User analysis method and system based on big data, product and medium

Abstract

A user analysis method and system based on big data, a product and a medium relate to the field of commodity data processing, and the method comprises the following steps: obtaining a specific operation sequence of a current user in a current full-decreasing sales promotion scene, extracting an order-on decision behavior feature in the specific operation sequence according to a preset user behavior analysis rule, and obtaining a decision-on decision behavior feature of the order-on decision behavior feature; the order-on decision behavior feature comprises an abnormal shopping behavior; on the basis of the make-on decision behavior characteristics, calculating a make-on possibility score for each commodity purchased by the user, determining a recommendation weight of each commodity according to the make-on possibility score, and when the make-on possibility score of a target commodity exceeds a preset score threshold value, determining that the target commodity is a recommended commodity; and reducing the recommendation weight of the target commodity in the current full-reduction sales promotion scene to a preset weight. By implementing the method, the accuracy of the e-commerce platform to push commodities to the user can be improved.

Description

Technical Field

[0001] The present application relates to the field of commodity data processing, and in particular to a user analysis method, system, product and medium based on big data. Background Art

[0002] With the development of mobile Internet technology, e-commerce platforms have invested a lot of resources in analyzing user consumption behavior. Through in-depth analysis of user behavior data, e-commerce platforms can better understand user needs. Among them, promotional activities, as an important marketing tool, can have a certain impact on users' shopping decisions.

[0003] In related technologies, an analysis system based on user historical behavior can be used to process user data in a full-discount scenario. The analysis system is based on user historical behavior data and, by establishing a user tag system, classifies and compiles user historical purchase records, browsing tracks and other data to form a fixed user portrait.

[0004] However, during the full-discount promotion, users' shopping decisions are often dynamically affected by promotion rules, and their behavior patterns are quite different from historical data, resulting in the recommendation results generated by the system being unable to meet the actual needs of users. Summary of the invention

[0005] The present application provides a user analysis method, system, product and medium based on big data, which are used to improve the accuracy of product push to users by e-commerce platforms.

[0006] In a first aspect, the present application provides a user analysis method based on big data, which is applied to a user analysis system, and the method includes: obtaining a specific operation sequence of the current user in a current full-reduction promotion scenario, wherein the current full-reduction promotion scenario is an e-commerce promotion rule including a full-reduction threshold amount and a discount amount for goods, and the specific operation sequence includes the time point when the user enters the full-reduction activity page, a record of changes in the amount of goods in the shopping cart, a path for switching between product categories, a record of adding, deleting, and modifying goods, and a browsing order for goods; extracting a purchase-matching decision behavior feature in the specific operation sequence according to a preset user behavior analysis rule, and the purchase-matching decision behavior feature includes abnormal shopping behavior; based on the purchase-matching decision behavior feature, calculating a purchase-matching possibility score for each product added by the user, and the purchase-matching possibility score is calculated based on the amount difference between the product and the full-reduction threshold when the product is added, the user operation frequency when the product is added, and the similarity between the product and the user's historical purchasing behavior; determining a recommendation weight for each product according to the purchase-matching possibility score, and when the purchase-matching possibility score of the target product exceeds a preset score threshold, reducing the recommendation weight of the target product in the current full-reduction promotion scenario to a preset weight.

[0007] By adopting the above technical solution, the specific operation sequence of users in the full-discount promotion scenario is obtained, including entry time, changes in shopping cart amount and other information. It can comprehensively record user behavior, extract the behavioral characteristics of decision-making to make up the order, especially abnormal shopping behavior, and accurately grasp the special operation mode generated by users to make up the order. Based on this, the possibility score of making up the order is calculated and the recommendation weight is determined. When the score exceeds the threshold, the recommendation weight is reduced to avoid excessive recommendation of make-up items, so that the recommendation is more in line with the real needs of users, improve the accuracy and effectiveness of user recommendations, and optimize the user shopping experience.

[0008] In combination with some embodiments of the first aspect, in some embodiments, the step of extracting the characteristics of the decision-making behavior of making up the order in the specific operation sequence according to the preset user behavior analysis rules specifically includes: obtaining the operation time sequence of the current user in the specific operation sequence after entering the full-reduction activity page; calculating the time interval between adjacent operations in the operation time sequence; when it is detected that the time interval between multiple consecutive operations in the operation segment is less than the preset time threshold, marking the operation segment as a high-frequency operation segment; analyzing the change records of the shopping cart item amount in the high-frequency operation segment, when the change direction of the shopping cart item amount undergoes multiple reverse adjustments and the adjustment amount is close to the full-reduction threshold amount, it is determined to be an abnormal shopping behavior; based on the current user's switching path between product categories in the high-frequency operation segment, the category switching distance of the current user between adjacent operations is counted, when the category switching distance is greater than the preset distance threshold and the number of switches exceeds the preset number threshold, it is determined to be an abnormal shopping behavior; integrating the time characteristics, shopping cart amount change characteristics and category switching characteristics of the high-frequency operation segment to obtain the characteristics of the current user's decision-making behavior of making up the order.

[0009] By adopting the above technical solution, the operation time series is obtained and the time intervals between adjacent operations are calculated, and high-frequency operation segments can be found. The changes in the shopping cart amount and the category switching path in the high-frequency operation segments are analyzed, and abnormal shopping behaviors are accurately determined by judging the reverse adjustment of the amount, the distance and number of category switching, etc. The relevant time, amount and category characteristics are used as the behavior characteristics of the decision-making behavior of adding to the order, which helps the system to more accurately identify the user's intention to add to the order, so as to make more reasonable decisions when recommending products and improve the accuracy and pertinence of recommendations.

[0010] In combination with some embodiments of the first aspect, in some embodiments, the step of calculating the possibility score of adding to the order for each item added by the user based on the decision-making behavior characteristics of adding to the order specifically includes: obtaining the total amount of the shopping cart when the current user adds each item, calculating the difference between the total amount of the shopping cart and the threshold amount for full discount, and calculating the first sub-item score based on the difference; obtaining the operation frequency within a preset time period when the current user adds each item, and calculating the second sub-item score based on the operation frequency; obtaining the current user's historical purchase list of items, calculating the category similarity and price range similarity between the current added item and the historically purchased items, and calculating the third sub-item score based on the category similarity and the price range similarity; weighted summing the first sub-item score, the second sub-item score and the third sub-item score according to preset weights to obtain a possibility score of adding to the order for each item.

[0011] By adopting the above technical solution, the difference between the total amount of the shopping cart and the threshold amount is calculated to obtain the first sub-item score, the operation frequency within the preset time period is obtained to calculate the second sub-item score, and the similarity between the current added goods and the historical purchased goods is calculated to obtain the third sub-item score to measure the compliance of the goods with user needs. The preset weighted sum is used to obtain the possibility score of completing the order, which improves the rationality and effectiveness of product recommendations and meets user shopping expectations.

[0012] In combination with some embodiments of the first aspect, in some embodiments, the step of determining the recommendation weight of each of the products based on the possibility score of adding to the order specifically includes: obtaining the score ranges and corresponding weight adjustment coefficients of preset multiple levels of possibility of adding to the order; matching the possibility score of each product with the score range, and determining the corresponding weight adjustment coefficient; obtaining the original recommendation weight of each of the products, multiplying the original recommendation weight by the weight adjustment coefficient, and obtaining the recommendation weight of each of the products.

[0013] By adopting the above technical solution, the recommendation weight distribution is made more reasonable, the recommendation list is more in line with the actual needs of users, the adaptability of the recommendation system to user behavior is improved, the recommended products can more accurately meet the shopping expectations of users in the full-discount scenario, and the shopping guidance effect is optimized.

[0014] In combination with some embodiments of the first aspect, in some embodiments, after the step of reducing the recommendation weight of the target product in the current full-reduction promotion scenario to a preset weight when the possibility score of the target product exceeds a preset score threshold, the method also includes: extracting the static features of the current user in the full-reduction activity based on a convolutional neural network, the static features including product category preferences and price sensitivity; extracting the temporal features of the current user during shopping cart operations based on a recurrent neural network RNN, the temporal features including a continuity pattern of adding, deleting and modifying products; and sorting the product recommendation list according to the static features and the temporal features.

[0015] By adopting the above technical solutions, the static features of the full-reduction activities, such as product category preferences and price sensitivity, can be extracted based on the convolutional neural network, which can deeply explore the long-term and stable shopping tendencies of users. The recurrent neural network RNN ​​extracts the temporal features of the shopping cart operation, such as the continuity mode of adding, deleting and modifying products, to capture the dynamic behavior changes of users. The static and temporal features are input into the preset fusion model, and the purchase tendency probability is obtained through comprehensive analysis. The optimal recommended price range is determined in combination with price sensitivity, so as to accurately grasp the user's needs, make the product recommendation more targeted, and improve the matching degree between the recommended products and user expectations.

[0016] In combination with some embodiments of the first aspect, in some embodiments, the step of sorting the product recommendation list according to the static feature and the time series feature specifically includes: inputting the static feature and the time series feature into a preset fusion model to obtain the purchase tendency probability of the current user for different product categories; determining the optimal recommended price range in combination with the price sensitivity feature of the current user; and sorting the product recommendation list according to the purchase tendency probability and the optimal recommended price range.

[0017] By adopting the above technical solution, static features and time series features are input into the fusion model to obtain the purchase tendency probability, which reflects the user's purchase possibility of different categories of goods. The optimal recommended price range is determined in combination with price sensitivity. The user's price acceptance range is comprehensively considered, and the product recommendation list is sorted according to the purchase tendency probability and the optimal price range. Products that users are more likely to buy and are reasonably priced are recommended first, making the recommendation order more scientific and reasonable, allowing users to find their favorite products more easily, improving user shopping efficiency, and enhancing the effectiveness of the recommendation system in guiding user shopping decisions. The recommendation results are more in line with the actual needs of users and optimize the user shopping experience.

[0018] In combination with some embodiments of the first aspect, in some embodiments, after the step of sorting the product recommendation list according to the static feature and the time series feature, the method also includes: monitoring the sorted product recommendation list in real time to obtain monitoring results; and adjusting the weight coefficient of the static feature to a preset weight threshold based on the monitoring results.

[0019] By adopting the above technical solution, the sorted product recommendation list is monitored in real time, and the interaction data between users and recommended products can be comprehensively collected, including key indicators such as click-through rate, dwell time, add-to-cart rate and purchase rate. The weight coefficient of static features is adjusted based on the monitoring results, so that the recommendation system can timely optimize the importance of features according to real-time user feedback. For example, if the click-through rate of a certain type of product is low, the weight of the relevant static features is reduced to make the recommendation more in line with the user's current needs, thereby improving the accuracy and effectiveness of the recommendation and enhancing the user's shopping experience.

[0020] In a second aspect, an embodiment of the present application provides a user analysis system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the user analysis system to perform the method described in the first aspect and any possible implementation method of the first aspect.

[0021] In a third aspect, an embodiment of the present application provides a computer program product comprising instructions, which, when the computer program product is run on a user analysis system, enables the user analysis system to perform the method described in the first aspect and any possible implementation method of the first aspect.

[0022] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, comprising instructions, which, when executed on a user analysis system, enable the user analysis system to execute the method described in the first aspect and any possible implementation of the first aspect.

[0023] It can be understood that the user analysis system provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the method provided in the embodiment of the present application. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding method, which will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. This application can comprehensively record user behavior and extract the characteristics of decision-making behavior for making up an order, especially abnormal shopping behavior, by obtaining the user's specific operation sequence in the full-discount promotion scenario, including entry time, changes in shopping cart amount and other information. It can accurately grasp the special operation mode generated by users to make up an order, and based on this, calculate the possibility score of making up an order and determine the recommendation weight. When the score exceeds the threshold, the recommendation weight is reduced to avoid excessive recommendation of make-up items, so that the recommendation is more in line with the user's real needs, improve the accuracy and effectiveness of user recommendations, and optimize the user shopping experience.

[0025] 2. This application can discover high-frequency operation segments by obtaining the operation time series and calculating the time intervals between adjacent operations. By analyzing the changes in the shopping cart amount and the category switching path in the high-frequency operation segments, the abnormal shopping behavior can be accurately determined by judging the reverse adjustment of the amount, the distance and number of category switching, etc., and the relevant time, amount and category characteristics can be used as the behavior characteristics of the decision-making behavior of adding to the order, which will help the system to more accurately identify the user's intention to add to the order, so as to make more reasonable decisions when recommending products and improve the accuracy and pertinence of the recommendations.

[0026] 3. This application can comprehensively collect the interaction data between users and recommended products, including key indicators such as click-through rate, dwell time, add-to-cart rate and purchase rate, by monitoring the sorted product recommendation list in real time. The weight coefficient of static features is adjusted based on the monitoring results, so that the recommendation system can timely optimize the importance of features according to real-time user feedback. For example, if the click-through rate of a certain type of product is low, the weight of the relevant static features can be reduced to make the recommendation more in line with the user's current needs, thereby improving the accuracy and effectiveness of the recommendation and enhancing the user's shopping experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flowchart of a user analysis method based on big data in an embodiment of the present application; Figure 2 is another flowchart of the user analysis method based on big data in an embodiment of the present application; Figure 3 It is a schematic diagram of the structure of a physical device of the user analysis system in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to be used as limitations to the present application. As used in the specification of the present application, the singular expressions "one", "a kind of", "above", "the" and "this" are intended to also include plural expressions, unless there is a clear indication to the contrary in the context. It should also be understood that the term "and / or" used in the present application refers to any or all possible combinations comprising one or more of the listed items.

[0029] In the following, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as suggesting or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features, and in the description of the embodiments of the present application, unless otherwise specified, "plurality" means two or more.

[0030] For ease of understanding, the application scenarios of the embodiments of the present application are introduced below.

[0031] A large e-commerce platform launched a "50 off for orders over 300" promotion during the "Double Eleven" shopping festival. Through data analysis, it was found that more than 40% of users showed obvious hesitation when the amount in the shopping cart was close to 300 yuan: repeatedly adding and deleting products, quickly switching between different categories to browse, and staying on a certain product detail page but delaying making a decision. According to conventional logic, the platform recommendation system prioritizes high-profit products and products that users often browse but do not buy. However, data shows that the conversion rate of these recommended products is only 2%, far lower than the usual average of 8%. In-depth analysis found that users' shopping decisions at this time are more influenced by the "completing the order" mentality. They need products with just the right amount of money to reach the full discount threshold, rather than high-value products recommended by the system. This mismatch between recommendations and users' actual needs leads to a decline in user experience, a prolonged shopping process, and ultimately the possibility of abandoning purchases.

[0032] An e-commerce platform previously used a recommendation system based on collaborative filtering to handle the full-reduction scenario. The system mainly builds user portraits based on the user's historical purchase records, browsing tracks, and collection data, and then recommends products that similar users have purchased. For example, when the shopping cart amount of a user who frequently purchases beauty products reaches 280 yuan, the system will continue to recommend high-end skin care products that the user frequently browses, with a price range of 200-500 yuan. However, the actual user's demand is to find a small item of about 20 yuan to make up the order. Because the system cannot recognize this temporary "make up the order" demand, users have to manually search for suitable products and screen the appropriate price range among many products. This process usually takes 3-5 minutes. What's worse is that when the user finally finds the right product, he may find that some of the products in the previous shopping cart have been sold out, and he has to start the screening process again.

[0033] The e-commerce platform has introduced an intelligent recommendation system based on user behavior analysis. When the system detects that the amount in the user's shopping cart is 285 yuan, it immediately starts the behavior analysis of buying together. Through real-time monitoring, it was found that the user performed 5 consecutive addition and deletion operations within 30 seconds, and switched back and forth between the "personal care and cleaning" and "snacks" categories. The system determined that this was a typical decision-making behavior of buying together. Based on this judgment, the system quickly adjusted the recommendation strategy, giving priority to displaying products with a price range of 15-20 yuan, and ranking similar products that the user has purchased in the past two months. At the same time, the system will pay special attention to the inventory status of the products to ensure that the recommended products are currently available for purchase. This precise recommendation enabled the user to find a toothpaste for 17.9 yuan in 10 seconds, which not only met the demand for buying together, but also was a category that users often use.

[0034] For ease of understanding, the following describes the process of the method provided by this implementation in combination with the above scenario. Figure 1, which is a flow chart of a user analysis method based on big data in an embodiment of the present application.

[0035] S101. Obtain the specific operation sequence of the current user in the current full-discount promotion scenario. The current full-discount promotion scenario is an e-commerce promotion rule that includes a full-discount threshold amount and a discount amount for goods. The specific operation sequence includes the time point when the user enters the full-discount activity page, a record of changes in the amount of goods in the shopping cart, a path for switching between product categories, a record of adding, deleting, and modifying goods, and a product browsing order.

[0036] Among them, the specific operation sequence refers to the ordered set of all interactive behaviors performed by users on the full-reduction activity page in chronological order. The behavior characteristics of the decision-making behavior of making up an order refer to the data characteristics of the specific behavior pattern exhibited by the user in order to reach the full-reduction threshold. Abnormal shopping behavior refers to the operation behavior that is significantly different from the user's normal shopping pattern. The operation time interval refers to the time difference between two adjacent operations. The time threshold is used to determine the time standard for high-frequency operations. High-frequency operation segments refer to a series of operations performed intensively in a short period of time. The shopping cart amount change record refers to the change in the total amount of the shopping cart with the user's operation. The product category switching path represents the browsing jump trajectory of the user between different product categories.

[0037] This step is triggered when the user enters the full-reduction activity page and starts shopping, and feature extraction is performed continuously throughout the shopping process. Specifically, the system first obtains the user's operation time series and calculates the time interval between adjacent operations. When it is found that the time interval between multiple consecutive operations is less than the preset time threshold (such as 2 seconds), this operation sequence is marked as a high-frequency operation segment. Then, the changes in the shopping cart amount within the high-frequency operation segment are analyzed, with special attention paid to the situation where the direction of the amount change is reversed and the adjusted amount is close to the full-reduction threshold. At the same time, statistics are collected on the user's switching of product categories within the high-frequency operation segment. When the category switching distance is large and the switching is frequent, it is determined to be an abnormal shopping behavior. Finally, these feature data are integrated to form the user's order-matching decision behavior feature vector.

[0038] In some embodiments, the extraction of the behavioral features of the decision to buy a small item can be achieved in a variety of ways: optionally, the sliding window method is used to scan the operation sequence, the window size is set to 30 seconds, and each slide is 10 seconds. The time interval distribution characteristics of the operations in the window are calculated, and the intensive operation segments with time intervals less than the threshold are identified. The shopping cart amount change trend and commodity category switching mode of these operation segments are analyzed, and the features that meet the abnormal conditions are recorded; optionally, a cluster analysis method is used to divide the user's operation sequence into multiple segments according to the time dimension, and the characteristic values ​​such as the operation frequency, amount change amplitude, and category switching distance of each segment are calculated. The K-means algorithm is used to cluster these features, and the segments that are far away from the normal shopping behavior cluster center are marked as abnormal behaviors. It can be understood that other feature extraction and anomaly detection methods can also be used to realize the identification of the behavioral features of the decision to buy a small item, which is not limited here.

[0039] S102: extracting a behavior feature of a purchase-adding decision in the specific operation sequence, wherein the behavior feature of the purchase-adding decision includes abnormal shopping behavior.

[0040] Among them, the score for the possibility of adding items to a purchase is a quantitative evaluation of the probability that a product is used for the purpose of adding items to a purchase, and the value range is 0-100. The amount difference indicates the difference between the total amount of the shopping cart and the threshold amount for full discount when adding items to a purchase. The user operation frequency refers to the number of times a user performs shopping operations within a preset time period. The similarity of historical purchase behavior is used to measure the degree of match between the current added items and the user's previously purchased items in terms of category attributes and price characteristics. The first sub-item score indicates the score calculated based on the amount difference. The second sub-item score indicates the score calculated based on the operation frequency. The third sub-item score indicates the score calculated based on historical similarity. The preset weight is used to control the importance ratio of each sub-item score in the final score.

[0041] This step is triggered when the user adds an item to the shopping cart, and the possibility score of each newly added item is calculated. Specifically, the system first obtains the total amount of the shopping cart when the item is added to the cart, calculates the difference with the full-reduction threshold, and determines the first sub-item score based on the difference. Then, the number of operations performed by the user within one minute before adding the item to the cart is counted, the operation frequency value is calculated, and the second sub-item score is determined. Then, the category and price characteristics of the current item are compared with the user's historical purchases, and the similarity value is calculated as the third sub-item score. Finally, the three sub-item scores are weighted and summed according to the preset weights to obtain the possibility score of the item.

[0042] In some embodiments, the calculation of the score of the possibility of making up the order can be achieved in a variety of ways: optionally, a piecewise function is used to calculate the score of the first sub-item, when the amount difference is within 10 yuan, the score is 90-100 points, 10-30 yuan is 70-89 points, 30-50 yuan is 50-69 points, and more than 50 yuan is 0-49 points; the scoring rules are designed according to the operation frequency, the frequency is greater than 1 time / second, the score is 90-100 points, 0.5-1 times / second is 70-89 points, and so on; the historical similarity score is calculated based on cosine similarity, and finally the weighted sum is performed according to the weight ratio of 0.4:0.3:0.3; optionally, a machine learning model is used for score prediction, and the features such as the amount difference, operation frequency, and historical similarity are input into the trained regression model, and the score of the possibility of making up the order is directly output. The model is regularly updated with the latest user behavior data. It can be understood that other scoring mechanisms or prediction methods can also be used to achieve quantitative evaluation of the possibility of making up the order, which is not limited here.

[0043] S103. Based on the behavior characteristics of the purchase decision, a purchase possibility score is calculated for each product added by the user. The purchase possibility score is calculated based on the difference in amount between the purchase price and the discount threshold when the product is added, the user's operation frequency when the product is added, and the similarity between the product and the user's historical purchase behavior.

[0044] Among them, the behavior characteristics of the decision-making behavior of adding items to a purchase represent the characteristics of the shopping behavior patterns that users exhibit in order to reach the threshold for full discounts. The probability score of adding items to a purchase refers to a quantitative measure of the probability that a product is used for the purpose of adding items to a purchase, and the value range is 0-100. The amount difference represents the absolute difference between the total amount of the shopping cart and the threshold amount for full discounts when adding items to a purchase. The operation frequency is used to represent the ratio of the number of operations of a user within a specified time window to the time. The similarity of historical purchase behavior refers to the degree of match between the current added items and the user's historical purchase records in terms of product category and price characteristics. The first sub-item score represents the single score value calculated based on the amount difference. The second sub-item score represents the single score value calculated based on the operation frequency. The third sub-item score represents the single score value calculated based on historical similarity. The weighting coefficient is used to control the proportion of each sub-item score in the final score.

[0045] This step is triggered when the user adds the product to the shopping cart. The system needs to calculate the possibility score of each added product in real time. Specifically, the system first obtains the total amount of the shopping cart when the product is added to the cart, and calculates the difference with the full discount threshold. When the difference is small (such as less than 30 yuan), the first sub-item score is higher. Then count the number of operations performed by the user within the specified time window (such as 60 seconds) before adding the product to the cart, and calculate the operation frequency value. The higher the frequency, the higher the second sub-item score. Then calculate the category similarity and price range similarity between the current product and the user's historical purchases, and comprehensively obtain the third sub-item score. Finally, the three sub-item scores are weighted and summed according to the preset weights (such as 0.4, 0.3, 0.3) to obtain the final possibility score of adding the product to the cart.

[0046] In some embodiments, the calculation of the score of the possibility of making up the order can be achieved in a variety of ways: Optionally, a rule-based scoring method is used: First, a scoring rule for the amount difference is designed. A difference of less than 10 yuan is scored 90-100 points, 10-30 yuan is scored 70-89 points, and so on; then a scoring rule for the operation frequency is designed. A frequency greater than 1 time / second is scored 90-100 points, and 0.5-1 times / second is scored 70-89 points; finally, the historical similarity score is calculated, and the matching degree of the product feature vector is calculated using cosine similarity, and weighted summation is performed according to preset weights; Optionally, a machine learning model is used for prediction: First, a feature vector is constructed, including dimensions such as amount difference, operation frequency, and historical similarity; then a trained neural network model is used, which includes multiple hidden layers and activation functions; finally, the possibility score of making up the order is directly output through the model, and the model parameters are updated regularly with new data. It can be understood that other scoring algorithms or prediction methods can also be used to achieve quantitative evaluation of the possibility of making up the order, which is not limited here.

[0047] S104. Determine the recommendation weight of each product according to the possibility score of adding to the order. When the possibility score of adding to the order of a target product exceeds a preset score threshold, reduce the recommendation weight of the target product in the current full-discount promotion scenario to a preset weight.

[0048] Among them, the recommendation weight represents the numerical indicator that determines the ranking position of the product in the recommendation list. The preset score threshold refers to the standard score used to determine whether a product has a strong tendency to add to the order. The target product refers to the specific product currently being evaluated. The current full-reduction promotion scenario refers to a promotional activity environment that includes a specific full-reduction threshold and discount amount. The preset weight represents the minimum recommended weight value pre-defined by the system. The score interval represents the multiple sub-intervals into which the range of 0-100 points is divided. The weight adjustment coefficient refers to the weight correction parameter corresponding to each score interval. The original recommendation weight represents the basic weight value of the product before the judgment of adding to the order is made.

[0049] This step is executed immediately after the calculation of the probability score of the product to be added to the order is completed, and is used to dynamically adjust the display priority of the product. Specifically, the system first sets multiple score intervals for the probability of adding to the order, such as 90-100 points for the extremely high probability interval, 70-89 points for the high probability interval, 50-69 points for the medium probability interval, and 0-49 points for the low probability interval. Then configure the corresponding weight adjustment coefficient for each interval, such as 0.3, 0.5, 0.7, and 1.0. Then obtain the original recommendation weight of the product, multiply it by the adjustment coefficient of the corresponding interval, and get the final recommendation weight. When the probability score of the product to be added to the order exceeds the preset threshold (such as 90 points), its recommendation weight is directly reduced to the preset minimum level (such as 30% of the original weight).

[0050] In some embodiments, the adjustment of the product recommendation weight can be achieved in a variety of ways: Optionally, a piecewise linear adjustment method is used: first, a weight mapping function is constructed to map the order possibility score to a weight adjustment coefficient of 0-1; then the original weight of the product is obtained, including a comprehensive score of sales volume, score, profit margin and other dimensions; finally, the original weight is multiplied by the adjustment coefficient to obtain the final recommendation weight; Optionally, a dynamic weight adjustment strategy is adopted: first, the trigger conditions for weight adjustment are set, including score thresholds and time windows; then the changes in the order possibility of the product are monitored in real time; finally, the weight attenuation rate is dynamically adjusted according to the change trend to achieve a smooth weight transition. It is understandable that other weight calculation and adjustment methods can also be used to achieve a reasonable allocation of product display priorities, which are not limited here.

[0051] The following is a more detailed description of the process of the method provided by this implementation. Figure 2 , is another flow chart of the user analysis method based on big data in an embodiment of the present application.

[0052] S201. Obtain the specific operation sequence of the current user in the current full-discount promotion scenario. The current full-discount promotion scenario is an e-commerce promotion rule that includes a full-discount threshold amount and a discount amount for goods. The specific operation sequence includes the time point when the user enters the full-discount activity page, the change record of the amount of goods in the shopping cart, the switching path between product categories, the operation record of adding, deleting and modifying goods, and the order of browsing goods.

[0053] The current full-reduction promotion scenario is a common marketing rule in e-commerce platforms. Specifically, it refers to a promotion activity in which a user can get a corresponding discount amount (such as 50 yuan off) when the total amount of goods in the shopping cart reaches a preset threshold amount (such as 300 yuan). The specific operation sequence refers to the user's entire behavior record during the full-reduction activity, including the following key elements: the time point of entering the activity page (timestamp accurate to seconds), the shopping cart amount change record (the total amount of the shopping cart after each addition or deletion of goods), the product category switching path (the user's browsing trajectory between different product categories), the product addition, deletion and modification operation record (adding products to the shopping cart, deleting products from the shopping cart, and modifying the quantity of products) and the product browsing order (the order in which users view product detail pages).

[0054] In the specific implementation, the system first records all the user's operation behaviors through the user behavior log collection module. For visits to the full-reduction activity page, the user's login ID, access timestamp, device information and other basic data are recorded. When the user performs a shopping cart operation, the operation type (add / delete / modify), operation time, product ID, product quantity, and the total amount of the shopping cart before and after the operation are recorded. For product browsing behavior, the product ID viewed by the user, the category to which the product belongs, the browsing time, the browsing order and other information are recorded. All of these data are organized into a complete operation sequence in chronological order. Each record contains fields such as operation time, operation type, operation object, and operation result, forming a standardized data structure for subsequent analysis and processing.

[0055] S202: Obtain the operation time sequence after the current user enters the full-discount activity page in the specific operation sequence.

[0056] The operation time series is the time dimension data extracted from the specific operation sequence. The specific operation sequence contains all the user's behavior records in the full-reduction activity, while the operation time series focuses on the time point information of these behaviors, recording the time when all operations of the user occurred from entering the activity page to the end of the shopping process. These time points are arranged in chronological order to form an orderly time series, and each time point corresponds to a specific user operation.

[0057] The system extracts time information from the complete operation sequence through the data processing module. First, determine the starting time point when the user enters the full-reduction activity page and use it as the starting point of the time series. Then extract the timestamp of each subsequent operation of the user in chronological order, including the time of browsing products, adding shopping carts, modifying product quantities, deleting products, etc. These time information are stored in a unified time format (such as Unix timestamp), and each time point is associated with the behavior data in the original operation sequence, which is convenient for subsequent analysis of the user's operation frequency and behavior pattern.

[0058] S203: Calculate the time intervals between adjacent operations in the operation time sequence.

[0059] The time interval refers to the time difference between two adjacent operations in the operation time series. This time difference reflects the user's pause or thinking time between two operations and is an important indicator for analyzing user behavior characteristics. Adjacent operations refer to the two operations that are arranged in chronological order in the operation time series. They may be operations of the same type (such as adding products twice in a row) or operations of different types (such as adding products to the shopping cart after browsing products).

[0060] During the specific calculation, the system processes each pair of adjacent time points in the operation time series. The timestamp of the latter operation is subtracted from the timestamp of the previous operation to obtain the time difference in milliseconds. These time interval data form a new sequence that records the rhythm changes of user operations. For example, if the user performs the first operation at 10:00:00 and the second operation at 10:00:03, the time interval between the two operations is 3000 milliseconds. The system stores these time interval data in an array or list for subsequent judgment of whether the user has high-frequency operation behavior.

[0061] S204: When it is detected that the time interval between multiple consecutive operations in an operation segment is less than a preset time threshold, mark the operation segment as a high-frequency operation segment.

[0062] An operation segment refers to a set of continuous operation behaviors of a user in a full-discount activity. The time interval refers to the time difference between two adjacent operations, measured in milliseconds. The preset time threshold is a benchmark time set based on normal user operation behavior statistics, for example, it is set to 2000 milliseconds. A high-frequency operation segment refers to a sequence of behaviors in which a user performs intensive operations in a short period of time. The operation interval in this sequence is significantly lower than the normal user's operation rhythm. For example, if a user adds or deletes an item 10 times in a row within 5 seconds, and the interval between each operation is less than 500 milliseconds, this operation segment is a high-frequency operation segment.

[0063] During the implementation process, the system first sets a time window (such as 10 seconds) as the observation unit and analyzes the operation sequence within this time window. The system reads the time interval data in the operation time series and uses the sliding window method to check each operation segment in turn. When it is found that the operation intervals for three or more consecutive times are less than the preset time threshold (such as 2000 milliseconds), the system marks this operation sequence as a high-frequency operation segment. The marking process includes recording the start time, end time, number of operations included, and average operation interval of the segment. For the marked high-frequency operation segments, the system also saves all the operation details in the segment to provide data support for subsequent behavior analysis.

[0064] S205: Analyze the change records of the shopping cart item amounts in the high-frequency operation segment. When the change direction of the shopping cart item amount undergoes multiple reverse adjustments and the adjusted amount is close to the full-discount threshold amount, it is determined to be an abnormal shopping behavior.

[0065] The change record of the shopping cart item amount refers to the change of the total amount of the shopping cart due to user operations, including the amount before the change, the amount after the change, and the difference in change. The direction of change refers to the increase or decrease trend of the shopping cart amount, with an increase being a positive change and a decrease being a negative change. Reverse adjustment refers to the user's alternating positive and negative changes to the shopping cart amount in a short period of time, such as increasing first and then decreasing, or increasing again and then decreasing. The threshold amount for full discount is the target amount set for the promotion, such as 300 yuan in "50 yuan off for purchases over 300 yuan". Abnormal shopping behavior refers to the operation mode in which the user approaches the threshold for full discount by frequently adjusting the items in the shopping cart.

[0066] The system first extracts all records involving changes in the amount of the shopping cart in the high-frequency operation segments, arranges these records in chronological order, and calculates the amount change value and direction caused by each operation. The system analyzes the amount change sequence and determines it as abnormal shopping behavior when the following situations are detected: within a single high-frequency operation segment (such as within 10 seconds), the direction of the change of the shopping cart amount has at least 4 reverse adjustments (such as increase-decrease-increase-decrease), and the difference between the total amount of the shopping cart and the threshold amount at any time does not exceed the preset range (such as 20 yuan). In specific implementation, the system records the amount change sequence through an array, uses a counter to count the number of direction changes, and calculates the difference with the threshold amount in real time. When the above conditions are met, the operation segment is marked as abnormal shopping behavior.

[0067] S206. Based on the switching path of the current user between product categories in the high-frequency operation segment, the category switching distance of the current user between adjacent operations is counted; when the category switching distance is greater than a preset distance threshold and the number of switching times exceeds a preset number threshold, it is determined to be an abnormal shopping behavior.

[0068] The product category switching path refers to the track record of users browsing and jumping between different product categories. The category switching distance indicates the degree of hierarchical difference between two product categories in the product classification system. For example, the category distance from "Mobile Digital-Mobile Phone-Apple Mobile Phone" to "Fresh Fruit-Imported Fruit-Apple" is the maximum hierarchical difference. The preset distance threshold is the upper limit of the normal category switching distance allowed, such as setting it to 2 levels. The preset number threshold refers to the upper limit of the number of category switching allowed, such as a maximum of 5 switches within 10 seconds. Abnormal shopping behavior here specifically refers to the behavior pattern of users frequently switching across categories in a short period of time.

[0069] The system first builds a standardized commodity classification tree, assigns a unique identifier to each category node, and defines the distance calculation rules between categories (such as the shortest path method). When analyzing high-frequency operation fragments, the system extracts the user's category switching sequence and calculates the distance value for each category switch between adjacent operations. During the implementation process, the system maintains a switching distance array and a switching count counter. Abnormal shopping behavior is determined when the following situations are detected: within a specified time window (such as 10 seconds), the user's category switching times exceed the preset number threshold (such as 5 times), and at least 3 of the switching categories have a distance greater than the preset distance threshold (such as 2 levels). The system marks the operation fragments that meet these conditions as category switching anomalies and records the specific switching path information.

[0070] S207: Integrate the time characteristics of the high-frequency operation segment, the shopping cart amount change characteristics and the category switching characteristics to obtain the current user's purchase decision behavior characteristics.

[0071] S208: Obtain the total amount of the shopping cart when the current user purchases each of the products, calculate the difference between the total amount of the shopping cart and the threshold amount for full discount, and calculate the first sub-item score based on the difference.

[0072] Time features include statistical indicators of the time dimension, such as the duration of the operation segment, the mean and variance of the operation interval, and the operation density. The shopping cart amount change features include statistical data of the amount dimension, such as the amplitude, direction, frequency, and degree of proximity to the full-reduction threshold. Category switching features include statistical features of the category dimension, such as switching frequency, distribution of switching distances, and regularity of switching paths. The order-completing decision behavior features are a combination of these three types of features, which are used to describe the user's operational behavior patterns in full-reduction activities, especially reflecting the various operational features performed by users to reach the full-reduction threshold.

[0073] The system organizes and stores these feature data in the form of feature vectors. The time feature part includes: operation segment duration (milliseconds), average value of operation interval (milliseconds), standard deviation of operation interval, number of operations per unit time and other numerical indicators. The shopping cart amount change feature part includes: cumulative value of amount change (yuan), number of changes, number of reversals of change direction, minimum difference with the full discount threshold (yuan), average amplitude of amount adjustment (yuan) and other numerical indicators. The category switching feature part includes: total number of switches, average switching distance, maximum switching distance, number of continuous large-distance switches and other numerical indicators. After normalizing these feature values, the system constructs a unified feature vector as the data basis for characterizing the user's decision-making behavior of making a purchase.

[0074] S209: Obtain the operation frequency of the current user in purchasing each of the products within a preset time period, and calculate the second sub-item score based on the operation frequency.

[0075] The total amount of the shopping cart refers to the sum of the prices of all items in the user's shopping cart. The time of adding to cart refers to the specific moment when the user adds a certain item to the shopping cart. The threshold amount for full discount is the target amount set for the full discount activity. The difference refers to the absolute difference between the total amount of the shopping cart and the threshold amount for full discount. The first sub-score is a standardized score calculated based on this difference, which is used to measure the degree of correlation between the user's add-to-cart behavior and the threshold amount for full discount. For example, in the "50 off for purchases over 300" activity, the user's shopping cart originally has 280 yuan of items, and after adding a 15 yuan item, the total amount is 295 yuan, and the difference with the threshold amount of 300 yuan is 5 yuan.

[0076] The system performs the following calculation process for each add-to-cart operation: first record the timestamp of the add-to-cart operation and the amount of the shopping cart before the add-to-cart operation, and then calculate the updated total amount of the shopping cart after adding the new item. Subtract this total amount from the threshold amount (such as 300 yuan) to get the difference value. Use a piecewise function to calculate the score of the first sub-item: when the difference is less than 10 yuan, the score is 90-100 points; when the difference is between 10-30 yuan, the score is 70-89 points; when the difference is between 30-50 yuan, the score is 50-69 points; when the difference is greater than 50 yuan, the score is 0-49 points. The specific score calculation formula is: score=base_score-(gap_value / threshold)*scale_factor, where base_score is the base score, gap_value is the difference, threshold is the segmentation threshold, and scale_factor is the scale factor.

[0077] S210: Obtain a list of historically purchased items of the current user, calculate the category similarity and price range similarity between the current added item and the historically purchased items, and calculate a third sub-item score based on the category similarity and the price range similarity.

[0078] The preset time period refers to a fixed time window for observing user operation behaviors, which is usually set to a period of time before the time of adding to the cart (such as the first 60 seconds). The operation frequency refers to the number of operations (including browsing, adding to cart, deleting, etc.) performed by the user during this time period divided by the length of time, in times / second. The second sub-score is a standardized score calculated based on the operation frequency, which is used to measure the intensity of user operation behaviors. For example, if a user performs 30 operations within 60 seconds, the operation frequency is 0.5 times / second.

[0079] The system performs the following calculation process: First, determine the start time (such as 60 seconds before the time of adding to the purchase) and the end time (the time of adding to the purchase) of the observation window. In this time window, count the total number of all types of user operations. Calculate the operation frequency value: frequency=operation_count / time_window. Set the segmented scoring standard according to the operation frequency: when the frequency is greater than 1 time / second, the score is 90-100 points; when the frequency is between 0.5-1 times / second, the score is 70-89 points; when the frequency is between 0.2-0.5 times / second, the score is 50-69 points; when the frequency is less than 0.2 times / second, the score is 0-49 points. The specific score calculation formula is: score=base_score+(frequency-baseline_frequency)*weight_factor, where base_score is the base score, baseline_frequency is the base frequency, and weight_factor is the weight coefficient.

[0080] S211. Perform a weighted summation of the first sub-item score, the second sub-item score and the third sub-item score according to preset weights to obtain a purchase possibility score for each product. The purchase possibility score is calculated based on the difference in amount between the product and the discount threshold when the product is added to the cart, the user operation frequency when the product is added to the cart, and the similarity between the product and the user's historical purchasing behavior.

[0081] The preset weight refers to the importance coefficient assigned to each sub-item score. The weight of the first sub-item score is usually set to 0.4, the weight of the second sub-item score is 0.3, and the weight of the third sub-item score is 0.3. Weighted summation refers to the calculation process of multiplying each sub-item score by the corresponding weight and adding them together. The order-matching possibility score is a standardized score of 0-100 that comprehensively reflects the possibility that the product will be used to make an order. Among them, the amount difference reflects the matching degree between the product price and the full-reduction threshold, the operation frequency reflects the urgency of user behavior, and the historical purchase similarity reflects the degree of conformity between the product and the user's real needs.

[0082] The system uses the following calculation process: First, the three sub-item scores are standardized to ensure that all scores are within the range of 0-100 points. Then a weighted calculation is performed: final_score=0.4*first_score+0.3*second_score+0.3*third_score, where first_score is the amount difference score, second_score is the operation frequency score, and third_score is the historical similarity score. For each product, the system maintains a score record, including the original sub-item score, weight value, and final order-matching possibility score. The higher the final score, the more likely it is that the product is a user's order-matching choice in order to reach the full discount threshold. For example, if the three sub-item scores of a product are 95 points, 80 points, and 70 points respectively, then its order-matching possibility score is: 0.495+0.380+0.3*70=83 points.

[0083] S212: Obtain the score ranges and corresponding weight adjustment coefficients for the preset multiple levels of order possibility.

[0084] The score range for the possibility of making a purchase is a detailed division of the range of 0-100 points, usually divided into four levels: 90-100 points for the extremely high possibility range, 70-89 points for the high possibility range, 50-69 points for the medium possibility range, and 0-49 points for the low possibility range. The weight adjustment coefficient is a correction parameter corresponding to each score range, which is used to adjust the recommendation weight of the product. For example, the extremely high possibility range corresponds to an adjustment coefficient of 0.3, which means that the original recommendation weight is reduced to 30%. The correspondence between the score range and the adjustment coefficient constitutes a complete weight adjustment plan.

[0085] The system maintains a score range mapping table in the configuration file, which contains the following specific values: 90-100 points correspond to an adjustment coefficient of 0.3, 70-89 points correspond to an adjustment coefficient of 0.5, 50-69 points correspond to an adjustment coefficient of 0.7, and 0-49 points correspond to an adjustment coefficient of 1.0. These values ​​are stored in a hash table structure, with the score range as the key and the adjustment coefficient as the value. The system also records the characteristic description of each range, including the range, the calculation basis of the adjustment coefficient, and the applicable conditions.

[0086] S213: Match the purchase possibility score of each product with the score range, and determine the corresponding weight adjustment coefficient.

[0087] The matching process is to compare the order possibility score of the product with the preset score range to determine which range the score falls into. The determination of the weight adjustment coefficient is to select the corresponding adjustment parameter based on the matching result. For example, the order possibility score of a certain product is 83 points, which falls into the range of 70-89 points, so its weight adjustment coefficient is determined to be 0.5. This process assigns a specific adjustment coefficient to each product for subsequent weight adjustment calculations.

[0088] The system performs the following matching process: First, read the item's order possibility score, and use the binary search algorithm to locate the score interval in the preset score interval. After determining the interval, the system obtains the corresponding weight adjustment coefficient from the hash table. To ensure the accuracy of the calculation, the system matches the score values ​​at the interval boundary (such as 70 points, 90 points) using the higher principle. The system stores the matching results in the item attribute table, including the item ID, order possibility score, interval to which it belongs, and the corresponding weight adjustment coefficient. These data will be used in the subsequent item recommendation weight calculation process. For example, when the item score is 83 points, the system determines that it belongs to the 70-89 point interval and records a weight adjustment coefficient of 0.5.

[0089] S214: Obtain the original recommendation weight of each of the products, and multiply the original recommendation weight by the weight adjustment coefficient to obtain the recommendation weight of each of the products.

[0090] The original recommendation weight refers to the basic weight value of the product in the recommendation system, with a value range of 0-1, and is calculated based on the basic attributes of the product (such as sales volume, rating, profit margin, etc.). For example, a product with a monthly sales volume of 10,000 pieces, a rating of 4.8 points, and a profit margin of 15% has an original recommendation weight of 0.85. The weight adjustment coefficient is a correction parameter determined based on the order possibility score in the previous step, with a value range of 0-1, which is used to adjust the original recommendation weight. The recommendation weight refers to the weight value ultimately used for product sorting, which is obtained by multiplying the original recommendation weight by the weight adjustment coefficient. For example, the original recommendation weight of 0.85 is multiplied by the weight adjustment coefficient of 0.5 to obtain a final recommendation weight of 0.425.

[0091] The system performs the following weight calculation process: First, read the sales volume, rating, profit margin and other data of each product from the product basic attribute table. Use the weighted calculation formula to get the original recommendation weight: original_weight=0.4*sales_score+0.3*rating_score+0.3*profit_score, where sales_score is the sales score (0-1), rating_score is the rating score (0-1), and profit_score is the profit margin score (0-1). Then, read the weight adjustment coefficient determined in the previous step and multiply it by the original recommendation weight: final_weight=original_weight*adjustment_coefficient. The system stores the calculation results in the recommendation weight table, which includes the product ID, original recommendation weight, weight adjustment coefficient and final recommendation weight. These weight values ​​will be directly used for the sorting and display of products. For example, if a product's sales score is 0.9, its rating score is 0.8, and its profit margin score is 0.7, the original recommendation weight is calculated to be 0.81; its order-matching possibility score is 83 points, and the corresponding weight adjustment coefficient is 0.5, then the final recommendation weight is 0.405. The system adjusts the ranking of the product in the recommendation list based on this final weight value, and the higher the weight value, the higher the ranking of the product.

[0092] A recommended product sequence is constructed based on the final recommendation weight of each product, and the products are sorted from high to low according to the weight value. The products are divided into three display levels: key recommendation, regular recommendation and ordinary display according to the preset weight threshold. The recommended product sequence can also be personalized filtered and rearranged in combination with user portrait data. The specific sorting of products in different display levels is adjusted based on the user's category preference, price sensitivity, purchase cycle and other characteristics. The personalized product recommendation sequence is pushed to users in real time through the WebSocket long link mechanism, and the content and order of the pushed products are dynamically adjusted according to factors such as the user's real-time interactive behavior, product inventory changes, price adjustments and activity rule updates. The pushed product sequence is monitored in real time, user interaction data including explicit feedback and implicit feedback are collected, core indicators such as click-through rate, conversion rate, and full-reduction achievement rate are calculated, and the recommendation strategy is dynamically optimized based on the monitoring results.

[0093] In some embodiments, the following steps may be further included after step S214: Based on the convolutional neural network, the static features of the current user in the full-discount activity are extracted, and the static features include product category preference and price sensitivity.

[0094] Specifically, a convolutional neural network (CNN) is a deep learning model that extracts data features through multi-layer convolution and pooling operations. Static features refer to the user's behavioral preference features that remain relatively stable over a long period of time. Product category preferences are expressed as the user's historical purchase ratio and browsing time distribution in different product categories. For example, the user's purchase ratio in the "digital electronics" category is 40%, and the "clothing" category is 30%. Price sensitivity reflects the user's reaction to the price of the product, which is measured by the price distribution of the user's historical purchases and the promotion participation. For example, the average price range of the user's purchases is 100-300 yuan, and the frequency of participating in promotion activities is 3 times a month. The system first constructs a standardized user behavior feature matrix, which contains multi-dimensional data such as the number of purchases, purchase amount, browsing time, and number of favorites of users in each product category. This feature matrix is ​​input into the pre-trained CNN model, which contains 3 convolutional layers and 2 pooling layers. The first convolution layer uses 32 3x3 convolution kernels to extract low-level features, the second convolution layer uses 64 3x3 convolution kernels to extract mid-level features, and the third convolution layer uses 128 3x3 convolution kernels to extract high-level features. Each convolution layer is followed by a maximum pooling layer with a pooling window size of 2x2. Finally, the fully connected layer maps the features into a category preference vector (the dimension is the number of product categories) and a price sensitivity scalar (value 0-1).

[0095] Based on the recurrent neural network (RNN), the temporal features of the current user's shopping cart operations are extracted, and the temporal features include a continuous pattern of adding, deleting and modifying commodities.

[0096] Specifically, the recurrent neural network (RNN) is a deep learning model that specializes in processing sequence data and can capture the temporal dependencies in the data. The temporal features describe the dynamic change pattern of user behavior. The continuity pattern of adding, deleting, and modifying items in the shopping cart refers to the time series features of the user's operations such as adding, deleting, and modifying the quantity of items in the shopping cart, including dimensions such as operation interval, operation sequence, and operation intensity. For example, the user adds 3 related items in succession within 10 minutes, then deletes 1 item, and finally modifies the quantity. The system uses the long short-term memory network (LSTM) to extract temporal features. First, the user's shopping cart operation sequence is converted into a vector sequence, each vector containing information such as the operation type (one-hot encoding), operation timestamp, product ID, and shopping cart status before and after the operation. The LSTM network contains 128 hidden units, the input layer receives the vector sequence, and learns long-term dependencies through a gating mechanism. The network's forget gate controls the degree of retention of historical information, the input gate determines the degree of update of current information, and the output gate generates the final temporal feature representation. The system divides the operation sequences in the last 24 hours into 10-minute units, and each sequence is independently input into the LSTM network to obtain the behavioral pattern characteristics within the time window.

[0097] The static feature and the time series feature are input into a preset fusion model to obtain the purchase tendency probability of the current user for different commodity categories.

[0098] Specifically, the preset fusion model is a multi-layer neural network that integrates different types of features and outputs prediction results. The purchase propensity probability indicates the possibility that a user will purchase a certain type of product in the future, and the value range is 0-1. The fusion model uses the attention mechanism to assign different weights to static features and temporal features to achieve adaptive combination of features. For example, when it is detected that a user is participating in a promotion activity of a certain category, the weight of the relevant temporal features will be automatically increased.

[0099] The fusion model adopts a three-layer structure: the first layer is the feature fusion layer, which uses the attention mechanism to calculate the weight coefficients of static features and temporal features, and the sum of the weights is 1. The second layer is the feature interaction layer, which contains 256 neurons, uses the ReLU activation function, and retains the original feature information through residual connections. The third layer is the prediction layer, which outputs a probability value for each product category and normalizes it using the Softmax function. The model training uses the negative log-likelihood loss function, uses the Adam optimizer for parameter updates, and the learning rate is set to 0.001. When running online, the model uses the latest user behavior data to update the prediction results every hour to provide basic data support for subsequent product recommendations.

[0100] Based on the price sensitivity characteristics of the current user, determine the optimal recommended price range.

[0101] Specifically, the price sensitivity feature is an indicator of the user's reaction degree to changes in commodity prices, including dimensions such as historical purchase price distribution, promotion participation frequency, and price comparison behavior. The optimal recommended price range is calculated based on the user's price sensitivity feature, and is the price range that is most likely to promote purchases, determined by two boundary values: the minimum price and the maximum price. For example, a user's historical purchase records show that 90% of successful purchases occurred in the price range of 200-500 yuan, and the promotion activity participation in this range is the highest. This range is determined as the optimal recommended price range, and the system determines the optimal price range through the following steps: First, calculate the price distribution of the user's historical purchases, and use the kernel density estimation method to obtain the price probability density function. Then analyze the user's purchase conversion rate in different price ranges, divide the price axis into multiple ranges (such as 0-100 yuan, 100-300 yuan, 300-500 yuan, etc.), and calculate the purchase probability of each range. Combined with the user's promotion sensitivity (purchase ratio of promotional products) and price comparison behavior (number of price comparisons of similar products), the comprehensive score of each price range is calculated using a weighted summation method. Select the price range with the highest score as the optimal recommended price range, and record the upper and lower boundary values ​​of the range.

[0102] The product recommendation list is sorted according to the purchase tendency probability and the optimal recommended price range.

[0103] Specifically, the purchase propensity probability is the user's purchase possibility for different commodity categories output by the prediction model, and the value range is 0-1. The commodity recommendation list is a set of candidate commodities to be sorted, and each commodity contains attributes such as ID, category, and price. Sorting is to sort the commodities in descending order according to the preset scoring rules, and the scoring rules comprehensively consider the purchase propensity probability and price matching. For example, the purchase propensity probability of a mobile phone accessory is 0.8, and the price of 299 yuan falls within the optimal interval, so it obtains a higher comprehensive score. The system performs the following sorting process: First, filter out the subset of commodities with prices within the optimal recommendation interval. Calculate the comprehensive score for each commodity, and the scoring formula is: score=w1*probability+w2*price_match_score, where probability is the purchase propensity probability of the category to which the commodity belongs, price_match_score is the price matching score (the closer the price is to the center of the optimal interval, the higher the score), w1 and w2 are weight coefficients and the sum is 1. Use the quick sort algorithm to sort the commodities in descending order according to the score to generate the final recommendation sequence. Each commodity record contains fields such as commodity basic information, comprehensive score, and sorting position.

[0104] The sorted product recommendation list is monitored in real time to obtain monitoring results.

[0105] Specifically, real-time monitoring refers to continuously tracking and recording the display effect and user feedback of recommended products. The monitoring results include multiple evaluation indicators: click-through rate (conversion ratio from display to click), dwell time (time users view products), add-to-cart rate (conversion ratio from click to add-to-cart), and purchase rate (conversion ratio from add-to-cart to purchase). For example, if a recommended product is displayed 1,000 times and clicked 100 times, the click-through rate is 10%. The system establishes a real-time monitoring mechanism: embedding code in the recommendation interface to record product display, click, add-to-cart, purchase and other events. Use a sliding time window (such as 15 minutes) to count various indicators and store the statistical results in the monitoring data table. Monitoring indicators include: average clicks per person, average dwell time, add-to-cart conversion rate, purchase conversion rate, etc. The system also records the indicator change trend of each product in different time periods and constructs a time series monitoring matrix. When an indicator is abnormal (such as a sudden drop in click-through rate), the abnormal event is recorded and the affected product is marked.

[0106] The weight coefficient of the static feature is adjusted to a preset weight threshold based on the monitoring result.

[0107] Specifically, the static feature weight coefficient is a parameter that controls the influence of the user's long-term behavior characteristics in the recommendation system, and its value range is 0-1. The preset weight threshold is the weight adjustment target value set according to the business rules, which is used to limit the adjustment range. Weight adjustment is a process of dynamically optimizing the importance of features based on the feedback of monitoring indicators. For example, when monitoring finds that the user's responsiveness to the historical preference category decreases, the weight of the corresponding static feature is reduced, and the system performs the following adjustment process: First, the change rate of each indicator in the monitoring period is calculated, and the change rate is mapped to the weight adjustment factor. For feature dimensions whose indicators such as click-through rate and dwell time are lower than expected, their weights are reduced according to the adjustment factor. The weight adjustment formula is: new_weight=old_weight*(1+adjustment_factor), where adjustment_factor is the adjustment factor, and its value range is [-0.5,0.5]. The adjusted weight is subject to the preset threshold and must not exceed the range of [0.1,0.9]. The system performs weight adjustment every 30 minutes and updates the new weight to the feature vector.

[0108] The user analysis system in the embodiment of the present invention is described below from the perspective of hardware processing. Figure 3 , which is a schematic diagram of a physical device structure of a user analysis system in an embodiment of the present application.

[0109] It should be noted that Figure 3 The structure of the user analysis system shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0110] like Figure 3 As shown, the user analysis system includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 302 or the program loaded from the storage part 308 to the random access memory (RAM) 303, such as executing the method described in the above embodiment. In the RAM 303, various programs and data required for system operation are also stored. The CPU 301, the ROM 302 and the RAM 303 are connected to each other through the bus 304. The input / output (I / O) interface 305 is also connected to the bus 304.

[0111] The following components are connected to the I / O interface 305: an input section 306 including an audio input device, a button switch, etc.; an output section 307 including a liquid crystal display (LCD) and an audio output device, an indicator light, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the I / O interface 305 as needed. A removable medium 311, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 310 as needed so that a computer program read therefrom is installed into the storage section 308 as needed.

[0112] In particular, according to an embodiment of the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present invention includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through the communication part 309, and / or installed from a removable medium 311. When the computer program is executed by the central processing unit (CPU) 301, various functions defined in the present invention are performed.

[0113] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in combination with an instruction execution system, apparatus, or device.

[0114] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. Each box in the flowchart or block diagram may represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box may also occur in an order different from that marked in the accompanying drawings.

[0115] Specifically, the user analysis system of this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, the user analysis method based on big data provided in the above embodiment is implemented.

[0116] As another aspect, the present invention further provides a computer-readable storage medium, which may be included in the user analysis system described in the above embodiment; or may exist independently without being assembled into the user analysis system. The above storage medium carries one or more computer programs, and when the above one or more computer programs are executed by a processor of the user analysis system, the user analysis system implements the user analysis method based on big data provided in the above embodiment.

[0117] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

[0118] As used in the above embodiments, the term "when..." may be interpreted to mean "if..." or "after..." or "in response to determining..." or "in response to detecting...", depending on the context. Similarly, the phrases "upon determining..." or "if (the stated condition or event) is detected" may be interpreted to mean "if determining..." or "in response to determining..." or "upon detecting (the stated condition or event)" or "in response to detecting (the stated condition or event)", depending on the context.

[0119] Those skilled in the art can understand that to implement all or part of the processes in the above-mentioned embodiments, the processes can be completed by computer programs to instruct related hardware, and the programs can be stored in computer-readable storage media. When the programs are executed, they can include the processes of the above-mentioned method embodiments. The aforementioned storage media include: ROM or random access memory RAM, magnetic disk or optical disk and other media that can store program codes.

Claims

1. A user analysis method based on big data, characterized in that: Applied to a user analysis system, the method comprises: Obtain the specific operation sequence of the current user in the current full-discount promotion scenario, where the current full-discount promotion scenario is an e-commerce promotion rule including a full-discount threshold amount and a discount amount for a product, and the specific operation sequence includes the time point when the user enters the full-discount activity page, a record of changes in the amount of products in the shopping cart, a path for switching between product categories, a record of product addition, deletion, and modification operations, and a product browsing order; Extracting a purchase-matching decision behavior feature in the specific operation sequence, wherein the purchase-matching decision behavior feature includes abnormal shopping behavior; Based on the above-mentioned buying decision behavior characteristics, a buying possibility score is calculated for each product added by the user. The buying possibility score is calculated based on the difference in amount between the product and the discount threshold when the product is added, the user operation frequency when adding the product, and the similarity between the product and the user's historical purchase behavior; The recommendation weight of each of the products is determined based on the possibility score of adding to the order. When the possibility score of adding to the order of a target product exceeds a preset score threshold, the recommendation weight of the target product in the current full-discount promotion scenario is reduced to a preset weight.

2. The method according to claim 1, characterized in that The step of extracting the order-matching decision behavior features in the specific operation sequence specifically includes: Obtaining the operation time sequence after the current user enters the full-discount activity page in the specific operation sequence; Calculating the time intervals between adjacent operations in the operation time series; When it is detected that the time interval between multiple consecutive operations in an operation segment is less than a preset time threshold, the operation segment is marked as a high-frequency operation segment; Analyze the change records of the shopping cart item amounts in the high-frequency operation segments, and determine it as abnormal shopping behavior when the change direction of the shopping cart item amounts undergoes multiple reverse adjustments and the adjusted amount is close to the full-reduction threshold amount; Based on the switching path of the current user between commodity categories in the high-frequency operation segment, the category switching distance between adjacent operations of the current user is counted, and when the category switching distance is greater than a preset distance threshold and the number of switching times exceeds a preset number threshold, it is determined to be an abnormal shopping behavior; The time characteristics of the high-frequency operation segments, the shopping cart amount change characteristics and the category switching characteristics are integrated to obtain the current user's purchase decision behavior characteristics.

3. The method according to claim 1, characterized in that The step of calculating the order-adding possibility score for each item purchased by the user based on the order-adding decision behavior characteristics specifically includes: Obtaining the total amount of the shopping cart when the current user purchases each of the products, calculating the difference between the total amount of the shopping cart and the threshold amount for full discount, and calculating the first sub-item score based on the difference; Acquire the operation frequency of the current user in purchasing each of the products within a preset time period, and calculate the second sub-item score based on the operation frequency; Obtain a list of historically purchased items of the current user, calculate the category similarity and price range similarity between the current added-to-purchase item and the historically purchased items, and calculate a third sub-item score based on the category similarity and the price range similarity; The first sub-item score, the second sub-item score and the third sub-item score are weighted and summed according to preset weights to obtain a purchase possibility score for each of the products.

4. The method according to claim 1, characterized in that: The step of determining the recommendation weight of each of the products according to the possibility score of placing an order specifically includes: Obtain the score ranges and corresponding weight adjustment coefficients for the preset multiple levels of order possibility; Matching the order-matching possibility score of each product with the score range, and determining the corresponding weight adjustment coefficient; The original recommendation weight of each of the commodities is obtained, and the original recommendation weight is multiplied by the weight adjustment coefficient to obtain the recommendation weight of each of the commodities.

5. The method according to claim 1, characterized in that After the step of reducing the recommendation weight of the target product in the current purchase-to-order discount promotion scenario to a preset weight when the purchase-to-order possibility score of the target product exceeds a preset score threshold, the method further includes: Extracting static features of the current user in the full-reduction activity based on a convolutional neural network, wherein the static features include product category preference and price sensitivity; Extracting the time series features of the current user's shopping cart operation based on a recurrent neural network (RNN), wherein the time series features include a continuous pattern of adding, deleting, and modifying commodities; The commodity recommendation list is sorted according to the static feature and the temporal feature.

6. The method according to claim 5, characterized in that The step of sorting the commodity recommendation list according to the static feature and the temporal feature specifically includes: Inputting the static features and the time series features into a preset fusion model to obtain the purchase tendency probability of the current user for different commodity categories; Determine the optimal recommended price range based on the price sensitivity characteristics of the current user; The commodity recommendation list is sorted according to the purchase tendency probability and the optimal recommended price range.

7. The method according to claim 5, characterized in that After the step of sorting the commodity recommendation list according to the static feature and the temporal feature, the method further includes: Monitor the sorted product recommendation list in real time to obtain monitoring results; The weight coefficient of the static feature is adjusted to a preset weight threshold based on the monitoring result.

8. A user analysis system, characterized in that: The user analysis system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the user analysis system to execute the method described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that: When the instructions are executed on a user analysis system, the user analysis system is caused to execute the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that When the computer program product is run on a user analysis system, the user analysis system is caused to execute the method according to any one of claims 1 to 7.

Images