A commodity recommendation optimization method based on big data analysis

Abstract

The present application relates to the technical field of intelligent recommendation, in particular to a commodity recommendation optimization method based on big data analysis, the present application constructs marketing pressure fingerprint by extracting full price reduction gear identification and price difference interval label, and forms a behavior event stream with pressure annotation in association with user behavior events; based on homogeneity pairing analysis of high-pressure users and low-pressure users, marketing pressure insensitive dimensions are identified, and a commodity intrinsic attraction feature library is constructed; combined with marketing activity constraints, the dimensions of commodity features are masked, a recommendation candidate set corresponding to the marketing state is generated, and the standardized hash value is stored in the distributed cache to realize fast response of the recommendation result; at the same time, through the calibration log analysis, the abandonment rate and the single-making failure mode are settled, the marketing pressure interval division threshold is dynamically corrected, and the recommendation model is continuously optimized. The present application can effectively reduce the interference of marketing factors, improve the recommendation accuracy, system response efficiency and self-adaptive optimization ability.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology, specifically to a product recommendation optimization method based on big data analysis. Background Technology

[0002] Online shopping platforms handle numerous merchants, product categories, and high-frequency transactions. To improve conversion rates and user experience, these platforms widely employ recommendation algorithms to provide personalized product recommendations. Existing recommendation systems typically build user-product interaction models based on data such as user historical purchase behavior, product attributes, and user interest tags. These models utilize methods like collaborative filtering, matrix factorization, or deep learning to predict user preferences and rank product recommendations. These methods have achieved some success in relatively stable business environments.

[0003] However, modern online marketplaces face a problem that current technologies have failed to effectively address: platforms continuously run multiple marketing campaigns that change frequently throughout their marketing cycles and are intricately coupled. Users' purchasing decisions are strongly driven by these campaigns, and their behavioral data encompasses both the inherent appeal of the products and the stimulating influence of the marketing activities. Existing recommendation systems, in their design, neglect the impact of the marketing campaign dimension, directly training models based on mixed-contamination user behavior data. This results in recommendation weights learned by the system exhibiting strong marketing campaign specificity; that is, the weights are only effective within a specific marketing mix context.

[0004] When marketing campaigns change, users' purchasing behavior patterns change significantly, but the weight structure of the existing recommendation system remains unchanged, leading to a severe mismatch between recommendation results and the new marketing environment. Specifically, the conversion rate of recommended products under the new marketing mix is ​​far lower than expected, even below historical averages. This phenomenon repeatedly occurs in the platform's daily operations, making it difficult to stabilize the actual performance of the recommendation system. Adjustments to recommendation strategies become a blind trial-and-error process, resulting in wasted marketing resources and fluctuations in user experience. Therefore, a dynamic recommendation optimization method that is marketing context-aware is needed to accurately identify and isolate the impact of marketing campaigns on recommendation effectiveness under conditions of frequent changes in marketing activities.

[0005] To address this, a product recommendation optimization method based on big data analysis is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a product recommendation optimization method based on big data analysis. By constructing a feature library of marketing pressure fingerprints, homogeneous user pairings, and intrinsic product attractiveness, the method eliminates the interference of marketing activities on user behavior. Combined with feature masking, cache preheating, and dynamic calibration mechanisms, it achieves highly accurate, fast-response, and adaptively optimized product recommendations.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A product recommendation optimization method based on big data analysis includes: In the settlement request response chain, extract the full reduction level identifier and price difference range label as marketing pressure fingerprint, and encapsulate them with user behavior events in the same session to form a behavior event stream; Using behavioral event streams as input, homogeneous matching is performed on users in different pressure zones according to their historical category preference vectors to construct a pairing queue. The interaction feature dimension whose difference value in the pairing queue is stable and close to zero is defined as the intrinsic feature dimension, and aggregated to form a product intrinsic attractiveness feature library. Read the intrinsic attractiveness feature library of products, perform feature dimension masking on products that do not meet the constraints according to the constraints, sort the effective products after masking according to the intensity of intrinsic attractiveness, generate a recommendation candidate set, and store it in the distributed cache with the normalized hash value of each marketing status as the index key. When the inventory status changes, the current marketing status hash value is recalculated, the corresponding recommendation candidate set is read from the distributed cache and returned, and the marketing status hash value of this settlement and the product composition in the recommendation candidate set are recorded as a calibration log to trigger the relabeling of the next round of behavior event stream.

[0008] Preferably, the steps for forming a behavioral event flow specifically include: after completing the activity calculation, mapping the currently activated discount tier identifier and the ratio of the difference between the current total price and the next discount tier to the current total price into discrete interval labels, and writing them, along with the session identifier and timestamp, into the extended field of the response message body; when forwarding the response message body, asynchronously reading the extended field and publishing a marketing pressure fingerprint anchor point event containing the discount tier identifier, interval label, session identifier, and timestamp to the message queue; starting from the timestamp, within a preset time window, associating and encapsulating all browsing, adding to cart, and favorite behavioral events generated by the same session with this point event, with each behavioral event carrying a marketing pressure fingerprint composed of the discount tier identifier and the interval label; writing behavioral events carrying high-pressure fingerprints into the high marketing pressure behavioral partition, and writing behavioral events carrying low-pressure fingerprints into the low marketing pressure behavioral partition, with the two partitions together forming a behavioral event flow labeled with pressure; the high-pressure fingerprint corresponds to a first interval where the difference ratio is lower than a first threshold, and the low-pressure fingerprint corresponds to a second interval where the difference ratio is higher than a second threshold.

[0009] Preferably, the steps for constructing the pairing queue specifically include: extracting active user sets from the high-marketing-pressure behavior partition and the low-marketing-pressure behavior partition respectively; reading historical category preference vectors and recent consumption frequency quantiles from user profile data to form multidimensional matching features of users; in the multidimensional matching feature space, using users in the high-pressure user set as query objects, retrieving users in the low-pressure user set that are closest to the features, and all ordered pairs together constitute the pairing queue.

[0010] Preferably, the steps for aggregating and forming a product intrinsic attractiveness feature library specifically include: for each pair of users in the pairing queue, separately calculating the click-through rate, dwell depth distribution, and add-to-cart rate of the same product for high-stress users and low-stress control users, forming feature value pairs for each pair in each interaction feature dimension; for each product, traversing all pairings containing product interaction data, calculating the difference between the feature values ​​of high-stress users and low-stress control users in each dimension, and summarizing to form the difference value distribution of the product in each interaction feature dimension; marking the interaction feature dimensions that have an absolute value of difference below the first stability upper limit, a variance below the second stability upper limit, and satisfy both conditions within the statistical window as marketing pressure insensitive dimensions of the product, and marking the interaction feature dimensions with a positive mean as marketing inducement sensitive dimensions; taking the marketing pressure insensitive dimension of each product as the effective dimension, taking the mean of the pairing feature values ​​of the product in the dimension to form the intrinsic feature vector of the product, and writing it into the product intrinsic attractiveness feature library.

[0011] Preferably, the steps of performing feature dimension masking on products that do not meet the constraints and generating a recommendation candidate set specifically include: reading marketing activity configuration data, extracting the applicable category set, lower and upper bounds of price ranges, and mutual exclusion relationship identifiers between activities for each activity, generating a marketing state combination list by traversing activity combinations and verifying mutual exclusion constraints one by one, where each record contains a constraint parameter set consisting of the category set, price range, and tier identifier of all activities in that state; reading the intrinsic feature vectors of all products in the product intrinsic attractiveness feature library, and performing constraint matching on each product for each marketing state in the list: comparing the product's category identifier and price with the category set and price range of all activities in the marketing state constraint parameter set one by one, and invalidating the values ​​of the activity-related interaction feature dimensions in the product's intrinsic feature vector for products that do not meet any activity constraints; and sorting the intrinsic feature vectors of all products after masking processing from high to low according to the comprehensive score of effective dimensions to generate a recommendation candidate set for the marketing state.

[0012] Preferably, the steps of storing the normalized hash value of each marketing status in the distributed cache as an index key, and recalculating the hash value of the current marketing status and reading the corresponding recommendation candidate set, specifically include: for the marketing status, concatenating the category set identifier, the lower and upper bounds of the price range, and the tier identifier in the constraint parameter set into a normalized string according to the predefined field order, performing a digest calculation on the string to obtain the index key of the marketing status, serializing the ordered list of product identifiers in the recommendation candidate set and writing it into the corresponding storage slot of the distributed cache using the index key to complete the preheating of the candidate set of the marketing status; when the product selection status, inventory status, and / or usage status change, obtaining the category set identifier, the lower and upper bounds of the price range, and the tier identifier of the current effective marketing status combination, concatenating them into a normalized string, and performing a digest calculation to obtain the index key of the current marketing status; If a query is initiated using the index key and a match is found, the ordered list of corresponding product identifiers will be returned as a set of recommended candidates. If no match is found, the intrinsic attractiveness feature library of the product will be read, and the feature dimension masking and effective dimension comprehensive score will be performed in real time to sort the current marketing status. The sorted results will be returned as a set of recommended candidates.

[0013] Preferably, the steps to trigger the next round of behavioral event stream relabeling include: after the user completes the settlement or abandons the settlement, the marketing status index key of this session, the product identifier of each product and the corresponding marketing pressure range label are written into the calibration log; The calibration logs are read in batches, and the number of settlement completions and settlement abandonments under each marketing state are counted by grouping the marketing state index key. The settlement abandonment rate of each marketing state is calculated. For marketing states where the settlement abandonment rate exceeds the preset abandonment rate threshold, the composition of shopping cart items in all abandoned settlement sessions under the state is extracted. The ratio of the number of items with the first interval label to the total number of items in the shopping cart is counted to identify the failure mode of combining orders. The boundary value of the difference ratio distribution that frequently occurs in the failure mode of combining orders is used as the update value and written into the marketing pressure level interval configuration table to replace the original dividing threshold between the first interval and the second interval, i.e., the first threshold and the second threshold, so that the labeling boundary of the marketing pressure fingerprint in the behavior event flow is corrected.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a marketing stress fingerprint by extracting discount tier identifiers and price difference range labels, and establishes homogeneous pairing relationships between high-stress and low-stress users. It further identifies intrinsic feature dimensions with stable differences approaching zero, thereby eliminating behavioral biases caused by marketing factors. By constructing a product intrinsic attractiveness feature library, recommendations are based more on the inherent value of the product than on promotional stimuli, effectively improving the authenticity, stability, and accuracy of recommendation results, reducing the misleading influence of marketing activities on the recommendation model, and enhancing the long-term learning ability of the recommendation system.

[0015] 2. This invention pre-codes marketing state combinations in a standardized manner, generates corresponding recommendation candidate sets through constraint matching and feature masking, and stores these sets in a distributed cache using the standardized hash value of the marketing state as an index key. When inventory changes or shopping cart changes trigger recommendation requests, only the current marketing state hash value needs to be recalculated to quickly hit the corresponding candidate set, eliminating the need to repeat the full recommendation calculation process. Furthermore, for cache misses, the system automatically reverts to a real-time calculation mechanism, balancing response speed and recommendation accuracy, thereby significantly reducing server computational pressure and improving system throughput and user experience.

[0016] 3. This invention establishes a calibration log mechanism to continuously record marketing status index keys, product identifiers, and marketing pressure interval labels, and performs statistical analysis in conjunction with settlement completion rate and abandonment rate. When a high abandonment rate is found in a specific marketing status, the invention further identifies the failure mode of bundling orders and extracts the frequently occurring difference percentage distribution boundary values ​​as the basis for new pressure interval division, dynamically correcting the marketing pressure level configuration table. By forming a closed-loop optimization mechanism of "behavior collection—recommendation generation—result feedback—parameter correction," the marketing pressure labels can continuously closely reflect the real user decision-making process, improving the model's adaptability to changes in user consumption psychology and its long-term evolution capability. Attached Figure Description

[0017] Figure 1 A schematic diagram of a product recommendation optimization method based on big data analysis provided by the present invention; Figure 2 This is a schematic diagram of the process for extracting intrinsic attractiveness features of a commodity provided by the present invention; Figure 3 This is a schematic diagram of the dynamic closed-loop and threshold self-calibration process provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0019] Example 1: Please see Figure 1This invention provides a product recommendation optimization method based on big data analysis. The technical solution is as follows: In the settlement request response chain, the full reduction level identifier and price difference range label are extracted as marketing pressure fingerprints, and associated and encapsulated with user behavior events in the same session to form a behavior event stream; using the behavior event stream as input, users in different pressure partitions are homogeneously matched according to the user's historical category preference vector to construct a pairing queue. The interaction feature dimension with a stable difference value close to zero in the pairing queue is defined as the intrinsic feature dimension, and aggregated to form a product intrinsic attractiveness feature library; the product intrinsic attractiveness feature library is read, and feature dimension masking is performed on products that do not meet the constraints according to the constraints. The effective products after masking are sorted according to the intrinsic attractiveness strength to generate a recommendation candidate set, which is stored in a distributed cache with the normalized hash value of each marketing state as the index key; when the inventory state changes, the current marketing state hash value is recalculated, the corresponding recommendation candidate set is read from the distributed cache and returned; the marketing state hash value of this settlement and the product composition in the recommendation candidate set are recorded as a calibration log to trigger the re-labeling of the next round of behavior event streams.

[0020] Further, the steps for forming a behavioral event flow specifically include: after completing the activity calculation, mapping the currently activated discount tier identifier and the ratio of the difference between the current total price and the next discount tier to the current total price into discrete interval labels, and writing them, along with the session identifier and timestamp, into the extended field of the response message body; when forwarding the response message body, asynchronously reading the extended field and publishing a marketing pressure fingerprint anchor point event containing the discount tier identifier, interval label, session identifier, and timestamp to the message queue; starting from the timestamp, within a preset time window, associating and encapsulating all browsing, adding to cart, and favorite behavioral events generated by the same session with this point event, with each behavioral event carrying a marketing pressure fingerprint composed of the discount tier identifier and interval label; writing behavioral events carrying high-pressure fingerprints into the high marketing pressure behavioral partition, and writing behavioral events carrying low-pressure fingerprints into the low marketing pressure behavioral partition, with the two partitions together forming a behavioral event flow labeled with pressure; the high-pressure fingerprint corresponds to the first interval where the difference ratio is lower than the first threshold, and the low-pressure fingerprint corresponds to the second interval where the difference ratio is higher than the second threshold.

[0021] Specifically, within the marketing computation microservice of the e-commerce platform's shopping cart, a state snapshot interceptor is deployed in the checkout request-response chain. The marketing pressure fingerprint is defined as a tuple consisting of a discount tier identifier and a difference range label, where the difference range label ∈ {high pressure range label, low pressure range label, middle range label}. At the gateway layer, the marketing pressure fingerprint is encapsulated into a marketing pressure fingerprint anchor event and published via a message queue. When the shopping cart page initiates a discount calculation request to the marketing computation service, this interceptor, during the serialization of the returned result, reads the intermediate calculation variables already existing in the service's memory, including the currently matched discount tier identifier and the price difference between the current total price of the shopping cart and the threshold of the next discount tier. It converts the ratio of the price difference to the current total price of the shopping cart into discrete range labels according to predefined tier range mapping rules, and writes both the discount tier identifier and the range labels into an extended field of the response message body. These two pieces of data together constitute the marketing pressure fingerprint, which is transmitted to the API gateway layer along with the response message.

[0022] The boundary thresholds for interval labels, namely the first and second thresholds for dividing high and low marketing pressure, are determined as follows: Upon initial system launch, the system administrator, based on the concentrated occurrence intervals of user bundled purchase behavior in historical activity data, preset the first threshold to 0.15 and the second threshold to 0.40 in the marketing pressure level interval configuration table. Both are ratios of the price difference to the current shopping cart's total price. When this ratio is below 0.15, the user is very close to the next discount threshold and is considered to be in a high marketing pressure state; the corresponding behavioral event is marked with a high-pressure fingerprint. When the ratio is above 0.40, the user is far from the next threshold and is considered to be in a low marketing pressure state; the corresponding behavioral event is marked with a low-pressure fingerprint. Behavioral events with ratios between 0.15 and 0.40 are not included in the current partition calculation.

[0023] The initial thresholds mentioned above were determined by referencing the statistical correlation between add-to-cart behavior and the percentage of difference in historical data: In past activity data, the frequency of add-to-cart behavior was significantly higher in sessions with a difference percentage below 0.15 than the overall average, while the frequency was significantly lower in sessions with a difference percentage above 0.40 than the overall average. Therefore, 0.15 and 0.40 were determined as the initial cutoff points. After the system goes live, the first and second thresholds will be automatically updated iteratively based on actual settlement behavior data through a calibration closed-loop mechanism, requiring no manual intervention. The initial thresholds are only used during the system's cold start phase.

[0024] Deploy a non-intrusive response message parsing filter at the API gateway layer. While forwarding the marketing computation response to the client, this filter asynchronously reads the marketing pressure fingerprint from the extended fields of the response message body. Combining this with the session identifier and current timestamp from the request context, it publishes a marketing pressure fingerprint anchor event to the specified topic in the message queue. The anchor event's message body contains four fields: a discount tier identifier, a difference range label, a session identifier, and a timestamp. The difference range label explicitly records the range category to which the current difference percentage belongs.

[0025] The streaming consumer of the message queue uses the timestamp of the anchor event as a starting point. Within a preset time window, it streams and encapsulates all browsing, adding to cart, and favorites events occurring under the same session identifier with that anchor event. Each event has a marketing pressure fingerprint field attached to its message body, consisting of a discount tier identifier and a difference range label. The preset time window length is initially configured to 60 seconds. This parameter is managed as an adjustable item in the system configuration file. The specific value is determined based on the average shopping cart dwell time of platform users in a single checkout session, to cover the time interval from when most users view the discount calculation result to when they generate the next explicit action event.

[0026] Behavioral events with a difference range label corresponding to a ratio below the first threshold are written to the high marketing pressure behavior partition; behavioral events with a difference range label corresponding to a ratio above the second threshold are written to the low marketing pressure behavior partition. The two partitions together constitute a behavioral event stream labeled with marketing pressure.

[0027] Furthermore, the steps for constructing the pairing queue specifically include: extracting active user sets from the high-marketing-pressure behavior partition and the low-marketing-pressure behavior partition respectively; reading historical category preference vectors and recent consumption frequency quantiles from user profile data to form multidimensional matching features of users; in the multidimensional matching feature space, using users in the high-pressure user set as query objects, and retrieving users in the low-pressure user set that are closest to the features; all ordered pairs together constitute the pairing queue.

[0028] Further, see Figure 2The specific steps for aggregating and forming a product intrinsic attractiveness feature library include: for each pair of users in the pairing queue, calculating the click-through rate, dwell depth distribution, and add-to-cart rate of the same product for high-stress users and low-stress control users respectively, forming feature value pairs for each pair in each interaction feature dimension; for each product, traversing all pairings containing product interaction data, calculating the difference between the feature values ​​of high-stress users and low-stress control users in each dimension, and summarizing to form the difference value distribution of the product in each interaction feature dimension; marking the interaction feature dimensions that meet the two conditions within the statistical window, with the absolute value of the difference value below the first stability upper limit and the variance below the second stability upper limit, as marketing pressure insensitive dimensions of the product, and marking the interaction feature dimensions with a positive mean as marketing inducement sensitive dimensions; taking the marketing pressure insensitive dimension of each product as the effective dimension, taking the mean of the pairing feature values ​​of the product in the dimension to form the intrinsic feature vector of the product, and writing it into the product intrinsic attractiveness feature library.

[0029] Specifically, taking the stream of behavioral events labeled with pressure as input, a homogeneous queue matching job is built in the streaming computing engine. This job extracts sets of active users who have generated behavioral events within a fixed time window from both the high-marketing-pressure behavior partition and the low-marketing-pressure behavior partition. The default length of the time window is 1 hour, which can be adjusted in the system configuration according to the actual traffic volume.

[0030] For each user in both sets, the task reads the following two types of matching features from the user profile service: First, a historical category preference vector, where each component represents the normalized weight of the user's historical interaction behavior in the corresponding primary product category, with values ​​ranging from 0 to 1 (real numbers), and the sum of all components is 1. The vector dimension is consistent with the number of primary product categories on the platform. Second, a recent consumption frequency quantile, which is the percentile ranking of the user's transaction count in the last 30 calendar days among all active users on the platform, with values ​​ranging from 0 to 100 (integers). These two types of features are concatenated to form a multi-dimensional matching feature vector for each user.

[0031] The task employs Locality Sensitive Hashing (LSH) to perform nearest neighbor retrieval in a multidimensional matching feature space, using Euclidean distance as the similarity metric. Each user in the high-marketing-pressure user set is used as the query object. The task retrieves the user in the low-marketing-pressure user set whose multidimensional matching feature vector has the closest Euclidean distance, forming an ordered pair between a high-pressure user and a low-pressure user. If a high-pressure user cannot find a corresponding user in the low-pressure set that meets the maximum Euclidean distance constraint, that high-pressure user is not included in the pairing queue to ensure the statistical comparability of paired users. The initial value of the maximum Euclidean distance constraint is determined based on the distribution range of user category preference vectors in historical data and is managed as an adjustable parameter in the system configuration. All valid ordered pairs together constitute the pairing queue for the current time window.

[0032] For each product in the matching queue, the system statistically analyzes the behavioral observations of high-stress users and low-stress control users on the following three interaction feature dimensions: click-through rate (the proportion of actual clicks when the product is exposed), dwell depth distribution (the product details page is divided into 10 equal height intervals, and the proportion of times users scroll to each interval is recorded to form a 10-component frequency vector), and add-to-cart rate.

[0033] For each interaction feature dimension of each product, the job iterates through all pairs containing interaction data for that product, calculates the difference between the feature values ​​of high-stress users and low-stress control users for each pair, and summarizes the difference values ​​of all pairs under the same product and the same feature dimension to form the difference value distribution for that dimension. The difference value for the dwell depth distribution dimension is the difference vector between the corresponding components of two 10-component frequency vectors, with its mean and variance calculated separately for each component and then averaged.

[0034] The stability of the mean and variance of the variance distribution is assessed separately: If the absolute value of the mean variance of a certain interactive feature dimension is lower than the first stability upper limit (initial value 0.05) and the variance is lower than the second stability upper limit (initial value 0.01), and both conditions are met for N consecutive statistical windows (N≥3, default N=3), then the dimension is marked as a marketing pressure-insensitive dimension; if the mean variance continuously exceeds the positive value of the first stability upper limit and the above conditions are met for N consecutive statistical windows, then the dimension is marked as a marketing inducement-sensitive dimension. The first stability upper limit (initial value 0.05) and the second stability upper limit (initial value 0.01) are the initial system configuration values, determined based on the natural fluctuation range of the variance distribution of different feature dimensions in historical data, and are uniformly managed in the system configuration file. They can be adjusted according to changes in the platform's data scale.

[0035] Using the interaction feature dimension that each product is marked as insensitive to marketing pressure as the effective dimension, for each effective dimension, the arithmetic mean of the feature values ​​of all paired high-pressure users and low-pressure control users for that product in that dimension is taken. Then, the arithmetic mean of all paired arithmetic means is calculated to obtain the intrinsic feature value of the product in that effective dimension. The intrinsic feature values ​​of all effective dimensions together constitute the intrinsic feature vector of the product.

[0036] The system records the product identifier, the list of valid dimensions, and the corresponding list of intrinsic feature values ​​as a single record in the product intrinsic attractiveness feature library. For products with fewer than 30 accumulated pairs, they are not added to the feature library until the number of pairs meets this lower limit, triggering supplementary calculations to ensure that the intrinsic feature vectors have sufficient statistical stability.

[0037] Furthermore, the steps of masking features for products that do not meet the constraints and generating a candidate set of recommendations specifically include: reading marketing campaign configuration data, extracting the applicable category set, lower and upper bounds of price ranges, and mutual exclusion relationship identifiers between activities for each activity; generating a marketing state combination list by traversing the activity combinations and verifying mutual exclusion constraints one by one; each record in the list contains a constraint parameter set consisting of the category set, price range, and tier identifiers of all activities in that state; reading the intrinsic feature vectors of all products in the product intrinsic attractiveness feature library; performing constraint matching for each product in each marketing state in the list: comparing the product's category identifier and price with the category set and price range of all activities in the marketing state constraint parameter set one by one; for products that do not meet any activity constraints, invalidating the values ​​of the activity-related interaction feature dimensions in the product's intrinsic feature vector; and sorting the intrinsic feature vectors of all products after masking processing from high to low according to the comprehensive score of effective dimensions to generate a candidate set of recommendations for marketing states.

[0038] Furthermore, the steps of storing the normalized hash value of each marketing status into the distributed cache as an index key, and recalculating the hash value of the current marketing status and reading the corresponding recommendation candidate set, specifically include: for the marketing status, concatenating the category set identifier, the lower and upper bounds of the price range, and the tier identifier in the constraint parameter set into a normalized string according to the predefined field order, performing a digest calculation on the string to obtain the index key of the marketing status, serializing the ordered list of product identifiers in the recommendation candidate set and writing it into the corresponding storage slot of the distributed cache using the index key to complete the preheating of the candidate set of the marketing status; when the product selection status, inventory status, and / or usage status change, obtaining the category set identifier, the lower and upper bounds of the price range, and the tier identifier of the current effective marketing status combination, concatenating them into a normalized string, and performing a digest calculation to obtain the index key of the current marketing status; If a query is initiated using the index key and a match is found, the ordered list of corresponding product identifiers will be returned as a set of recommended candidates. If no match is found, the intrinsic attractiveness feature library of the product will be read, and the feature dimension masking and effective dimension comprehensive score will be performed in real time to sort the current marketing status. The sorted results will be returned as a set of recommended candidates.

[0039] Specifically, the offline marketing status analysis module is automatically triggered after the operations staff publishes the daily marketing campaign schedule. This module reads the configuration data of all valid campaigns within the current operating cycle from the marketing campaign configuration database and extracts the constraint parameters for each campaign: the set of applicable product categories (represented by a list of product category identifiers), the lower and upper bounds of the price range, the tier identifier (the full reduction rule number), and the mutual exclusion relationship identifier with other campaigns.

[0040] This module iterates through all subsets of activities and verifies mutual exclusion constraints one by one (i.e., checks whether the mutual exclusion identifiers of any two activities within a combination contain each other). The verified subsets of activities are then written as valid marketing status records to the marketing status combination list. Each record in the list contains a set of constraint parameters for all activities in that status, namely a list of category set identifiers, a list of lower and upper bounds for price ranges, and a list of price tier identifiers.

[0041] The candidate set pre-construction engine reads the intrinsic feature vectors of all products from the product intrinsic attractiveness feature library. For each valid marketing state in the marketing state combination list, the engine performs constraint matching for each product: comparing the product's category identifier and price with the applicable category set and price range of all activities in the constraint parameter set for that marketing state. If the product's category identifier is not in the applicable category set of any activity, or the product's price is not within the price range of any activity, then the feature value of the activity-related interaction feature dimension in the product's intrinsic feature vector, as specified by the system configuration, is set to zero, i.e., a feature dimension masking operation is performed, while the feature values ​​of the remaining valid dimensions remain unchanged. The activity-related interaction feature dimension is specified as follows: in the marketing activity configuration data, each type of activity is labeled with its corresponding set of interaction feature dimensions. The masking engine reads this configuration when performing constraint matching and performs the zeroing operation on the corresponding dimension for products that do not meet the conditions.

[0042] After the masking operation is completed, all products in the current marketing state are sorted from highest to lowest based on their comprehensive scores across effective dimensions. The comprehensive score is calculated as follows: the feature value of each effective dimension in the product's intrinsic feature vector is multiplied by its corresponding weight coefficient, and the weighted results of all effective dimensions are summed to obtain the product's comprehensive score in the current marketing state. The weight coefficients for each dimension are determined through regression analysis based on the contribution of each interaction feature dimension to the final add-to-cart behavior in historical conversion data during the initial system configuration. These weight coefficients are stored in a weight configuration table and can be updated in the configuration with system iterations without requiring a regeneration of the feature library. After sorting, the ordered list of the top N product identifiers by comprehensive score is taken as the recommendation candidate set for this marketing state. The default value for N is 50, which can be adjusted in the system configuration.

[0043] For each valid marketing status, the category set identifier list (arranged in ascending order by identifier lexicographical order and concatenated into a substring separated by commas), the lower bound of the price range, the upper bound of the price range, and the price tier identifier in its constraint parameter set are concatenated into a normalized string according to the above predefined field order and separated by vertical bars. The normalized string is then subjected to SHA-256 digest calculation, and the first 32 hexadecimal characters of the digest result are taken as the index key of the marketing status.

[0044] After serializing the ordered list of product identifiers in the recommended candidate set into a JSON array format, the result is written to a hash slot using the index key and the serialization result. An expiration time is set to the current marketing campaign's effective deadline plus an 8-hour buffer. Once all candidate sets of valid marketing states are written, the engine registers a cache expiration flag with the state management service, indicating that the current cached content is consistent with the current marketing schedule.

[0045] When the shopping cart server detects changes in the product selection status, inventory status, or coupon usage status, it calls the marketing calculation service to obtain the latest valid marketing status combination for the current shopping cart. Following the same standardized concatenation rules as in the pre-building phase, it generates an index key for the current marketing status and sends a Get request to the Redis cluster using this index key. If the cache is hit, the ordered list of corresponding product identifiers is returned directly to the recommendation response body as a recommendation candidate set. If not, the server reads the product's intrinsic attractiveness feature library, performs real-time feature dimension masking and effective dimension comprehensive score ranking on the current marketing status, returns the ranking result as a recommendation candidate set, and asynchronously triggers the pre-building engine to perform supplementary cache preheating for that marketing status.

[0046] Further, see Figure 3 The steps that trigger the next round of behavioral event flow relabeling include: after the user completes the settlement or abandons the settlement, the marketing status index key of this session, the product identifier of each product and the corresponding marketing pressure range label are written into the calibration log. The calibration logs are read in batches, and the number of settlement completions and settlement abandonments under each marketing state are counted, using the marketing state index key as the grouping dimension. The settlement abandonment rate for each marketing state is calculated. For marketing states where the settlement abandonment rate exceeds the preset abandonment rate threshold, the composition of the shopping cart items in all abandoned settlement sessions under the state is extracted. The ratio of the number of items carrying the first interval tag to the total number of items in the shopping cart is counted to identify the order-combination failure mode. The boundary value of the difference ratio distribution that frequently occurs in the order-combination failure mode is used as the update value and written into the marketing pressure level interval configuration table to replace the original division threshold between the first interval and the second interval, i.e., the first threshold and the second threshold. The step of writing into the marketing pressure level interval configuration table also simultaneously triggers the invalidation of the storage slot corresponding to the marketing state index key in the distributed cache, so that when the recommendation request is received next time, the recommendation candidate set is regenerated according to the updated first threshold and the second threshold.

[0047] Specifically, after a user completes checkout or abandons checkout, the shopping cart service writes the following data for this session into the calibration log: the marketing status index key at the time of checkout, the product identifier of each item in the shopping cart and the corresponding marketing pressure range label, and the user's final behavior result (checkout completed or checkout abandoned). This data together constitutes a calibration log record, written to the calibration log database using the marketing status index key and timestamp as the combined primary key.

[0048] The calibration log consumption service reads calibration logs in daily batches. Using the marketing status index key as the grouping dimension, it counts the number of settlement completions and settlement abandonments for each marketing status, and calculates the settlement abandonment rate for each marketing status (number of settlement abandonments divided by the total number of settlements).

[0049] For marketing campaigns where the checkout abandonment rate exceeds a preset threshold, the consumer service extracts the composition of the shopping cart items from all abandoned checkout sessions in that campaign. It then calculates the ratio of the number of items with a price difference range label corresponding to a value lower than the current first threshold to the total number of items in the shopping cart. If this ratio exceeds 0.50, the marketing campaign is marked as exhibiting a "buy-to- ...

[0050] The initial value for the preset abandonment rate threshold is configured as 0.30. This means that if more than 30% of checkout transactions in a certain marketing scenario are ultimately abandoned by users, it is considered that there is a systemic mismatch between the recommended candidate set and the actual needs of users in that marketing scenario, requiring further analysis. This initial value is determined based on the average abandonment rate in historical normal checkout scenarios plus one standard deviation, and can be updated in the configuration during system operation based on the statistical distribution of the platform's overall checkout data.

[0051] The consumer service extracts the percentage of the difference in the shopping cart items from all session data marked as "failed to reach the minimum purchase amount" mode, and calculates the quantile distribution of this percentage across all failed sessions. The 25th percentile of this distribution is used as a new candidate value for the first threshold, and the 75th percentile as a new candidate value for the second threshold. The arithmetic mean of the candidate values ​​obtained from three consecutive daily batches is calculated, and this mean is used as the final updated value, written into the marketing pressure level configuration table, replacing the original first and second thresholds.

[0052] After the configuration table is updated, the storage slots corresponding to all affected marketing status index keys in the distributed cache are simultaneously invalidated. This ensures that the next time a recommendation request is received, feature dimension masking and candidate set sorting are re-performed based on the updated first and second thresholds, guaranteeing that the cached content is consistent with the latest thresholds. In the next batch of behavioral event processing, consumers in the behavioral event stream automatically read the latest first and second thresholds from the configuration table and re-execute the write judgment for high-marketing-pressure behavior partitions or low-marketing-pressure behavior partitions based on the updated interval boundaries, thus correcting the labeling boundaries of the marketing pressure fingerprint in the behavioral event stream. The updated thresholds are also synchronously written to the running configuration of the API gateway layer filter, ensuring that the interval label mapping rules published in subsequent marketing pressure fingerprint anchor events remain consistent, completing the system's end-to-end self-calibration closed loop.

[0053] Example 2: This embodiment, based on Embodiment 1, focuses on the cross-session pressure attenuation weighting process, the lightweight re-ranking process, and the fine-grained attribution process at the category dimension. The specific implementation process is as follows: The settlement request response chain extracts the full-discount tier identifier and the difference range label, forming a behavioral event stream with marketing pressure fingerprints, which is written to the high-marketing-pressure behavioral partition and the low-marketing-pressure behavioral partition. When the total price of the shopping cart has reached or exceeded the highest full-discount tier, the difference percentage is recorded as zero and a full-tier range label is assigned. The corresponding behavioral event is written to the full-tier behavioral partition and does not participate in subsequent matching queue construction. Based on the user's historical category preference vector, recent purchase frequency quantiles, and price sensitivity quantiles, homogeneous matching is performed between active users in the high-pressure and low-pressure partitions to construct a matching queue. After performing cross-session pressure attenuation weighting on the matching queue, it is aggregated to form a product intrinsic attractiveness feature library. The feature library is read, and feature dimension masking is performed on products that do not meet the constraints, generating a recommendation candidate set. Each product identifier and its effective dimension comprehensive score form a key-value pair, which is then arranged in descending order of score and serialized, and written to the distributed cache using the marketing status index key. The candidate set is returned after being filtered by real-time product status and then overlaid with a lightweight re-ranking. The calibration log batch processing performs fine-grained attribution at the category level, and the attribution results are fed back to the score calculation stage, triggering the corresponding marketing status cache invalidation.

[0054] Furthermore, before aggregating to form the intrinsic attractiveness feature library of goods, the process includes a step of performing cross-session pressure attenuation weighting on each pair in the pairing queue, specifically including: Extract the historical session records of each paired user within the current statistical period from the behavioral event stream, and count the cumulative number of exposure sessions for each user under the same discount tier, as the user's repeated exposure count under the current marketing pressure. Using the repeated exposure count as the independent variable, the cross-session pressure attenuation coefficient is calculated according to a preset attenuation function. The attenuation function satisfies the following: it decreases monotonically with the repeated exposure count, and the attenuation coefficient takes values ​​in the range of [δ, 1], where δ is a preset lower bound for attenuation and 0 < δ < 1. When the repeated exposure count is 1, the attenuation coefficient is 1. When calculating the difference in feature values ​​between high-stress users and low-stress control users on a dimension-by-dimensional basis, the feature values ​​of high-stress users are multiplied by their corresponding cross-session stress attenuation coefficient before being included in the difference value calculation. This makes the contribution weight of user behavior data that has been repeatedly exposed to the same marketing pressure over a long period of time to the intrinsic feature vector decrease as the number of exposures increases. The original difference value distribution is replaced with the attenuated weighted difference value distribution, and the subsequent intrinsic feature dimension labeling and intrinsic feature vector generation steps are performed.

[0055] Specifically, after the pairing queue is built but before the difference value calculation is executed, cross-session pressure attenuation weighting is performed on high-pressure users in each pair. From the persistent storage of the behavior event stream, using the user identifier and the full / reduction tier identifier as the joint query key, the number of sessions in which the user has appeared under the same full / reduction tier identifier within the current statistical period (the statistical period length is a preset time length, the initial value is set to a natural month, and it is adjusted in the system configuration according to the system data accumulation) is retrieved, and recorded as the user's duplicate exposure count.

[0056] For users with a repeat exposure count of 1, the decay coefficient is set to 1, and the feature value is used in the difference value calculation with full weight. For users with a repeat exposure count greater than 1, the decay coefficient decreases monotonically as the count increases: for each increase in the count, the decay coefficient is multiplied by a fixed ratio, which is a preset value greater than zero and less than one. The fixed ratio is obtained as follows: in historical data, the difference in add-to-cart rate across each interaction feature dimension is statistically analyzed between the user groups with repeat exposure counts of 1, 2, 3, 4, and 5 and the low-stress control group. The folding ratio fitting value with the largest decrease in absolute value between the third and fifth exposures is used as the initial value of the fixed ratio, and it is updated quarterly in the same way as the calibration logs accumulate. The reasonable range of the fixed ratio should ensure that the decay coefficient only reaches the lower bound of decay within at least five repeat exposures to ensure that the decay curve has sufficient discriminative power.

[0057] A lower bound is set for the attenuation coefficient. When the calculated attenuation coefficient is lower than the lower bound, the lower bound value is used instead. The lower bound is obtained as follows: In historical data, the absolute value of the mean difference between high-frequency users with repeat exposure times above the 95th percentile and low-stress control users across each interaction feature dimension is statistically analyzed. The minimum attenuation coefficient value is selected as the initial value of the lower bound, ensuring that the proportion of dimensions whose mean absolute difference value exceeds one-tenth of the standard deviation of the entire sample for that dimension does not exceed 5%. This lower bound is updated in the same way on a quarterly basis.

[0058] When calculating the difference value dimension by dimension, for each pairing, the feature value of the high-stress user in a certain interaction feature dimension is multiplied by its corresponding cross-session stress attenuation coefficient, and then the difference is taken with the feature value of the low-stress control user in the same dimension to obtain the attenuated weighted difference value. After traversing all pairs in the pairing queue, the attenuated weighted difference value distribution of each product in each dimension is summarized.

[0059] The statistical stability is determined as follows: After the cumulative number of paired samples for a product in the pairing queue reaches a preset minimum sample size, the moving mean and moving variance of the difference value sequence of that product in a certain dimension are calculated with a sliding step of 10 pairs. The minimum sample size is obtained by taking the number of pairs that can be matched for a product with the 25th percentile of exposure on the platform within the statistical period, rounding this number down to the nearest multiple of 10 as the initial value, and updating it quarterly as the platform's data scale changes. When the absolute value of the mean of three consecutive sliding windows does not exceed one-tenth of the standard deviation of the difference values ​​of the entire sample in that dimension, and the variance of three consecutive sliding windows does not exceed one-fifth of the variance of the entire sample, the dimension is deemed to meet the statistical stability condition and is marked as a marketing pressure-insensitive dimension. For products with a cumulative number of paired samples less than the minimum sample size, the intrinsic feature dimension marking is not performed; they are recorded as products to be accumulated, and the determination will be re-evaluated after data is supplemented in subsequent statistical periods. The above determination is performed using the attenuated weighted difference value distribution.

[0060] Furthermore, after the distributed cache hits and the recommendation candidate set is returned after real-time state filtering, the process also includes a step of performing a lightweight reordering of the candidate set based on the user's real-time session characteristics, specifically including: Extract real-time session feature vectors from the current user session context. The real-time session feature vectors include the category distribution vector of the products in the current shopping cart, the difference vector between the total price of the current shopping cart and the threshold of each discount tier, and the category identifier sliding window of the N most recent products in the browsing sequence in the current session. For each product in the recommended candidate set, a session matching score is calculated between the intrinsic feature vector of the product and the real-time session feature vector. The session matching score is obtained by weighted summation based on the complementarity between the product category and the existing category distribution in the shopping cart, the filling efficiency of the product price on the current difference vector, and the category co-occurrence frequency of the product category and the browsing sequence sliding window. The intrinsic attractiveness score and conversation matching score of each product in the candidate set are linearly weighted according to preset weights to generate a personalized re-ranking score. The candidate set is then re-ranked from high to low according to the personalized re-ranking score and returned. The lightweight reordering process only reads the stored intrinsic feature vectors and real-time session context, without re-accessing the full data of the product intrinsic attractiveness feature library, and without triggering new feature dimension masking calculations.

[0061] Specifically, when the distributed cache hits the marketing status index key and returns the recommended candidate set after being filtered by the real-time product status, a lightweight reordering is immediately performed synchronously in the response assembly stage of the settlement service. No full data read request is initiated for the product intrinsic attractiveness feature library. The intrinsic attractiveness score of each product is directly obtained from the product identifier and effective dimension comprehensive score key-value pair read from the cache.

[0062] The real-time session feature vector consists of three parts. The first part is the shopping cart category distribution vector: iterates through the category identifiers of the selected items in the current shopping cart, calculates the ratio of the number of items in each category to the total number of items in the shopping cart, forming a fixed-length vector. The length of the vector is equal to the total number of categories on the platform, and the component value of categories that do not appear is zero. The second part is the difference filling vector: for the current total price of the shopping cart, calculate the difference with each discount threshold in turn. The difference value corresponding to the threshold with a positive difference (i.e., the threshold has not yet been reached) is taken as this component, and the component with a negative or zero difference (the threshold has been reached or exceeded) is set to zero. When all components of the difference filling vector are zero, it indicates that the shopping cart has reached all thresholds, the difference filling efficiency score is recorded as zero, and it does not participate in the subsequent score weighting. The third part is the browsing sequence sliding window: it takes the category identifier sequence of the N browsing events with the most recent timestamp in the current session. N is a preset positive integer. The method of obtaining it is: take the 50th percentile value of the browsing depth distribution of the platform's historical sessions, round it up to the nearest integer as the initial value of N, and update it quarterly according to the platform's user browsing behavior data.

[0063] For each product in the candidate set, a session matching score is calculated: The complementarity score is calculated by subtracting the corresponding component value of the product category in the shopping cart category distribution vector from one; the difference filling efficiency score is calculated by the ratio of the number of positive components whose prices do not exceed the total number of positive components in the difference filling vector (this score is zero if the difference filling vector has no positive components); and the browsing co-occurrence score is calculated by dividing the number of times the product category identifier appears in the browsing sequence sliding window by N. The session matching score is obtained by multiplying each of the three scores by a preset weight and summing them. All three weights are preset values ​​greater than zero, and their sum is one. The initial weight values ​​are obtained as follows: In the historical calibration log, with the settlement completion rate as the optimization objective, a grid search is performed on the three weights within their respective value ranges. The weight combination with the highest settlement completion rate is used as the initial value, and subsequently updated monthly in the same way as the calibration log accumulates.

[0064] The intrinsic attractiveness score and conversation matching score of each item in the candidate set are multiplied by preset sovereign weight and subordinate weight, respectively, and then summed to obtain the personalized re-ranking score. The sovereign weight is not less than 0.5 to ensure the dominant position of the global intrinsic ranking, and the sum of the sovereign weight and subordinate weight is one. The candidate set is re-ranked from high to low according to the personalized re-ranking score and then returned to the front-end display module.

[0065] Furthermore, after identifying the failure patterns of bundled purchases, the process also includes a step of fine-grained attribution of these failure patterns by marketing campaign category dimension, specifically including: From the calibration log of marketing status where the checkout abandonment rate exceeds the preset abandonment rate threshold, the shopping cart items are grouped according to the product category identifier, and the concentration of the number of items with the first interval label in each category is calculated as a percentage of the total number of items in that category. Categories with a concentration of combined purchases exceeding a preset category concentration threshold are marked as high-risk categories for combined purchase failures. The combination of the corresponding discount tier identifier and the category identifier is written into the category tier failure record table. In the effective dimension comprehensive score calculation stage after the feature dimension masking step, the category level failure record table is read. For products that hit the failure record, their effective dimension comprehensive score is multiplied by a preset discount coefficient. The discount coefficient ranges from (0,1). This reduces the ranking score of products in high-risk categories of bundled purchase failures in the current marketing state without affecting the stored value of their intrinsic feature vector, thus preserving the product's ability to participate in recommendations in non-bundled purchase trigger scenarios. Write the update timestamp of the category and price tier expiration record table to the expiration flag key of the distributed cache, trigger the cache expiration of the corresponding marketing status index key, so that the feature dimension masking and sorting calculations are re-executed in the next query, and the discount adjustment takes effect.

[0066] Specifically, after identifying the failed order-pushing pattern during calibration log batch processing, fine-grained category-level attribution is performed on marketing states where the settlement abandonment rate exceeds the abandonment rate threshold. The abandonment rate threshold is obtained by using the 75th percentile of the platform's historical overall settlement abandonment rate at the initial stage of system launch as the initial value, and updating the threshold with the 75th percentile of the settlement abandonment rate distribution of all marketing states within the sliding window during each calibration batch processing.

[0067] For calibration logs that hit the invalid marketing status, extract all abandoned checkout sessions (excluding sessions with payment failure event records), group them by the category identifier of the items in the shopping cart, and count the ratio of the number of items with the first interval tag under each category to the total number of items in the shopping cart of all invalid sessions for that category, as the category's intra-category order concentration.

[0068] The method for obtaining the concentration threshold of order-combining within a category is as follows: the initial value is the 80th percentile of the concentration distribution of order-combining within all categories in the current batch, and the average of the 80th percentiles of the distribution in the most recent three batches is taken as the current effective threshold to smooth the fluctuation of data in a single batch.

[0069] Categories with a concentration of add-on orders exceeding a threshold are marked as high-risk categories for add-on order failures. The combination of the category identifier and the corresponding discount tier identifier is written into the category tier failure record table, and the record is timestamped.

[0070] The category tier failure record table performs a validity check during each batch processing: For each record in the table, the current activation status of the corresponding discount tier identifier is read. If the tier has been deactivated, the record is deleted directly. For records that are still active, it is checked whether the concentration of orders within the corresponding category in the two most recent calibration batches has been lower than the category concentration threshold. If so, the record is marked as pending recovery. If it is still lower than the threshold in the next batch, it is deleted from the failure record table, and the corresponding marketing status cache is invalidated.

[0071] In the effective dimension comprehensive score calculation stage after feature dimension masking, the category tier failure record table is read. For products whose category identifier and full reduction tier identifier match the failure record, their effective dimension comprehensive score is multiplied by the discount coefficient. The discount coefficient is a preset value greater than zero and less than one. The initial value of the discount coefficient is obtained as follows: In the historical calibration log, with the settlement completion rate of products that match the failure record as the optimization target, a grid search with a step size of 0.1 is performed on the discount coefficient within the range of greater than zero and less than one, and the discount coefficient value that results in the highest settlement completion rate is used as the initial value. The discount coefficient is dynamically adjusted as follows: In each calibration batch after the discount takes effect, the settlement completion rate of products that match the failure record is calculated. If the settlement completion rate increases relative to the average of the two batches before the discount takes effect, the current discount coefficient is maintained. If it does not increase for two consecutive batches, the discount coefficient is reduced by a preset step size until the settlement completion rate increases or the discount coefficient reaches a preset lower bound. If the settlement completion rate is higher than the level before the discount takes effect for two consecutive batches after the discount takes effect, the discount coefficient is increased by a preset step size until the discount coefficient reaches a preset upper bound. The above operations only apply to the score calculation stage and do not modify the intrinsic feature vectors stored in the product's intrinsic attractiveness feature library, thus preserving the product's ability to participate in recommendations normally under other marketing conditions.

[0072] After the category-level invalidation record table is written or updated, the timestamp of this operation is written to the invalidation flag key of the corresponding marketing status index key in the distributed cache, triggering the invalidation of the marketing status cache entry. This causes the feature dimension masking and effective dimension comprehensive score calculation to be re-executed in the next query, so that the discount adjustment takes effect in the next recommendation response.

[0073] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A product recommendation optimization method based on big data analysis, characterized in that, include: In the settlement request response chain, extract the full reduction level identifier and price difference range label as marketing pressure fingerprint, and encapsulate them with user behavior events in the same session to form a behavior event stream; Using behavioral event streams as input, homogeneous matching is performed on users in different pressure zones according to their historical category preference vectors to construct a pairing queue. The interaction feature dimension whose difference value in the pairing queue is stable and close to zero is defined as the intrinsic feature dimension, and aggregated to form a product intrinsic attractiveness feature library. Read the intrinsic attractiveness feature library of products, perform feature dimension masking on products that do not meet the constraints according to the constraints, sort the effective products after masking according to the intensity of intrinsic attractiveness, generate a recommendation candidate set, and store it in the distributed cache with the normalized hash value of each marketing status as the index key. When the inventory status changes, the current marketing status hash value is recalculated, the corresponding recommendation candidate set is read from the distributed cache and returned, and the marketing status hash value of this settlement and the product composition in the recommendation candidate set are recorded as a calibration log to trigger the relabeling of the next round of behavior event stream.

2. The product recommendation optimization method based on big data analysis according to claim 1, characterized in that: The steps for forming a behavioral event flow specifically include: after completing the activity calculation, mapping the currently activated discount tier identifier and the ratio of the difference between the current total price and the next discount tier to the current total price into discrete interval labels, and writing them, along with the session identifier and timestamp, into the extended field of the response message body; when forwarding the response message body, asynchronously reading the extended field and publishing a marketing pressure fingerprint anchor event containing the discount tier identifier, interval label, session identifier, and timestamp to the message queue; starting from the timestamp, within a preset time window, associating and encapsulating all browsing, adding to cart, and favorite behavioral events generated in the same session with this point event, with each behavioral event carrying a marketing pressure fingerprint composed of the discount tier identifier and interval label; writing behavioral events carrying high-pressure fingerprints into the high marketing pressure behavioral partition, and writing behavioral events carrying low-pressure fingerprints into the low marketing pressure behavioral partition, with the two partitions together forming a behavioral event flow labeled with pressure; the high-pressure fingerprint corresponds to the first interval where the difference ratio is lower than a first threshold, and the low-pressure fingerprint corresponds to the second interval where the difference ratio is higher than a second threshold.

3. The product recommendation optimization method based on big data analysis according to claim 2, characterized in that: The specific steps for constructing the pairing queue include: extracting active user sets from the high-marketing-pressure behavior partition and the low-marketing-pressure behavior partition respectively; reading historical category preference vectors and recent consumption frequency quantiles from user profile data to form multidimensional matching features of users; in the multidimensional matching feature space, using users in the high-pressure user set as query objects, and retrieving users in the low-pressure user set that are closest to the features; all ordered pairs together constitute the pairing queue.

4. The product recommendation optimization method based on big data analysis according to claim 3, characterized in that: The specific steps for aggregating and forming a product intrinsic attractiveness feature library include: for each pair of users in the pairing queue, calculating the click-through rate, dwell depth distribution, and add-to-cart rate of the same product for high-stress users and low-stress control users respectively, forming feature value pairs for each pair across each interaction feature dimension; for each product, iterating through all pairings containing product interaction data, calculating the difference between the feature values ​​of high-stress users and low-stress control users dimension by dimension, and summarizing to form the difference value distribution of the product across each interaction feature dimension; marking interaction feature dimensions whose absolute value of the difference value is lower than the first stability upper limit, whose variance is lower than the second stability upper limit, and which meet both conditions within the statistical window as marketing pressure insensitive dimensions of the product, and marking interaction feature dimensions with a positive mean as marketing inducement sensitive dimensions; using the marketing pressure insensitive dimension of each product as the effective dimension, taking the mean of the pairing feature values ​​of the product across the dimension to form the product's intrinsic feature vector, and writing it into the product intrinsic attractiveness feature library.

5. The product recommendation optimization method based on big data analysis according to claim 4, characterized in that: The specific steps for masking features of products that do not meet constraints and generating a candidate set of recommendations include: reading marketing campaign configuration data, extracting the applicable category set, lower and upper bounds of price ranges, and mutual exclusion relationship identifiers between activities for each activity; generating a marketing state combination list by traversing activity combinations and verifying mutual exclusion constraints one by one; each record in the list contains a constraint parameter set consisting of the category set, price range, and tier identifiers of all activities under the marketing state; reading the intrinsic feature vectors of all products in the intrinsic attractiveness feature library; performing constraint matching for each product under each marketing state in the list: comparing the product's category identifier and price with the category set and price range of all activities in the marketing state constraint parameter set one by one; for products that do not meet any activity constraints, invalidating the values ​​of the activity-related interaction feature dimensions in the product's intrinsic feature vector; and sorting the intrinsic feature vectors of all products after masking processing from high to low according to the comprehensive score of effective dimensions to generate a candidate set of recommendations for the marketing state.

6. The product recommendation optimization method based on big data analysis according to claim 5, characterized in that: The steps of storing the normalized hash value of each marketing status into the distributed cache as an index key, and recalculating the hash value of the current marketing status and reading the corresponding recommendation candidate set, specifically include: for each marketing status, concatenating the category set identifier, lower and upper bounds of the price range, and grade identifier in the constraint parameter set into a normalized string according to the predefined field order; performing a digest calculation on the string to obtain the index key of the marketing status; serializing the ordered list of product identifiers in the recommendation candidate set and writing it into the corresponding storage slot of the distributed cache using the index key to complete the preheating of the candidate set for the marketing status; when the product selection status, inventory status, and / or usage status change, obtaining the category set identifier, lower and upper bounds of the price range, and grade identifier of the current effective marketing status combination, concatenating them into a normalized string, and performing a digest calculation to obtain the index key of the current marketing status; If a query is initiated using the index key and a match is found, the ordered list of corresponding product identifiers will be returned as a recommendation candidate set. If no match is found, the intrinsic attractiveness feature library of the product will be read, and the feature dimension masking and effective dimension comprehensive score will be performed in real time to sort the current marketing status. The sorted results will be returned as a recommendation candidate set.

7. The product recommendation optimization method based on big data analysis according to claim 1, characterized in that: The steps to trigger the next round of behavioral event flow relabeling include: after the user completes the settlement or abandons the settlement, the marketing status index key of this session, the product identifier of each product and the corresponding marketing pressure range label are written into the calibration log. The calibration logs are read in batches, and the number of settlement completions and settlement abandonments under each marketing state are counted by grouping the marketing state index key. The settlement abandonment rate of each marketing state is calculated. For marketing states where the settlement abandonment rate exceeds the preset abandonment rate threshold, the composition of shopping cart items in all abandoned settlement sessions under the state is extracted. The ratio of the number of items with the first interval label to the total number of items in the shopping cart is counted to identify the failure mode of combining orders. The boundary value of the difference ratio distribution that frequently occurs in the failure mode of combining orders is used as the update value and written into the marketing pressure level interval configuration table to replace the original dividing threshold between the first interval and the second interval, i.e., the first threshold and the second threshold, so that the labeling boundary of the marketing pressure fingerprint in the behavior event flow is corrected.

Images