it eliminates interference from abnormal average order values ​​and group-buying behavior through consumption amount correction logic, restoring the user's true individual consumption contribution. Then, it eliminates time deviations in cross-time zone orders and pre-booked orders through consumption time correction logic, accurately identifying users' stable consumption habits during peak hours. On this basis, a consumption trend gain item is superimposed, which can capture the dynamic changes in user interests, rather than relying solely on static historical consumption data. The additional cross-business interest migration factor mines users' potential interests in related businesses through business similarity and historical migration rate. When new users have no historical data, the platform's general migration rate can be directly called, solving the cold start problem in new business and new user scenarios from the bottom up.

[0018] This system changes the logic of traditional systems that judge preferences based solely on the absolute value of single-business consumption. The output of user interest includes not only long-term stable consumption habits, but also short-term trend changes and potential migration needs, providing accurate user-side basic characteristics for system-wide recommendations and realizing a full-dimensional and dynamic characterization of user preferences.

[0019] Second, this invention constructs a real-time discount adaptation system centered on user interests. It introduces a discount perception threshold factor designed based on Weber's Law, realizing the personalized value quantification of the same discount amount for users with different spending power, avoiding the bias of uniform discount value judgment; the timeliness factor matches users' near-expiration decision-making psychology in a linearly decreasing manner, improving the utilization efficiency of discount resources at the end of the validity period; the multi-scenario configurable inventory weight item can flexibly switch between three strategies—hot-selling, inventory clearance, and normal sales—according to the merchant's operational goals, adapting to the needs of different business stages without modifying the core calculation logic; the push fatigue penalty item avoids the damage to user experience caused by excessive pushes through multi-dimensional correction of the effective push count, while taking into account the weight differences of cross-channel pushes; the consumption time matching gain item solves the boundary judgment problem of cross-midnight time periods through cyclic difference calculation, realizing dynamic matching between recommendation time and users' immediate needs.

[0020] This invention optimizes the problem of traditional system discounts being disconnected from users and scenarios. The priority of the output discounts can be adjusted in real time based on factors such as timeliness, inventory, time, and user acceptance. This not only ensures the personalization of discount recommendations but also enables the flexible implementation of platform operation strategies, greatly improving the scenario adaptability and real-time effectiveness of discount recommendations.

[0021] Third, the basic modules of this invention safeguard the core weights of user interest and preferential value, ensuring that the recommendation results always match the user's real needs; the enhancement factor module, through a dynamically adjustable weight system, simultaneously supports positive enhancements across five dimensions: merchant promotion needs, protection of high-value user rights, guidance of platform business growth trends, social trust endorsement conversion, and identification of user hesitation period intentions, achieving a balance of value among users, merchants, and the platform; the penalty factor module, through geographical distance attenuation logic, ensures fulfillment efficiency and user consumption experience; the final score cutoff mechanism avoids the solidification of recommendation results caused by extremely high values, and also ensures the minimum recommendation probability for all services through a minimum threshold of 0.2, preventing the formation of information cocoons and preserving the possibility for users to explore new services. Attached Figure Description

[0022] Figure 1 This is a logic diagram of the operation identification component of the present invention;

[0023] Figure 2 This is a schematic diagram of the calculation process of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example: Please refer to Figures 1 to 2 This invention provides a personalized discount recommendation system based on big data user consumption and merchant habits, including an operation management component and a data processing component. The operation management component acquires big data related to user consumption and merchant habits, and sends the collected big data to the data processing component for cleaning. The discount recommendation system also includes an operation identification component. The cleaned big data is divided into interest-related data, discount-related data, and comprehensive related data, and then sent to the operation identification component, which includes a dynamic interest judgment module.

[0026] The processing logic of the dynamic interest judgment module is as follows: based on the corrected consumption amount C in the interest-related data. u,b corr Total user spending C u,total Time preference weight, month-on-month change in spending amount ΔC u,b Spending amount C in the first 30 days u,b prev Smoothing constant e, interest decay coefficient μ, migration interest degree I u,bmigrate and migration influence coefficient γ migrate Output user dynamic interest level I u,b Through user dynamic interest I u,b Quantifying the degree of user preference for specific services, that is:

[0027]

[0028] The time preference weight is calculated as the user's spending percentage during peak hours divided by the maximum spending percentage across all time periods. Starting from the user's historical spending behavior, it first corrects for confounding factors such as spending amount and time, then incorporates changes in spending trends and cross-business interest migration effects. The final output is the user's overall preference for a single business. This overall preference serves as the fundamental user-side feature of the entire recommendation system, specifically including:

[0029] The basic consumption percentage is calculated by dividing the user's adjusted spending amount in service b by the user's total spending across all services. This percentage measures the user's relative preference for different services, avoiding absolute value biases caused by differences in the user's total spending level, and directly reflecting the user's share of spending on that service. The time preference weight is obtained by dividing the user's peak-hour consumption percentage by the maximum consumption percentage across all time periods. The more concentrated the user's consumption periods, the more stable their consumption habits, and the higher the matching degree of service recommendations for those periods. This ratio amplifies the time preference weight of highly concentrated users, achieving personalized time matching. The consumption trend gain is obtained by dividing the difference between the spending amount in the last 30 days and the previous 30 days by the spending amount in the previous 30 days, multiplying by the interest decay coefficient, and adding 1. A month-on-month increase in spending amount indicates rising user interest, receiving a positive gain; a month-on-month decrease indicates weakening interest, receiving a negative decay, enabling dynamic prediction of changes in user interest. The cross-business migration gain term is obtained by multiplying the cross-business migration interest by the migration impact coefficient and adding 1. Users' consumption behavior in similar businesses will generate interest migration. For example, users who often order afternoon tea may have a potential interest in dining in. This term can be used to explore potential needs that users have not directly expressed, thus solving the problem of cold start for new businesses.

[0030] The revised consumption amount C u,b corr Time preference weight and migration interest I u,b migrate All of these were calculated, and the total user spending amount C was obtained by summing the order table. u,total And the amount of spending in the previous 30 days C u,b prev e is a smoothing constant, taken as 1, and the interest decay coefficient μ and migration influence coefficient γ are also mentioned. migrateAll parameters were determined through parameter tuning. The range of 0.3-0.6 reflects short-term consumption trend changes without significantly altering long-term interest judgments due to fluctuations in single consumption transactions, balancing historical consumption stability and sensitivity to the latest trends. The range of 0.1-0.3 indicates that migrating interests are potential predictions and have a lower weight than direct consumption behavior, avoiding over-amplification of potential interests that could cause recommendations to deviate from users' actual needs. The remaining parameters can be obtained through conventional data collection methods, such as obtaining order tables from merchant service platforms, and will not be elaborated further.

[0031] Furthermore, the revised consumption amount C u,b corr The calculation method is as follows:

[0032]

[0033] Where P i,b P represents the actual payment amount made by user u for the i-th order in business b. avg,b For business b, the average order value across all platforms, O i Let represent the number of people in the group purchase for the i-th order, and n represent the total number of orders placed by user u in business b. Both are obtained through order information collection from the merchant service platform. β is the price deviation penalty coefficient, ranging from 0.5 to 1.5. If there is group purchase behavior in an order, by eliminating the interference of abnormal orders and group purchase behavior on the judgment of user spending power, the original order amount, which includes large abnormal consumption and group purchase by multiple people, is converted into a corrected amount that reflects the user's true spending power. This avoids false consumption behavior from misleading the calculation of interest and ensures the objectivity of user consumption contribution statistics.

[0034] When calculating time preference weights, it is necessary to count user consumption orders during peak hours. The calculation of peak hours requires the exclusion of abnormal orders. Specifically: valid order set = all orders - abnormal time order set. The consumption period is specifically defined as the scheduled delivery period (scheduled orders) and the order completion period (instant orders). By eliminating interference from cross-time zone business trips and travel, and the inconsistency between the order time and the consumption time of scheduled orders, the true distribution of user consumption periods can be restored. This accurately identifies the user's stable consumption habit periods and provides reliable basic data for subsequent time preference weight and period matching calculations, avoiding the distortion of time preference statistics caused by abnormal orders in special scenarios.

[0035] Furthermore, the transfer of interest I u,b migrate The calculation method is as follows:

[0036]

[0037] Where I u,b1 To calculate the user's interest in related service b1, the formula is used in the same way as above, S b1,b反映The similarity between business b1 and business b, M b1→b This represents the historical migration rate from business b1 to b, i.e., the percentage of users who consumed b1 and then consumed b. The platform-wide migration rate reflects the average migration rate of all users across the platform from business b1 to b. Both are derived through business configuration and user behavior chain statistics. α is used as the credibility weight of user historical data, taking values ​​of [0,1]. Considering the phenomenon of "interest migration" in user consumption behavior, for example, if a user frequently consumes "afternoon tea," they may be cultivating a potential interest in "dine-in." Traditional systems calculate the interest level of each business independently, ignoring the correlation migration effect between businesses. This factor uses business correlation to predict potential interests and solves the cold start problem.

[0038] By capturing the cross-business correlation effect of user consumption behavior, and quantifying users' potential interests in related businesses through business similarity and historical migration rate, the platform can directly connect to the general migration rate when new users have no historical data. It can predict users' potential preferences without them having consumption records in the target business, thus solving the gap in interest calculation when there is no user behavior data in the early stage of new business launch.

[0039] Furthermore, the operation identification component also includes a discount recommendation judgment module, the processing logic of which is as follows:

[0040] Based on user dynamic interest u,b And the remaining days of the discount (D) in the discount-related data. b Maximum number of activity days D max Perceived Discount Strength V perceived Maximum perceived intensity V max Inventory weight, number of valid pushes after correction (N) u,b corr Push frequency limit N max Push fatigue penalty coefficient ω, time period matching degree T match The influence coefficient λ of matching time period match Output user discount recommendation priority P u,b Recommendation priority P based on user discounts u,b The recommendation ranking is based on user interests and discount features, that is:

[0041]

[0042]

[0043]

[0044] The inventory weight is determined by the operations staff based on the appropriate inventory strategy chosen according to business objectives;

[0045] Based on users' dynamic interests, and incorporating five adjustment factors—the urgency of the offer itself, the user's perceived value of the offer, inventory management strategies, user push fatigue, and the current time relevance—a real-time recommendation priority for each offer is output for each individual user. This serves as an intermediary layer connecting user interests and operational strategies, specifically including:

[0046] The urgency factor is obtained by subtracting the remaining days of the discount from the maximum number of days of the promotion, and then dividing by the maximum number of days of the promotion. Users face greater decision-making pressure and higher conversion rates when the discount is about to expire. Prioritizing near-expiration discounts is achieved through a linear decreasing approach, leveraging users' loss aversion to boost conversion rates. The discount perception factor is obtained by dividing the user's perceived discount intensity by the platform's maximum perceived intensity. Based on Weber's Law, user perception of discounts is positively correlated with their spending power. For example, a 10 yuan discount is far more attractive to users with an average order value of 20 yuan than to those with an average order value of 100 yuan. Normalizing the perceived intensity allows for a horizontal comparison of the attractiveness of different users and different discounts.

[0047] Inventory weights are calculated for three scenarios: hot-selling, clearance, and normal. These weights are tailored to different operational goals. In the hot-selling scenario, scarcity is used to improve conversion rates; in the clearance scenario, slow-moving inventory turnover is accelerated; and in the normal scenario, no additional adjustments are made. (θ and θ') 1 The values ​​are all between 0.1 and 0.3. This range reflects the guiding role of inventory strategies without allowing inventory factors to completely overshadow user interests, balancing operational goals and user experience. The push fatigue penalty is calculated by dividing the corrected effective push count by the push frequency limit, multiplying by the fatigue penalty coefficient, and then subtracting this value from 1. Repeatedly pushing similar offers can annoy users and decrease click-through rates. This variable dynamically reduces the priority of excessively pushed offers to avoid harassing users. ω is between 0.5 and 0.8. This range effectively suppresses excessive pushes without drastically reducing the priority of offers due to a few pushes, balancing reach efficiency and user experience. The time-matching gain is calculated by multiplying the match between the current time and the user's consumption time by the time-matching influence coefficient and then adding 1. Users have the highest demand during their usual consumption times; for example, the conversion rate of a 11:30 push for takeout offers is much higher than that of 14:00. This variable dynamically matches the recommendation time with user needs, improving instant conversion. λ match The value is 0.2-0.4. Time period matching is a short-term real-time factor, and its weight is lower than that of long-term interest and preferential value, so as to avoid the excessive influence of time period factors on the rationality of recommendations during off-peak hours.

[0048] Among them G b corr G represents the inventory tightness after the adjustment of business b. avg θ represents the average inventory tightness across all business segments on the platform, and θ represents the impact coefficient of inventory urgency in high-selling scenarios. 1This represents the inventory impact coefficient and perceived discount intensity V in the inventory clearance scenario. perceived 、Business b revised inventory tightness G b corr N valid push notifications after user u receives the corrected offer from service b u,b corr The degree of match between the current time and the user's consumption habit time period T match All parameters were calculated, while the remaining parameters were collected in real time by the system operation. Since these parameters are common parameters in the statistical tables, the specific collection process will not be described in detail.

[0049] Furthermore, the perceived strength of the discount, V perceived The calculation method is as follows:

[0050]

[0051] Where M b P represents the discount amount for service b. u,avg V represents the average order value for user u, e is a smoothing constant with a value of 1, and V upper To determine the upper limit of the perceived discount intensity of 3-5, and to avoid abnormally high perceived discount intensity due to high discounts on ultra-low-priced goods, the perceived discount factor is kept within a reasonable range.

[0052] Based on Weber's Law of Psychology, the logic of judging the value of discounts is reconstructed, and a fixed discount amount is converted into personalized perceived value for users with different spending power. This avoids the logical bias that "the same amount of discount has the same attraction for all users" and realizes the personalized quantification of discount value on the user side, ensuring that users at different spending levels can receive discount recommendations that meet their perceived expectations.

[0053] Furthermore, the matching degree T of users' consumption habits during different time periods match The calculation method is as follows:

[0054]

[0055] The time-period difference is calculated using cyclic interpolation:

[0056]

[0057] Specifically, ΔT match T represents the cyclic difference between the current time and the user's usual time period. now T represents the current time. u,b peakThis represents the peak time period for user u's consumption of business b, i.e., the time period with the highest probability density (24-hour timeframe). All the above parameters are obtained from internal information of the merchant service platform. σ is the time period tolerance width, with a value of 1-2 hours. By establishing a dynamic correlation between the current time and user consumption habits, the boundary judgment error across midnight time periods is solved through cyclic difference calculation. The intensity of user demand at different time points is quantified, and the recommendation priority of corresponding businesses is automatically increased during the time period when users are most likely to make consumption behavior. This realizes time-sensitive dynamic recommendation and improves the real-time fit of promotional pushes.

[0058] Furthermore, the adjusted inventory tightness G for business b b corr The calculation method is as follows:

[0059]

[0060] Q sold,b Q represents the quantity sold for business b. cancel,b malice Q indicates the number of orders maliciously canceled by business b. total,b Q represents the total inventory quantity of business b. reserve,b E represents the safety stock reserve for business b. b E represents an abnormal indicator in the inventory system of business b. threshold The above parameters can be obtained through the order system in the merchant service platform, and x is the anomaly penalty coefficient with a value of 0.3-0.5.

[0061] The actual sales volume is obtained by subtracting the number of maliciously canceled orders from the number of sold orders. This process eliminates fake sales volume caused by malicious order cancellations. The total inventory is then subtracted from the safety reserve inventory to obtain the total inventory actually involved in sales. The actual sales volume is divided by the actual available inventory to obtain the original inventory tightness. When the inventory system data fluctuates beyond the abnormal threshold, the reliability of the inventory tightness is reduced proportionally. The higher the degree of abnormality, the lower the correction factor, with a minimum of 0. The original inventory tightness is multiplied by the abnormality correction factor to obtain the final corrected inventory tightness.

[0062] By eliminating interference from abnormal system data and safety stock reservations, the true inventory tightness of the business is restored, providing a reliable basis for subsequent inventory weight calculations. This avoids misleading recommendation strategies with false hot-selling and false inventory data, ensuring that inventory-related operational strategies can be accurately implemented; the maximum value of 0 is taken to ensure that the correction factor is not negative.

[0063] Furthermore, the number of valid push notifications N after user u receives the corrected offer from service b. u,b corr The calculation method is as follows:

[0064]

[0065] Where m represents the total number of pushes, Ri represents the channel deduplication coefficient of the i-th push (1 for duplicates within the same channel, 0.5 for duplicates across channels), Vi represents the effective reach indicator of the i-th push (1 for users who have viewed it, 0.3 for those who haven't), and di represents the time decay coefficient of the i-th push (1.0 for intervals less than 1 day, 0.5 for intervals of 1 to 3 days, and 0.2 for intervals greater than 3 days). All of the above content is collected through push logs. By combining three factors—push channel, reach status, and time decay—the original number of pushes is converted into the effective number of pushes that users actually perceive. This avoids the problem of artificially high fatigue levels caused by duplicate pushes across channels and pushes that have not been reached. It accurately measures the user's acceptance of similar offers and provides an objective basis for push frequency control, improving reach efficiency while avoiding excessive user harassment.

[0066] Furthermore, the operational identification component also includes a final score calculation module, the processing logic of which is as follows:

[0067] Based on the base score F base Enhancement factor weighted sum F boost and the penalty factor weighted sum F penalty Output the final recommended score Final(F) u,b ), through the final recommendation score Final(F) u,b This serves as the final basis for recommendation ranking.

[0068] A layered weighted architecture is adopted. The base layer retains the core weights of user interest and the timeliness of offers. The enhancement layer adds five positive dimensions: merchant operational needs, user value, market trends, social influence, and user intention. The penalty layer adds a negative dimension: geographical distance. Finally, the results are normalized to a fixed range, which makes the score calculation take into account recommendation accuracy, operational flexibility, and result stability.

[0069]

[0070] Where ω base A value of 0.4 indicates that the basic priority of user interest and the timeliness of offers accounts for 40% of the total weight;

[0071]

[0072]

[0073] A is assigned a weight of 0.2, B a weight of 0.1, C a weight of 0.1, D a weight of 0.1, and E a weight of 0.1, which together account for 60% of the total weight.

[0074]

[0075] Where F is set to 0.1, it means that the distance decay penalty accounts for 10% of the total weight;

[0076]

[0077] After normalizing the five enhancement dimensions to intervals, multiply them by their corresponding weights and sum them to obtain the weighted sum of enhancement factors. Divide the distance between the user and the business by the maximum service radius, take the minimum value of the result and 1, multiply it by the distance decay weight to obtain the weighted sum of penalty factors. After adding the base score, enhancement score and penalty score, use a truncation function to restrict the result to an interval to obtain the normalized final score.

[0078] Final(F) u,b The final recommendation score is ), where Wb is the merchant's promotional weight for business b, ranging from 1 to 10, set by the merchant, and L... u For user u, the membership level is 0-5, ΔS b corr S represents the adjusted month-over-month increase in conversion rate for business B. b For business B, the current conversion rate is... u Let B be the geographic coordinates of user u. b To provide service coverage for the center point coordinates of business B, B max To maximize the platform's service radius, K u,b corr K represents the corrected number of valid users who have made purchases in user's social circle using service b. max H is the reference number with the greatest influence in the social circle. u,b For user u's decision-making hesitation regarding business b, the above represents the month-on-month increase ΔS after adjusting for the conversion rate of business b. b corr The corrected number of valid users K who have consumed services in user u's social circle (as per business b) u,b corr User u's decision-making hesitation H regarding business b u,b The results, after calculation and correction, indicate that the rest can be obtained through existing merchant service platforms;

[0079] Furthermore, user u's decision-making hesitation H regarding business b u,b The calculation method is as follows:

[0080]

[0081] Where k is the total number of uncompleted actions by user u for service b, A j view This is the identifier for the j-th visit to the details page of service b; 1 for viewed, 0 for not viewed. j collect This is the identifier for the j-th collection; 1 represents a collection, and 0 represents no collection. j cartLet ω be the identifier for the j-th time the item is added to the cart; 1 for adding to cart and 0 for not adding to cart. view The browsing weight coefficient is set to 0.2-0.4, ω cart The weighting factor for adding items to the cart is set to 0.5-0.8, ω. collect The collection weighting coefficient is set to 0.3-0.5, d j Let t be the time decay coefficient for the j-th action, with a value of 1.0 for intervals less than 1 day, 0.5 for intervals of 1 to 3 days, and 0.2 for intervals of 3 to 7 days. j Let T be the number of days since the j-th action. hesitate The cooling-off period is 7 days. tj≤Thesitate This represents a time window indicator function. The behavior is 1 within the window and 0 outside the window. All of the above parameters can be obtained from the window logs of the merchant service platform.

[0082] Users often exhibit a "hesitation period" before making a purchase: repeatedly viewing product details pages, adding items to their cart but not placing an order, or adding items to their favorites but not buying them. Traditional recommendation systems fail to recognize this "hesitation state," potentially missing the optimal time for push notifications. This factor identifies users' potential purchase intentions through incomplete behaviors such as browsing, adding to cart, and adding items to favorites, quantifying the degree of user decision-making hesitation. It automatically increases the recommendation weight of corresponding services for users in the critical purchase decision-making period, achieving a "final push" to promote conversion and filling the gap in traditional recommendation systems' recognition of deep user behavioral intentions.

[0083] Furthermore, the adjusted month-on-month increase in the conversion rate of business B is ΔS. b corr The calculation method is as follows:

[0084]

[0085] Where S b current S represents the current conversion rate of business b. b prev S represents the conversion rate of business b in the previous cycle. avg current S represents the platform's current average conversion rate. avg prev F represents the average conversion rate over the previous period on the platform. b For abnormal transaction characteristics of business b, such as concentrated order placement from the same IP address or duplicate device fingerprints, F max The threshold for identifying abnormal features is τ, which is the penalty coefficient for fraudulent transactions, ranging from 0.5 to 0.8. All of the above parameters can be obtained from the platform logs of the merchant service platform.

[0086] The conversion rate of the previous period was adjusted according to the overall increase of the platform before comparison. This eliminated the interference of seasonal market fluctuations and fake transactions, restored the real conversion rate growth trend of the business, accurately identified high-quality businesses in the growth phase, provided an objective basis for guiding the platform's business growth, avoided misjudgments of trends caused by external environmental interference, and ensured that traffic was tilted towards real high-growth businesses.

[0087] Furthermore, the corrected effective number K of users who have consumed services within user u's social circle (b) u,b corr The calculation method is as follows:

[0088]

[0089] The strong ties coefficient ranges from 2 to 3, and the weak ties coefficient ranges from 0.5 to 1, dynamically selected based on the user's relationship type with friends. Strong ties have a higher influence. This parameter can be obtained based on the chat interaction frequency set within the merchant service platform. k is the total number of friends within the social circle who have purchased service b. A j This is the activity indicator for the j-th friend; 1 is assigned if they logged in within the last 30 days, and 0.2 is assigned if they haven't logged in. u,j反映 The interaction intensity between user u and friend j, with chat / like / comment frequencies normalized to 0-1, C j C represents the amount spent by friend j. avg P represents the average spending amount on the platform. j This indicates that friend j is the payer's identifier. If the payment account is not equal to the receiving account, use 1; otherwise, use 0. All of the above parameters can be obtained through the social interaction logs of the merchant service platform.

[0090] By eliminating interference from inactive friends, weak ties, and payment-on-behalf behavior, the influence of strong and weak ties is differentiated, the true influence of social circles on users' consumption decisions is quantified, friend consumption behavior is transformed into actionable recommendation enhancement signals, social trust is used to improve recommendation conversion rates, and ineffective social relationships are avoided from misleading recommendation judgments.

[0091] It should be emphasized that the content in this embodiment can be directly used by a business management platform with catering characteristics. In the existing technology, the management operating system for merchants' offline and online operations is already mature. Most of the data in this embodiment can be obtained through conventional collection methods, such as relevant logs and behavior records. The sources of non-collected items have been identified above.

[0092] After obtaining the final recommended score Final(F) u,b After that, the recommendation scores of all candidate services for the same user are sorted from high to low. The top N services are selected as the final recommendation objects according to the needs of the business scenario. The value of N can be flexibly configured according to the number of display positions of the push channel. For example, 1 service is selected for SMS push and 2-3 services are selected for homepage information flow push.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A personalized discount recommendation system based on big data user consumption and merchant habits, including a merchant service platform, characterized by: It also includes an operations management component and a data processing component. The operations management component obtains relevant operations data from the merchant service platform and sends the collected operations data to the data processing component. The data processing component cleans the operations data. The discount recommendation system also includes an operations identification component. The cleaned operations data is divided into interest-related data, discount-related data, and comprehensive-related data, and then sent to the operations identification component. The operational identification component: Based on interest-related data such as adjusted spending amount, total user spending amount, time preference weight, month-on-month change in spending amount, spending amount in the previous 30 days, smoothing constant, interest decay coefficient, migration interest degree, and migration impact coefficient, the user dynamic interest degree is output, which measures the user's preference for specific services. Based on user dynamic interest and discount-related data such as remaining discount days, maximum activity days, discount perception intensity, maximum perception intensity, inventory weight, corrected effective push count, push frequency limit, push fatigue penalty coefficient, time period matching degree, and time period matching influence coefficient, the user discount recommendation priority is output, and the user discount recommendation priority is used as the recommendation ranking basis for user interests and discount features. The final recommendation score is output based on the base score, the weighted sum of the enhancement factor and the penalty factor, and is used as the final basis for recommendation ranking.

2. The personalized discount recommendation system based on big data user consumption and merchant habits according to claim 1, characterized in that: The operational identification component includes a dynamic interest judgment module, a discount recommendation judgment module, and a final score calculation module.

3. The personalized discount recommendation system based on big data user consumption and merchant habits according to claim 2, characterized in that: The processing logic of the dynamic interest determination module is as follows: Starting from users' historical consumption behavior, the system first corrects for interference factors such as consumption amount and consumption time, then adds the effects of consumption trend changes and cross-business interest migration, and finally outputs the user's comprehensive preference for a single business. The user's comprehensive preference for a single business is used as the user-side basic feature of the entire recommendation system.

4. The personalized discount recommendation system based on big data user consumption and merchant habits according to claim 3, characterized in that: The processing logic of the discount recommendation judgment module is as follows: Based on users' dynamic interests, and with the addition of five adjustment factors—the timeliness and urgency of the offer itself, the user's perceived value of the offer, inventory operation strategy, user push fatigue, and the current time matching degree—the system outputs the real-time recommendation priority of a single offer for a single user, thus serving as an intermediary layer connecting user interests and operational strategies.

5. The personalized discount recommendation system based on big data user consumption and merchant habits according to claim 4, characterized in that: The processing logic of the final score calculation module is as follows: The system adopts a hierarchical weighted architecture. The basic layer retains the core weights of user interest and the timeliness of offers. The enhancement layer adds five positive dimensions: merchant operational needs, user value, market trends, social influence, and user intention. The penalty layer adds the negative dimension of geographical distance. Finally, the system is normalized to a fixed range output, which makes the score calculation take into account recommendation accuracy, operational flexibility, and result stability.

6. The personalized discount recommendation system based on big data user consumption and merchant habits according to claim 5, characterized in that: The dynamic interest judgment module includes: a basic consumption ratio item obtained by dividing the user's adjusted consumption amount in business b by the user's total consumption amount across all businesses; a time preference weight item obtained by dividing the user's peak-hour consumption ratio by the maximum consumption ratio across all time periods; a consumption trend gain item obtained by dividing the difference between the consumption amount in the last 30 days and the consumption amount in the previous 30 days by the consumption amount in the previous 30 days, multiplying by the interest decay coefficient, and adding 1; and a cross-business migration gain item obtained by multiplying the cross-business migration interest degree by the migration impact coefficient and adding 1.

7. The personalized discount recommendation system based on big data user consumption and merchant habits according to claim 6, characterized in that: The discount recommendation judgment module includes: a timeliness factor obtained by subtracting the remaining days of the discount from the maximum number of activity days and dividing by the maximum number of activity days; a discount perception factor obtained by dividing the user's perception intensity of the discount by the platform's maximum perception intensity; an inventory weight item calculated for three scenarios: hot-selling, clearance, and normal; a push fatigue penalty item obtained by dividing the corrected number of effective pushes by the upper limit of push frequency, multiplying by the fatigue penalty coefficient, and subtracting the value from 1; and a time period matching gain item obtained by multiplying the matching degree between the current time and the user's consumption time period by the time period matching influence coefficient and adding 1.

8. The personalized discount recommendation system based on big data user consumption and merchant habits according to claim 7, characterized in that: The final score calculation module includes: normalizing the five enhancement dimensions to intervals, multiplying them by their corresponding weights, and summing them to obtain a weighted sum of enhancement factors; dividing the distance between the user and the service by the maximum service radius, taking the minimum value of the result and 1, and multiplying it by the distance decay weight to obtain a weighted sum of penalty factors; adding the base score, enhancement score, and penalty score, and then using a truncation function to restrict the result to an interval to obtain the normalized final score.

Images